Advancements in Breast Cancer Detection: A Review of Global Trends, Risk Factors, Imaging Modalities, Machine Learning, and Deep Learning Approaches

Abstract

Breast cancer remains a critical global health challenge, with over 2.1 million new cases annually. This review systematically evaluates recent advancements (2022–2024) in machine and deep learning approaches for breast cancer detection and risk management. Our analysis demonstrates that deep learning models achieve 90–99% accuracy across imaging modalities, with convolutional neural networks showing particular promise in mammography (99.96% accuracy) and ultrasound (100% accuracy) applications. Tabular data models using XGBoost achieve comparable performance (99.12% accuracy) for risk prediction. The study confirms that lifestyle modifications (dietary changes, BMI management, and alcohol reduction) significantly mitigate breast cancer risk. Key findings include the following: (1) hybrid models combining imaging and clinical data enhance early detection, (2) thermal imaging achieves high diagnostic accuracy (97–100% in optimized models) while offering a cost-effective, less hazardous screening option, (3) challenges persist in data variability and model interpretability. These results highlight the need for integrated diagnostic systems combining technological innovations with preventive strategies. The review underscores AI’s transformative potential in breast cancer diagnosis while emphasizing the continued importance of risk factor management. Future research should prioritize multi-modal data integration and clinically interpretable models.

1. Introduction

Breast cancer (BC) is the leading form of cancer among women, comprising 25% of global female cancer cases. In 2020, there were roughly 2.3 million new cases and 685,000 deaths worldwide. Early and accurate detection remains critical, as malignant tumors can spread aggressively, significantly reducing the probability of survival if left untreated [1]. Advances in artificial intelligence (AI) and machine learning (ML) have transformed medical image analysis, offering innovative tools for BC diagnosis. Techniques such as convolutional neural networks (CNNs) have shown promise in classifying and diagnosing BC using diverse imaging modalities, including mammography, CT, and ultrasound. By incorporating CAD systems and utilizing deep learning (DL), the fatality rate of breast carcinoma decreased by 40% from the 1980s to 2020, with projections suggesting that up to 2.5 million lives could be saved by 2040 if global mortality rates decline by 2.5% annually [2]. Figure 1 and Figure 2 illustrate the global incidence and mortality rates of BC among females in 2022. Figure 1 shows that Asia has the highest incidence of BC, accounting for 42.92% of cases, followed by Europe at 24.27%. Oceania reports the lowest incidence rate, at just 1.24%. Regarding mortality, 47.34% of female deaths occurred in Asia, while Europe accounted for 21.68%. These statistics demonstrate that over 65% of BC cases and deaths are concentrated in Asia and Europe.

Figure 1. Global breast cancer incidence rate in females in 2022 [3].

Figure 2. Global breast cancer mortality rate in females in 2022 [3].

Figure 3 presents the mortality rates of females from the fourteen leading cancers worldwide in 2022, including ‘Breast’; ‘Trachea, bronchus, and lung’; ‘Colorectum’; ‘Cervix uteri’; ‘Liver and intrahepatic bile ducts’; ‘Stomach’; ‘Pancreas’; ‘Ovary’; ‘Leukaemia’; ‘Brain, central nervous system’; ‘Corpus uteri’; ‘Kidney’; ‘Bladder’; and ‘Thyroid’. The figure highlights that 19.71% of female cancer-related deaths are due to BC, making it the leading cause of cancer-related mortality among women [3].

Figure 3. Female mortality rates from all cancers worldwide in 2022 [3].

1.1. Motivation

Breast cancer remains a significant global health concern, particularly among women, where it stands as one of the most prevalent and life-threatening forms of cancer. Despite advances in medical imaging and therapeutic strategies, there are still persistent challenges to early and accurate diagnosis, representing a major hurdle to effective treatment.

While DL models achieve high accuracy, their ‘black-box’ nature remains a barrier to clinical trust. Recent advancements in explainable AI (XAI) techniques, such as Grad-CAM and SHAP, are critical for bridging this gap. This review evaluates not only performance metrics but also the interpretability of state-of-the-art models, highlighting their clinical applicability.

Below are the key motivations for this study.

A1.

Breast cancer affects millions annually, with increasing prevalence across diverse age groups. Traditional diagnostic techniques such as mammography are often inadequate, particularly in dense breast tissues, leading to missed or delayed diagnoses.

A2.

DL techniques have revolutionized medical imaging by enabling precise analysis of complex, high-dimensional datasets. These methods reduce human error, increase diagnostic consistency, and improve efficiency.

A3.

Existing screening tools have inherent strengths and weaknesses. Investigating these methods to identify the most effective approaches for breast cancer detection remains critical for advancing diagnostic quality.

A4.

Innovations such as transfer learning, hybrid DL models, and explainable AI have enhanced diagnostic accuracy while addressing trust and interpretability concerns, essential factors for clinical integration.

A5.

By integrating advanced technologies into clinical practice, we can increase early detection rates, potentially saving countless lives and transforming breast cancer treatment outcomes.

1.2. Contributions

This paper offers a comprehensive review of advancements in BC detection, with a strong focus on the transformative impact of machine learning and DL. The key contributions are as follows:

A1.

The global BC trends and risk factors are examined to inform preventive strategies.

A2.

We present a visualization of current trends in BC research, review the impacts of early detection and technology on mortality rates, and highlight the potential of lifestyle changes to reduce the incidence of breast cancer in high-risk groups.

A3.

This review examines various BC imaging techniques, including mammograms, ultrasound, MRI, CT scans, histopathology, and thermal imaging. We discuss their advantages and disadvantages and present examples of specific cases.

A4.

We explore the use of structured data (CSV) for BC detection, focusing on feature engineering, preprocessing, and algorithmic approaches.

A5.

The study presents a summary of hybrid frameworks combining deep feature extraction and machine learning classifiers to enhance diagnostic accuracy.

A6.

We report the level of accuracy achieved to date by various models using different datasets.

2. Background Study

Breast cancer is a major global health issue, with the World Health Organization (WHO) reporting 7.8 million women being diagnosed in the previous three years, making it the most prevalent cancer worldwide. Breast cancer accounts for 4–5.5% of all new cancer cases annually, significantly contributing to morbidity rates [4,5]. In the United States alone, over 250,000 individuals are diagnosed with BC each year; the disease accounts for 41,690 deaths, making it the most frequent and fatal cancer type among women. BC represents approximately 25% of all cancer diagnoses in women globally [6]. Early diagnosis is critical, as patients diagnosed in the early stages have a five-year survival rate of about 99% [7,8].

Breast cancer typically begins in the ducts or lobules and manifests through symptoms such as lumps, changes in breast size or shape, and abnormal discharge. Traditional diagnostic methods, including mammography, histopathological analysis, and other imaging techniques, are widely employed but are time-consuming, requiring skilled professionals, and are prone to human error. For example, mammograms rely on radiologists for interpretation, while histopathological analysis demands experienced pathologists to accurately classify tissues [9,10,11,12]. Breast cancer was first identified in Egypt around 1600 BC and has been responsible for approximately 15% of female deaths [13,14]. While BC affects men, women, and transgender individuals, its prevalence is significantly higher among females. The risk for transgender individuals, particularly the impact of gender-affirming hormonal therapy (GAHT) on BC, remains unclear, and guidelines for screening in this group are still undefined [15,16].

The WHO projects that BC cases will increase to 2.7 million by 2040. The National Breast Cancer Coalition (NBCC) has reported a rise in the lifetime risk of developing invasive breast cancer among women in the US since 1975, with 287,850 new cases identified in women and 2710 in men. Additionally, the American Chemical Society (ACS) notes that a woman’s likelihood of being diagnosed with invasive breast cancer increased from 9.09% to 12.9% during the same period [17,18,19,20]. Early and accurate BC detection is essential for improving survival rates, but early-stage diagnosis remains challenging. Diagnostic tools such as mammography, magnetic resonance imaging (MRI), computed tomography (CT), and ultrasonography are critical, with mammography being the most widely used method. Mammography has a sensitivity range of 77–95% and a specificity range of 92–95%, yet its accuracy can be affected by dense breast tissue. To address these limitations, advancements in technology, particularly machine learning (ML), are being utilized to enhance diagnostic precision, offering significant promise for early detection and personalized treatment [15].

In India, breast cancer ranks among the top five most common cancers, with new cases rising from 159,417 in 2017 to 178,361 in 2020, comprising nearly 26% of all cancer cases in women [14].

The remainder of the paper is organized as follows: Section 3 presents the research questions and includes a comparison table of our review with existing review papers on breast cancer detection. Section 4 discusses the most popular datasets used in breast cancer detection research, while Section 5 details the machine learning techniques applied for breast cancer detection. Section 6 provides statistics on various cancers, focusing on mortality and incidence rates for breast cancer across different categories. Section 7 outlines current diagnostic techniques for breast cancer and highlights key disease-associated factors. Section 8 analyzes different breast cancer detection approaches, followed by Section 9, which examines their applications, strengths, and weaknesses. Section 10 presents an overview of publication trends across journals. Finally, Section 11 concludes the paper, and Section 12 proposes future research directions to advance breast cancer detection and management.

3. Research Methodology

This section outlines the research methodology used to conduct a comprehensive review of breast cancer detection techniques using machine learning and deep learning approaches.

3.1. Study Selection Strategy

To ensure a rigorous and unbiased review, we implemented a structured selection process for identifying relevant studies:

3.1.1. Search Strategy

SS1.

Conducted searches across major academic databases including IEEE Xplore, SpringerLink, ScienceDirect, Nature, Wiley Online Library, and MDPI.

SS2.

Used keywords: (“breast cancer detection” OR “breast cancer diagnosis”) AND (“machine learning” OR “deep learning” OR “artificial intelligence”) combined with modality-specific terms (“mammography”, “ultrasound”, “MRI”, “CT”, “Histopathological”, “Thermal”, etc.).

3.1.2. Inclusion Criteria

IC1.

Primary focus on studies published between 2023 and 2024.

IC2.

Included a few foundational papers from 2019 to 2022 that introduced significant methodological advances.

IC3.

Included studies were required to report quantitative performance metrics such as accuracy, AUC, sensitivity, and specificity.

3.1.3. Exclusion Criteria

EC1.

Non-English publications, conference abstracts, and editorials.

EC2.

Duplicate or overlapping studies.

3.2. Research Questions

The article addresses ten key research questions, as outlined below. Additionally, Table 1 provides a comparison of existing review papers on breast cancer detection with our own, highlighting the main differences and contributions. The key research questions explored in this review include the following:

Table 1. Comparative analysis of recently published review papers and our review of literature on breast cancer detection.

Review Papers	Year	Research Questions
		Q1	Q2	Q3	Q4	Q5	Q6	Q7	Q8	Q9	Q10
Abhisheka et al. [21]	2023		✓			✓	✓	✓		✓
Nasser et al. [22]	2023		✓				✓	✓		✓
Thakur et al. [23]	2024		✓			✓	✓	✓		✓
Rautela et al. [15]	2024		✓	✓			✓	✓		✓
Sharafaddini et al. [24]	2024		✓			✓	✓	✓		✓
Sushanki et al. [25]	2024							✓
Our review	2025	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓

Q1.

What are the incidence and death rates for breast cancer across various categories?

Q2.

What diagnostic methods are currently used for breast cancer detection?

Q3.

What are the key risk factors for breast cancer?

Q4.

How can lifestyle changes help reduce the risk of breast cancer?

Q5.

Which breast cancer datasets have been used in the development of machine and deep learning models?

Q6.

What are the reasons for adopting machine and deep learning techniques in breast cancer detection?

Q7.

What are the different medical imaging modalities used for breast cancer classification, and what is the highest accuracy achieved by various models on these modalities to date?

Q8.

What are the various CSV datasets employed for breast cancer classification, and what is the best accuracy attained by different models on these datasets so far?

Q9.

What are the strengths, weaknesses, and applications of imaging modalities in breast cancer detection techniques?

Q10.

What are the benefits, drawbacks, and use cases of breast cancer detection techniques based on CSV datasets?

4. Overview of Breast Cancer Datasets

Researchers utilize diverse breast cancer datasets across multiple modalities—including X-ray, ultrasound, MRI, histopathology, thermography, and structured clinical records—for developing and validating detection algorithms. These datasets vary significantly in scale (from hundreds to thousands of samples), content complexity (from basic images to comprehensive multimodal clinical data), and accessibility (ranging from open public repositories to restricted institutional databases). For systematic evaluation, Table 2 provides a short description for the modalities, detailing key characteristics such as sample size, benign/malignant class distribution, annotation quality, and access protocols, enabling researchers to identify optimal resources for their specific computational analysis needs.

Table 2. Comprehensive overview of multimodal breast cancer datasets.

Ref.	Dataset	Description	Source	Modality
[26]	The Chinese Mammography Database (CMMD)	This dataset contains 3728 mammograms from 1775 patients (China, 2012–2016), with biopsy-confirmed diagnoses and molecular subtypes (Luminal A/B, HER2+, Triple-negative) for 749 cases.	The dataset is publicly available at https://www.cancerimagingarchive.net/collection/cmmd/ (accessed on 5 April 2025)	X-Ray
[27]	Mammographic Image Analysis Society (MIAS)	The MIAS (Mammographic Image Analysis Society) database contains 322 digitized mammogram images from the UK National Breast Screening Program. Stored on a 2.3 GB 8 mm (Exabyte) tape, all images are clipped, padded, and resized to 1024 × 1024 pixels with a 200-micron resolution. It includes annotations on breast density, abnormality type, location, and severity, and is widely used for breast cancer detection research.	The dataset is publicly available at https://www.kaggle.com/datasets/kmader/mias-mammography (accessed on 5 April 2025)	X-Ray
[28]	Digital Database for Screening Mammography (DDSM)	The Digital Database for Screening Mammography (DDSM) is a publicly accessible resource created to aid research in mammographic image analysis and breast cancer detection. It consists of 2620 digitized mammography cases, which include a variety of benign, benign without callback, typical, and cancer cases carefully selected for the database. Each case includes mammographic images of both breasts, along with detailed annotations specifying lesion types, locations, and pathology outcomes.	The DDSM is hosted by the University of South Florida and available at http://www.mammoimage.org/databases/ (accessed on 5 April 2025)	X-Ray
[29]	Curated Breast Imaging Subset of DDSM (CBIS-DDSM)	The CBIS-DDSM dataset comprises 1566 standardized mammography studies selected from the original DDSM database of 2620 cases. It features DICOM-formatted images, precise lesion annotations, and verified pathology data to support CAD research.	The dataset is publicly available for academic use through The Cancer Imaging Archive: https://www.cancerimagingarchive.net/collection/cbis-ddsm/ (accessed on 5 April 2025)	X-Ray
[30]	INbreast	The INbreast dataset is a public digital mammography database containing 115 cases (410 images), including 90 bilateral cases (4 images each) and 25 mastectomy cases (2 images each). It features diverse breast lesions (masses, calcifications, asymmetries, distortions) with precise radiologist annotations in XML format.	The INbreast dataset was developed at Centro Hospitalar de São João in Porto, Portugal, with ethical approval from the Portuguese National Data Protection Commission and the Hospital Ethics Committee, using Siemens MammoNovation full-field digital mammography systems.	X-Ray
[31]	Breast Ultrasound and US-Guided Interventional Techniques: A Multimedia Teaching File	This specialized collection contains 306 breast ultrasound images with an average resolution of 377 × 396 pixels, extracted from 51 clinical cases. The images cover various breast conditions and were originally compiled for professional medical education purposes.	The “Breast Ultrasound and US-Guided Interventional Techniques” dataset was commercially distributed as a CD-ROM (2003) through BreastUltrasound.com (Thessaloniki, Greece).	Ultrasound
[32]	Dataset B	This dataset contains 163 breast ultrasound images (size 760 × 570 pixels) collected in 2012 at UDIAT Diagnostic Center, Parc Taulí Corporation (Sabadell, Spain). It includes 53 malignant and 110 benign lesion cases from different patients.	The breast cancer imaging dataset can be accessed online at goo.gl/SJmoti for research and development purposes.	Ultrasound
[33]	Open Access Series of Breast Ultrasound Data (OASBUD)	The Open Access Series of Breast Ultrasound Data (OASBUD) contains raw radiofrequency ultrasound scans from 78 female patients, featuring 98 breast lesions (52 malignant and 46 benign).	Researchers may access the publicly available dataset at https://zenodo.org/records/545928#.XYp6mShKi9w (accessed on 6 April 2025)	Ultrasound
[34]	Breast Ultrasound Dataset	This breast ultrasound dataset contains 780 images (500 × 500 pixels) from 600 female patients (aged 25–75), collected in 2018 at Baheya Hospital. It includes 437 benign, 210 malignant, and 133 normal cases	The dataset is publicly available for research purposes and can be accessed at https://www.kaggle.com/datasets/aryashah2k/breast-ultrasound-images-dataset (accessed on 6 April 2025)	Ultrasound
[35]	UDAIT	The UDAIT breast ultrasound collection contains 163 images (110 benign, 53 malignant masses) from the UDAIT Diagnostic Centre (Parc Taulí, Spain). Captured via Siemens ACUSON scanners, each image depicts a single mass.	The dataset is publicly available at https://www.kaggle.com/datasets/jarintasnim090/udiat-data (accessed on 6 April 2025)	Ultrasound
[36]	Reference Image Database to Evaluate Therapy Response (RIDER)	The RIDER Breast MRI dataset contains 500 dynamic contrast-enhanced MRI examinations stored in The Cancer Imaging Archive, with each 288 × 288 pixel scan comprising 60 slices and corresponding expert-delineated tumor contours.	The dataset is publicly available at https://www.cancerimagingarchive.net/collection/rider-breast-mri/ (accessed on 11 April 2025)	MRI
[37]	The Cancer Genome Atlas (TCGA)	The Cancer Genome Atlas (TCGA) is a landmark NCI-led initiative that molecularly characterized over 20,000 primary tumors and matched normal samples across 33 cancer types, generating 2.5 petabytes of multi-omics data (genomics, transcriptomics, epigenomics, proteomics) from 2006 to 2018.	Researchers may obtain the dataset from the public repository at https://www.cancer.gov/ccg/research/genome-sequencing/tcga (accessed on 11 April 2025)	MRI
[38]	QIN-BREAST	The QIN-BREAST dataset provides longitudinal PET/CT and MRI scans (DWI, DCE, T1-mapping) acquired during neoadjuvant breast cancer therapy. Collected using standardized protocols (GE PET/CT, Philips 3T MRI), it enables quantitative treatment response assessment.	The data, including multi-center extension QIN-BREAST-02, is available at https://www.cancerimagingarchive.net/collection/qin-breast/ (accessed on 11 April 2025)	MRI
[39]	BreAst Cancer Histology (BACH)	The dataset comprises 400 microscopy images (TIFF format, 2048 × 1536 pixels, 0.42 µm/pixel) equally distributed across four classes: normal (100), benign (100), in situ carcinoma (100), and invasive carcinoma (100). Each RGB image (10–20 MB) includes image-wise labels.	The dataset is publicly available at https://iciar2018-challenge.grand-challenge.org/Dataset/ (accessed on 13 April 2025)	Histopathology
[40]	BreakHis	The BreakHis dataset contains 9109 breast tumor histopathology images (700 × 460 pixels, RGB, PNG) from 82 patients, including 2480 benign and 5429 malignant samples at four magnifications (40×, 100×, 200×, 400×).	The dataset is openly accessible at https://web.inf.ufpr.br/vri/databases/breast-cancer-histopathological-database-breakhis/ (accessed on 13 April 2025)	Histopathology
[41]	DBT-TU-JU	The DBT-TU-JU Breast Thermogram dataset contains 1100 thermal images from 100 patients, collected at GBP Hospital, Tripura. It includes 6 views per patient (frontal, lateral, oblique, supine) with 45 normal, 49 abnormal (33 benign, 13 malignant), and 6 unknown cases, validated via mammography/FNAC reports. Expert-annotated hotspot regions are provided for abnormal cases.	Internal patient databases are only accessible to the individual patients and their treating physicians.	Thermography
[42]	Digital Mammography and Radiology – Infrared (DMR-IR)	The DMR-IR dataset contains 3534 infrared (IR) breast images from 141 patients, collected at Hospital Universitário Antônio Pedro (Brazil).	The images are stored in the Database for Research Mastology with Infrared Image—DMR-IR, accessible on the website https://visual.ic.uff.br/dmi/ (accessed on 13 April 2025)	Thermography
[43]	Wisconsin Diagnostic Breast Cancer (WDBC)	The Wisconsin Diagnostic Breast Cancer (WDBC) dataset contains 569 samples (357 benign, 212 malignant) with 30 numerical features derived from cell nuclei characteristics	The data is available at https://www.kaggle.com/datasets/mohaiminul101/wisconsin-diagnostic-breast-cancer-wdbc (accessed on 13 April 2025)	CSV
[44]	SEER	The SEER Breast Cancer Dataset contains clinical records of 4024 female patients diagnosed with infiltrating ductal/lobular carcinoma (2006–2010), curated from the NCI’s SEER Program (2017 update).	The dataset is available at https://ieee-dataport.org/open-access/seer-breast-cancer-data (accessed on 13 April 2025)	CSV
[45]	METABRIC	The METABRIC (Molecular Taxonomy of Breast Cancer International Consortium) dataset is a widely used breast cancer resource containing clinical, genomic, and transcriptomic data from 2000 patients. It includes tumor characteristics, survival outcomes, gene expression, and copy number alterations, helping researchers study molecular subtypes, biomarkers, and treatment responses in breast cancer.	The dataset is available at https://ega-archive.org/studies/EGAS00000000083 (accessed on 13 April 2025)	CSV
[46]	Mammographic Mass Dataset (MM-Dataset)	The Mammographic Mass dataset contains 961 cases (516 benign, 445 malignant) with BI-RADS scores (1–5) and patient age. It helps evaluate CAD systems in reducing unnecessary biopsies ( 70% false positives) by predicting lesion severity from mammograms.	The dataset is available at https://archive.ics.uci.edu/dataset/161/mammographic+mass (accessed on 13 April 2025)	CSV

5. Machine Learning Techniques

The three main categories of machine learning are supervised learning, unsupervised learning, and reinforcement learning. Supervised learning trains models using labeled data to map inputs to known outputs, making it ideal for predictive tasks like classification and regression. The process involves analyzing training data, computing statistical features, and establishing input–output relationships. While effective for accurate predictions, it requires large labeled datasets and significant training resources. Unsupervised learning discovers hidden patterns in unlabeled data through techniques like clustering and association rule mining. Valuable for exploratory analysis, it requires less preprocessing than supervised methods but can yield more challenging-to-interpret results. Reinforcement learning trains autonomous agents through environmental interaction and reward feedback, excelling in dynamic domains like robotics and game AI. This powerful approach demands substantial computation and careful reward mechanism design. Table 3 summarizes the key characteristics of machine learning algorithms, including their types, categorizations, approaches, advantages, and limitations.

Table 3. Machine learning algorithm taxonomy (type, categorization, approach, advantages, limitations).

Type & Ref.	Categorization	Approach	Advantage	Limitation
Supervised Learning
Classification & [47]	Discrete Output	Assigns input to predefined labels	Effective for clear categories; works well with algorithms like SVM, Decision Trees	Struggles with ambiguous cases; requires balanced datasets
Binary Classification & [47]	Two-class output (Yes/No)	Simple models like Logistic Regression	Easy to implement; computationally efficient	Cannot handle multi-class problems
Multiclass Classification & [48]	Multiple classes (>2)	Algorithms like Random Forest, Neural Networks.	Solves complex real-world problems with many categories	Requires more data and computational power
Multilabel Classification & [48]	Single instance → Multiple labels	Specialized algorithms like ML-kNN	Useful for text tagging, medical diagnosis	Complex evaluation; harder to train
Regression & [49]	Continuous Output	Predicts numerical values	Provides interpretable results	Sensitive to outliers; assumes linear relationships
Linear Regression & [49]	Linear modeling	Fits a straight line to data	Simple and fast	Poor performance on non-linear data
Non-linear Regression & [49]	Non-linear modeling (polynomials, kernels)	Captures complex patterns	Flexible and powerful for curved trends	Prone to overfitting; harder to tune
Ordinal Regression & [49]	Ordered categories (e.g., ratings 1–5)	Treats labels as ranked	Preserves order information	Difficult to model unequal intervals between ranks
Time Series Forecasting & [50]	Sequential data prediction	Models like ARIMA, LSTM	Critical for stock prices, weather forecasting	Sensitive to missing data; requires careful feature engineering
Ensemble Methods & [51]	Combines multiple models (e.g., Bagging, Boosting)	Random Forest, XGBoost	Reduces variance; improves accuracy	Computationally intensive; less interpretable
Stacking & [51]	Meta-Ensemble	Combines multiple models via a meta-learner (e.g., predictions of base models become input to a final model)	Often outperforms single models; leverages strengths of diverse algorithms	Computationally expensive; risk of overfitting if not properly tuned
Neural Networks & [52]	Deep Learning (CNN, RNN)	Learns hierarchical features	State-of-the-art for images, text, and complex data	Requires huge datasets; “black box” nature
Naïve Bayes & [53]	Probabilistic (Gaussian, Multinomial)	Based on Bayes’ theorem	Fast training; works well with small data	Assumes feature independence (often untrue)
Decision Trees & [54]	Predictive (Discrete/Continuous)	Rule-based splits	Easy to visualize and interpret	Prone to overfitting; unstable with small data changes
Unsupervised Learning
K-means & [55]	Clustering	Groups data into k clusters	Simple and scalable	Sensitive to outliers; requires predefined k
PCA (Principal Component Analysis) & [56]	Dimensionality Reduction	Transforms data into uncorrelated components	Reduces noise and overfitting	Loses interpretability of original features
Gaussian Mixture Models (GMM) & [57]	Probabilistic clustering	Soft clustering (data points can belong to multiple clusters)	Flexible cluster shapes	Slow convergence; assumes Gaussian distribution
Hidden Markov Model (HMM) & [58]	Sequential data modeling	Models time-dependent processes	Powerful for speech recognition, genomics	Complex parameter tuning
Reinforcement Learning
Q-Learning & [59]	Model-Free	Learns action–value function (Q-table)	Works in environments with unknown dynamics	Struggles with large state spaces
Deep Q Network (DQN) & [60]	Combines Q-Learning + Deep Learning	Uses neural networks to approximate Q-values	Handles high-dimensional states	Requires massive computational resources
Policy Gradient & [61]	Direct policy optimization	Updates policy parameters via gradient ascent	Can handle continuous action spaces	High variance; sample-inefficient

6. Cancer Statistics and Incidence and MortalityRates

This section presents a statistical analysis of data for cancer rankings from 1990 to 2021. We highlight the mortality rates for both males and females, along with the global death rate per 100,000 individuals. Additionally, the section includes the incidence and death rates for both sexes categorized by age groups (20+ years and under 20 years) specifically for Bangladesh. Table 4 presents the year-wise rankings of six major cancers—Breast, Lung, Stomach, Cervical, Colorectal, and Brain—from 1990 to 2021, obtained from the Institute for Health Metrics and Evaluation (IHME) [62]. In the table, lower values indicate a higher impact and higher values indicate a lower impact. The rankings highlight significant changes in the incidence of these cancers over three decades. Breast Cancer ranked third in 1990 and 2000, worsened to become the most impactful cancer by 2011, and it had retained its top position by 2021, underscoring its increasing global prevalence and burden. Lung Cancer consistently ranked among the top two, taking the first position in 2000 and 2010 before being overtaken by Breast Cancer. Stomach Cancer, initially the most impactful cancer (ranked first) in 1990, had declined to fifth by 2021, reflecting advancements in prevention and treatment. Cervical Cancer remained stable at fourth until 2011 before worsening to third in 2021, while Brain Cancer consistently held the sixth position, indicating a lower comparative global impact. Meanwhile, Colorectal Cancer worsened from fifth in 1990 to fourth by 2021, reflecting its global increase. These patterns illustrate the evolving global cancer landscape, with BC emerging as the most impactful over time, particularly during e period 2011 to 2021. The worsening of BC highlights its growing prevalence and significance, underscoring the urgent need for enhanced research, early detection, and effective treatment strategies.

Table 4. Year-wise ranking of the six top cancers in the world [62].

Year	Breast	Lung	Stomach	Cervical	Colorectal	Brain
1990	3	2	1	4	5	6
2000	3	1	2	4	5	6
2010	2	1	3	4	5	6
2011	1	2	3	4	5	6
2015	1	2	3	4	5	6
2018	1	2	3	4	5	6
2021	1	2	5	3	4	6

Figure 4 shows the death rates for males, and Figure 5 depicts the death rates for females per 100,000 individuals aged 15–49 between 1990 and 2021, using data obtained from the Institute for IHME [62]. The data reveal that the average death rate for females was 6.6401, while for males this was 0.0829. This indicates that the female death rate for BC is 98.75% higher than that of males during this period. Figure 6 illustrates the global death rates per 100,000 individuals. The data demonstrate an increase from 2.9141 in 1990 to 3.3183 in 2021, marking a 12.181% increase over this period using data provided by the same institute.

Figure 4. Male death rates (1990–2021) due to BC [62].

Figure 5. Female death rates (1990–2021) due to BC [62].

Figure 6. Global death rates per 100,000 individuals (1990–2021) [62].

Figure 7 and Figure 8 depict the death rates and incidence rates of BC in Bangladesh from 1980 to 2021 categorized by age groups (20+ years and under 20 years) using data provided by the Institute for IHME, Seattle, United States [63]. Both figures clearly show a consistent increase in the mortality and incidence rates over time.

Figure 7. Breast cancer death rate in Bangladesh (1980–2021) [63].

Figure 8. Incidence rates of breast cancer in Bangladesh (1980–2021) [63].

The integration of ML and DL technologies has emerged as a transformative approach, enabling rapid and more accurate diagnosis. Research has shown that ML techniques such as CNNs can achieve up to 90% accuracy in classifying mammographic images as benign, malignant, or normal. Similarly, artificial neural networks have demonstrated high efficiency in analyzing ultrasound images to distinguish between benign and malignant cases. Some studies have reported that ML techniques can improve the prediction accuracy of BC recurrence by up to 25% compared to traditional methods [24,64,65]. These advancements in ML and DL offer the potential to address the limitations of traditional diagnostics by automating image interpretation, enhancing predictive analytics, and tailoring individual treatment plans. This review summarizes commonly used ML and DL algorithms and methodologies used in BC research, providing insights into their applications, strengths, and prospects.

7. Breast Cancer Diagnosis and Classification

Determining the type and stage of breast cancer is crucial for patients, as this directly impacts treatment planning and prognosis, helping physicians develop optimal strategies and predict the chance of survival. Figure 9 illustrates the types of BC diagnosis, spanning multiple categories. Biopsy methods such as fine needle aspiration (FNA) for extracting fluid or cells, core needle biopsy for removing small tissue samples, Surgical Biopsy for excising part or all of a suspicious mass, and vacuum-assisted biopsy for obtaining larger samples play key roles in diagnosis. Imaging techniques include mammography, an X-ray of the breast; ultrasound, which uses sound waves to differentiate masses; MRI, a technology that provides detailed images of soft tissues, and tomosynthesis, a 3D mammography employed for enhanced clarity. Pathological analysis involves histopathology to examine tissue, immunohistochemistry (IHC) to identify protein markers such as HER2 or hormone receptors, and molecular testing to detect genetic mutations or biomarkers. Clinical evaluation includes a physical examination to check for abnormalities and a review of the patient’s medical history, while lab reports involve tests such as tumor markers (e.g., CA 15-3, CEA) for cancer indicators and a complete blood count (CBC) to assess general health. These methods provide a comprehensive framework for accurate identification and evaluation [2,66].

Figure 9. Overview of breast cancer diagnostic methods.

7.1. Breast Cancer Variants

Breast cancer can be categorized according to the location, biological characteristics, and molecular subtype. Non-invasive types such as Ductal Carcinoma In Situ (DCIS) and Lobular Carcinoma In Situ (LCIS) are confined to their point of origin and serve as warning signs for potential invasive cancers. Invasive types include Invasive Ductal Carcinoma (IDC), the most common form, and Invasive Lobular Carcinoma (ILC) that often responds to hormone therapy. Rare types include Inflammatory Breast Cancer, a highly aggressive form that causes redness and swelling, and Paget’s Disease that affects the nipple and areola. Breast cancer is also classified into molecular subtypes based on the hormone receptor (ER, PR) and HER2 status; these include Luminal A, Luminal B, HER2-enriched, and Triple-Negative, and these subtypes can be used to guide treatment strategies. Assignment of stages from 0 to 4 and grading further assess the progression and behavior of the cancer, with stages indicating the degree of spread and grades reflecting growth patterns. Understanding the meanings of the classifications is crucial for diagnosis and treatment planning. Figure 10 displays the different types of breast cancer.

Figure 10. Types of breast cancer.

7.2. Key Risk Factors Associated with Breast Cancer

Breast cancer is influenced by several non-modifiable risk factors, including gender, age, family history, genetics, hormonal factors, and personal history, with women being at a significantly higher risk, particularly after the age of 50. A family history of BC, especially in first-degree relatives, along with known genetic mutations such as BRCA1 and BRCA2, further increases the likelihood of developing the disease. Hormonal factors such as early menstruation, late menopause, or hormone replacement therapy also contribute to this risk. Additionally, individuals who have had BC in one breast are at higher risk. Beyond these, other risk factors including tobacco use (smoking, chewing tobacco, and secondhand smoke), poor dietary habits (low intake of fruits, vegetables, whole grains, and seafood omega-3 fatty acids, along with a high intake of red meat, processed meat, sugar-sweetened beverages, and trans fats), and metabolic factors (high levels of fasting plasma glucose and LDL cholesterol, high blood pressure and body mass index, low bone mineral density, and kidney dysfunction) also play significant roles in the development of BC. Figure 11 illustrates the statistics of BC deaths linked to various risk factors, showing that, from 1990 to 2021, tobacco contributed to 3.92% of female and 0.88% of male deaths; dietary factors accounted for 11.87% of male and 12.62% of female deaths, while metabolic risk factors were responsible for 10.48% of female deaths but none for males among all BC patients [62,63].

Figure 11. Global breast cancer mortality by risk factor (1990–2021) [63].

Figure 12 presents the seven key factors associated with the occurrence of breast cancer in females of all ages in 2021. The figure highlights that high consumption of red meat was the leading factor associated with BC, followed by high body mass index and high fasting plasma glucose as the second and third most significant factors, respectively. Furthermore, the figure shows that other critical factors contributing to BC risk include high alcohol use, low physical activity, smoking, and exposure to secondhand smoke. Therefore, avoiding these risk factors could significantly reduce the likelihood of developing BC in females.

Figure 12. Key afctors contributing to breast cancer risk in females in 2021 [62].

Beyond hormonal and lifestyle factors, genetic mutations significantly influence breast cancer risk. BRCA1/2 mutations impair DNA repair, conferring up to 65% lifetime risk and association with triple-negative subtypes. TP53 and PTEN mutations dysregulate cell-cycle control and PI3K signaling, respectively, while PROC variants may promote metastatic survival via coagulation pathways [67,68].

8. Breast Cancer Detection Approaches

The diagnosis of breast cancer employs both traditional methods, such as imaging (mammograms, ultrasounds, and MRIs) and biopsies, as well as advanced computational techniques like machine learning and deep learning (DL). Deep learning, particularly methods using convolutional neural networks (CNNs), has shown exceptional performance in analyzing medical images and tissue samples. Emerging approaches, including liquid biopsies, radiomics, and AI-powered multimodal systems, are further transforming the accuracy and efficiency of BC diagnosis.

8.1. Mammogram Image-Based Breast Cancer Detection

Several publicly available datasets are widely used for mammogram image-based BC detection, including the Curated Breast Imaging Subset of DDSM (CBIS-DDSM) [69], the Digital Database for Screening Mammography (DDSM) [70], the Mammographic Image Analysis Society (MIAS) dataset [71], and the INbreast dataset [30]. Figure 13 shows a sample mammogram image with a marked areaindicated by the red square box on the right side highlighting a potential abnormality for further analysis.

Figure 13. A sample mammogram image [72].

The approach proposed by Zhao et al. involved identifying eligible patients, preparing image data, extracting labels and BI-RADS scores, and applying AlexNet, VGG, ResNet, and DenseNet models, with DenseNet achieving the best performance. The evaluation was conducted under three settings: General (BI-RADS 1–5), False Positive Reduction (BI-RADS 4–5), and Difficult Cases (BI-RADS 0) [73]. Jamil et al. proposed a BC detection technique using mammography images, achieving an accuracy of 0.9719. Their approach applied three techniques in sequence: Log Ratio, Gabor Filter, and Fuzzy C-Means (FCM) [74]. Nemade et al. proposed an ensemble-based model for BC classification, utilizing VGG16, InceptionV3, and VGG19 as base learners, with an artificial neural network for classification. The model achieved an accuracy of 98.10%, a specificity of 99.12%, and a sensitivity of 97.01% [75].

Another study conducted five experiments using different preprocessing techniques. The researchers applied six classifiers—SVM, Random Forest, ANN, KNN, Naive Bayes, and Decision Tree—and achieved 100% for all metrics (accuracy, sensitivity, specificity, and F1-score) with SVM and ANN. This was achieved when preprocessing was performed using contrast limited adaptive histogram equalization (CLAHE) and unsharp masking (USM) on the mini-MIAS dataset [76].

Yaqub et al. developed a two-stage system for BC diagnosis using mammogram images, where the first stage involves image segmentation through the Atrous Convolution-based Attentive and Adaptive Trans-Res-UNet (ACA-ATRUNet) model, and the second stage utilizes the Atrous Convolution-based Attentive and Adaptive Multi-scale DenseNet (ACA-AMDN) model for cancer detection, with hyperparameter optimization carried out using the Modified Mussel Length-based Eurasian Oystercatcher Optimization (MML-EOO) algorithm. They achieved an accuracy of 89.13%, a specificity of 0.8937, a precision of 0.8226, and an F1-score of 0.8536 for the MIAS mammography dataset, and accuracy, precision, and F1-score values of 89.061%, 80.348%, and 84.424%, respectively, for the CBIS-DDSM BC image dataset [77]. In [78], EfficientNet achieved the highest accuracy of 0.9829 on the MaMaTT2 dataset. For the DDSM dataset, a CNN classifier with different fine-tuning achieved an accuracy of 0.9996, a sensitivity of 1.0, a precision of 0.9992, and an F1-score of 0.9996 [79]. Additionally, the Inception Recurrent Residual Convolutional Neural Network (IRRCNN) achieved an accuracy of 0.9676 on the BreakHis dataset (factor ×40) and 0.9711 on the BC Classification Challenge 2015 dataset [10].

Table 5 summarizes recently published research on mammography image-based studies.

Table 5. Comparison of recent mammography-based breast cancer detection techniques.

Ref.	Year	Dataset	Image Size	Proposed Model	Proposed Approach	Results
[77]	2024	MIAS mammography & CBIS-DDSM	1024 × 1024 & Non-fixed	MML-EOO-ACA-ATRUNet-MDN	Two-stage system: ACA-ATRUNet for segmentation, ACA-AMDN for cancer detection, hyperparameter optimization with MML-EOO.	MIAS mammography Accuracy: 0.8913 Precision: 0.8226 Specificity: 0.8937 F1-Score: 0.8536 CBIS-DDSM Accuracy: 0.8906 Precision: 0.8035 F1-Score: 0.8442
[78]	2024	MaMaTT2	224 × 224	EfficientNet	Data augmentation, EfficientNet.	Accuracy: 0.9829
[74]	2024	MIAS breast cancer mammography images	1024 × 1024	Mammogram Breast Cancer Image Change Detection (MBCICD)	Log Ratio + Gabor Filter + Fuzzy c-means (FCM)	Accuracy: 0.9719
[75]	2024	DDSM	$299\times 299$	Ensemble Model 2	Ensemble model with VGG16, InceptionV3, VGG19 as base learners, and ANN for classification.	Accuracy: 0.9810 Sensitivity: 0.9701 Specificity: 0.9912
[76]	2023	mini-MIAS	1024 × 1024	SVM, ANN	CLAHE and USM preprocessing, SVM and ANN classifiers.	Performance metrics: Accuracy, Sensitivity, Specificity, and F1-Score all = 1.0
[79]	2022	DDSM	299 × 299	CNN	CNN classifier with different fine-tuning.	Accuracy: 0.9996 Sensitivity: 1.0 Precision: 0.9992 F1-Score: 0.9996
[73]	2022	Digital mammography	1024 × 832	DenseNet-121	Dataset preparation, label extraction, BI-RADS score extraction, model training with DenseNet-121, and performance evaluation were performed.	Setting BI-RADS score 1–5 CV AUC: 0.953 Test AUC: 0.94 Specificity: 0.85 Sensitivity: 0.9 Setting BI-RADS score 4–5 CV AUC: 0.932 Test AUC: 0.927 Setting BI-RADS score 0 CV AUC: 0.758 Test AUC: 0.787
[10]	2019	BreakHis & BC Classification Challenge 2015	×460 (RGB) & 2040 × 1536	IRRCNN + Aug	Data augmentation, Inception Recurrent Residual Convolutional Neural Network (IRRCNN).	BreakHis dataset Accuracy: 0.9676 ± 0.0111 (Factor × 40) BC Classification Challenge 2015 Accuracy: 0.9711 Sensitivity: 0.9371 Specificity: 0.9809 AUC: 0.9905

Figure 14 shows the highest accuracies obtained by different models applied to mammogram datasets in 2023 and 2024, including CBIS-DDSM [80], CMMD [80], DDSM [81], INbreast [82], MIAS [81], mini-MIAS [76], Breast Cancer Wisconsin [83], Categorized Digital CESM [84], Oncology Hospital Ho Chi Minh [85], Cropped DDSM [86], MaMaTT2 [78], RSNA [87], and BNS [88], as reported in the articles, while the accompanying correlation figure uses color intensity to represent accuracy levels.

Figure 14. Highest detection accuracy across multiple mammogram datasets for breast cancer.

8.1.1. Model Performance on Mammogram Image: Pros, Cons, and Future Directions

Table 6 provides an overview of the key benefits, limitations, and future research directions for leading mammogram-based breast cancer detection models, enabling easy comparison of their capabilities and areas for improvement.

Table 6. Comprehensive overview of top-performing models on mammogram breast cancer datasets.

Ref.	Advantages	Limitations	Future Work Recommendations
[78]	EfficientNet outperformed CNN and Inception models on the MaMaTT2 dataset, achieving 98.29% overall accuracy (99.27% benign, 98.46% malign, 97.13% normal) using an augmented dataset of 22,500 images.	The original dataset was small (408 images), which may affect model generalizability. Legal and privacy concerns limit access to large, real-world medical datasets needed for broader validation.	Future work should expand datasets and incorporate more pre-trained models to boost accuracy. Overcoming data privacy challenges will be key for clinical implementation.
[74]	The method achieved 97.19–97.38% PCC on MIAS, outperforming FDA-RMG (94.81%) using Mean/Log Ratio, Gabor Filter, and FCM clustering.	Reliance on limited MIAS images may reduce real-world applicability, while high false negatives (93.31%) suggest missed subtle changes.	Larger real-patient datasets and improved algorithms could reduce false negatives, while exploring advanced feature extraction may boost accuracy further.
[75]	Ensemble Model 2 (neural net) achieved 98.10% accuracy on DDSM/CBIS-DDSM, outperforming InceptionV3 (95.71%). Combining VGG16/InceptionV3/VGG19 boosted specificity (99.12%) and sensitivity (97.01%).	The study relied on pre-curated, fixed-size images, limiting generalizability to diverse clinical settings. High computational resources were required for training ensemble models, which may hinder real-time deployment in resource-constrained environments.	Expanding validation to diverse datasets (e.g., INBreast or MIAS) could enhance robustness. Exploring lightweight meta-learners (e.g., SVM or Random Forest) may improve efficiency without compromising accuracy.
[79]	The proposed modified ResNet50 model, trained on the DDSM dataset with 13,128 images (5970 benign and 7158 malignant), achieved excellent classification performance, with test accuracy up to 99.96%, AUC of 1.0000, and F1-score of 0.9996 using SGD and 100% fine-tuning.	The training time for a single model was around 3.7 h, indicating high computational cost. Additionally, selecting optimal hyperparameters remains a challenge due to the exhaustive trial-and-error process.	Future research should focus on reducing computational time and developing methods for automated hyperparameter tuning to enhance efficiency and scalability.

8.1.2. Summary

Several publicly available datasets, such as CBIS-DDSM, DDSM, MIAS, and INbreast, have been widely used for breast cancer (BC) detection from mammogram images. Recent studies have employed advanced machine learning and deep learning techniques, including DenseNet, VGG16, AlexNet, and InceptionV3, to achieve high accuracy and specificity in BC detection. Preprocessing methods like contrast limited adaptive histogram equalization (CLAHE) and unsharp masking (USM) have been utilized for image enhancement, while ensemble models and hybrid systems combining deep learning with traditional image processing techniques have also shown promising results. The highest reported accuracy reached 99.96% on the DDSM dataset. Table 7 presents a summary of key findings in mammogram-based breast cancer detection.

Table 7. Summary of key findings in mammogram-based BC detection.

Aspect	Details
Datasets	CBIS-DDSM, DDSM, MIAS, INbreast, mini-MIAS, MaMaTT2, BreakHis, etc.
Models	DenseNet, VGG16, AlexNet, InceptionV3, EfficientNet, IRRCNN, etc.
Preprocessing	CLAHE, USM, Log Ratio, Gabor Filter, Fuzzy C-Means (FCM)
Highest Accuracy	99.96% (CNN with fine-tuning on DDSM)
Key Techniques	Ensemble models, hybrid systems, hyperparameter optimization

8.1.3. Analysis

The application of deep learning models, particularly CNNs and ensemble models, has significantly improved BC detection performance. Preprocessing techniques like CLAHE and USM, combined with classifiers such as SVM and ANN, have enabled researchers to achieve near-perfect metrics. Advanced architectures like DenseNet and EfficientNet have demonstrated robust performance across multiple datasets, including DDSM, MIAS, and CBIS-DDSM. Hybrid systems, such as those combining deep learning with traditional image processing techniques (e.g., Log Ratio, Gabor Filter, and FCM), have also shown remarkable accuracy. Table 8 provides an analysis of techniques and their impact based on mammogram-based BC detection.

Table 8. Analysis of techniques and their impact based on mammogram-based BC detection.

Technique	Impact
Deep Learning Models	High accuracy and specificity; handles complex patterns in mammograms.
Preprocessing	Enhances image quality, leading to improved model performance.
Ensemble Models	Combines strengths of multiple models, improving robustness and accuracy.
Hybrid Systems	Integrates traditional and deep learning methods for better performance.
Hyperparameter Optimization	Improves model tuning, leading to higher accuracy and efficiency.

8.1.4. Critique

While current models demonstrate impressive performance, several limitations need addressing. First, the interpretability of these models remains a challenge, which is critical for clinical acceptance. Second, the generalization ability of models across diverse datasets and imaging conditions is often untested, raising concerns about their real-world applicability. Additionally, the lack of diversity in datasets (e.g., limited representation of different populations and imaging devices) may lead to overfitting and reduced robustness. Table 9 presents a critique of current limitations.

Table 9. Critique of current limitations based on mammogram-based BC detection.

Limitation	Details
Interpretability	Lack of explainability in model predictions hinders clinical adoption.
Generalization	Models may not perform well on datasets from different populations or devices.
Dataset Diversity	Limited diversity in datasets may lead to overfitting.
Real-World Applicability	Performance in clinical settings may differ from experimental results.

8.1.5. Comparison

The results of various studies highlight differing approaches and outcomes. For instance, Yaqub et al. (2024) [77] achieved 89.13% accuracy on the MIAS dataset using a two-stage system, while Nemade et al. (2024) [75] reported 98.10% accuracy with an ensemble model. Avci et al. (2023) [76] demonstrated the importance of preprocessing, achieving 100% accuracy with CLAHE, USM, SVM, and ANN. EfficientNet, as used by Dada et al. (2024) [78], outperformed other models with 98.29% accuracy on the MaMaTT2 dataset. These comparisons underscore the importance of dataset quality, preprocessing, and algorithm choice in achieving high detection accuracy. Table 10 provides a comparison of key studies.

Table 10. Comparison of key studies based on mammogram-based BC detection.

Study	Dataset	Model/Technique	Accuracy	Key Findings
Yaqub et al. (2024) [77]	MIAS, CBIS-DDSM	ACA-ATRUNet + ACA-AMDN + MML-EOO	89.13%	Two-stage system with hyperparameter optimization.
Nemade et al. (2024) [75]	DDSM	Ensemble (VGG16, InceptionV3, VGG19) + ANN	98.10%	High specificity and sensitivity.
Avci et al. (2023) [76]	mini-MIAS	CLAHE + USM + SVM/ANN	100%	Preprocessing significantly improves metrics.
Dada et al. (2024) [78]	MaMaTT2	EfficientNet	98.29%	Outperforms other models on MaMaTT2.
Mudeng et al. (2022) [79]	DDSM	CNN with fine-tuning	99.96%	Highest reported accuracy on DDSM.

The advancements in mammogram-based BC detection are promising, with deep learning models and preprocessing techniques driving significant improvements in accuracy and performance. However, challenges related to interpretability, generalization, and dataset diversity must be addressed to ensure these models are clinically viable and applicable across diverse populations and imaging conditions. Future research should focus on developing more interpretable models, expanding dataset diversity, and testing models in real-world clinical settings.

While mammography models achieve >99% accuracy on curated datasets (Table 5), two critical gaps persist: (1) performance drops significantly on dense breast tissue, (2) most studies lack real-world testing with radiologist workflows.

8.2. Ultrasound Image-Based Breast Cancer Detection

Ultrasound images are used to visualize internal structures through high-frequency sound waves. There are several publicly available datasets for ultrasound image-based BC detection, including the Breast Ultrasound Images (BUSI) dataset [34], BUS-BRA [89], BUS-UCLM [90], and Breast-Lesions-USG [91]. Figure 15 presents a sample ultrasound image with an array-marked region, highlighting a potential abnormality for further examination.

Figure 15. A sample ultrasound image [92].

The BUSI dataset was created in 2018 and includes 780 images from 600 female patients aged 25–75 categorized as normal, benign, or malignant, with PNG images averaging $500\times 500$ pixels and ground truth masks. The BUS-BRA dataset features 1875 anonymized images from 1064 patients in Brazil, including biopsy-proven cases (722 benign and 342 malignant), BI-RADS assessments, manual segmentations, and cross-validation partitions to standardize CAD system evaluations. The BUS-UCLM dataset, collected between 2022 and 2023, contains 683 images with RGB segmentation masks for normal, benign, and malignant cases, enabling the development of machine learning models for the detection and diagnosis of breast lesions. Lastly, the Breast-Lesions-USG dataset offers 256 scans with annotated tumors, BIRADS classifications, and histopathological diagnoses, serving as an external testing set and providing support for AI development in BC research.

Mehedi et al. introduced a DL strategy for BC detection using ultrasound images, based on a publicly available breast ultrasound image dataset from Rodrigues [93] that includes 250 images (100 benign and 150 malignant cases). They incorporated Grad-CAM and occlusion mapping to evaluate feature extraction and developed a custom CNN model with fewer parameters for improved efficiency. For classification, they applied DenseNet201, ResNet50, and VGG16 models, achieving 100% accuracy after fine-tuning using this dataset. Specifically, DenseNet201 and ResNet50 performed best with the Adam and RMSprop optimizers, while VGG16 reached 100% accuracy using Stochastic Gradient Descent. The original images with varying sizes (57 × 75 to 161 × 199) were resized to 224 × 224, 227 × 227, and 299 × 299 for different pre-trained models [94].

The DeepBreastCancerNet model consists of 24 layers, with six convolutional layers, nine inception modules, and a fully connected layer. The model achieved 99.35% accuracy on the initial ultrasound dataset [34] and 99.63% accuracy on a second publicly available dataset [93], demonstrating its robustness. By applying transfer learning, a method that utilizes pre-trained models to avoid overfitting, the researchers improved detection accuracy, surpassing state-of-the-art models in BC classification. The model incorporates clipped ReLU and leaky ReLU activation functions, along with batch and cross-channel normalization, that enhanced its performance on both datasets [95].

A study of 603 patients from three institutions (2018–2021) trained and validated four DCNNs on ultrasound images (420 training, 183 validation) to predict the responses to NAC. ResNet50, the best model using only images, achieved an AUC of 0.879 and 82.5% accuracy. When combined with clinical–pathologic variables, the integrated DLR model outperformed both the image-based and clinical models, achieving AUCs of 0.962 and 0.939 in training and validation, surpassing radiologists’ predictions and improving their accuracy as a support tool [96]. The study introduces a novel ROI-free system for BC diagnosis using ultrasound images that incorporates prior anatomical knowledge of the spatial relationships between malignant and benign tumors, captured through the HoVer-Transformer. This model, which extracts inter- and intra-layer spatial information both horizontally and vertically, was evaluated on the GDPH&SYSUCC dataset and compared to four CNN-based and three vision transformer models. The proposed model achieved state-of-the-art results, with an AUC of 0.924, an accuracy of 0.893, a specificity of 0.836, and a sensitivity of 0.926, surpassing two senior sonographers in BC diagnosis [97].

Madhusudan et al. presented a method for BC detection using ultrasound images (USIs) [98], combining transfer learning (TL) models (MobileNetV2, ResNet50, VGG16) with LSTM for feature extraction and SMOTETomek for feature balancing. The method with VGG16 achieved an F1 score of 99.0%, MCC and kappa coefficients of 98.9%, and an AUC of 1.0. Using K-fold cross-validation, the same method achieved an average F1-score of 96%. Grad-CAM and LIME were applied for visualization and interpretability, and confidence intervals were calculated via NAI (LCI 96.50%, UCI 99.75%, MCI 98.13%) and bootstrapping (LCI 93.81%, UCI 96.00%, MCI 94.90%) [99]. Rao et al. proposed Inception V3 + Stacking, a model that combined transfer learning with ensemble stacking of ML models (including MLP and SVM with RBF and polynomial kernels) and applied the model to the ultrasound breast cancer (USBC) image dataset. Their model achieved excellent results, with an AUC of 0.947 and an accuracy of 0.858, outperforming existing diagnostic systems [100].

Appiah et al. explored the use of DL for classifying BC from ultrasound images, focusing on low- and middle-income countries (LMIC) with limited healthcare access. Their aim was to deploy a simple classifier on a mobile device using affordable handheld ultrasound systems to identify cases needing medical attention. Their model achieved 64% accuracy when trained from scratch and 78% accuracy after pre-training, highlighting DL’s potential to aid early BC detection with minimal image data [101]. Table 11 provides a summary of recent research on ultrasound image-based studies.

Table 11. Comparison of recent ultrasound-based breast cancer detection techniques.

Ref.	Year	Dataset	Image Size	Proposed Model	Proposed Approach	Results
[94]	2021	Breast Ultrasound Image	57 × 57 & 161 × 199	DenseNet201, ResNet50, VGG16	Grad-CAM, occlusion mapping for feature extraction, DenseNet201, ResNet50 (Adam, RMSprop optimizers), VGG16 (Stochastic Gradient Descent).	Accuracy: 100%
[95]	2023	Initial Dataset [34] & Second Dataset [93]	500 × 500 & Non-Fixed	DeepBreastCancerNet	DeepBreastCancerNet model, 24 layers, 6 convolutional layers, 9 inception modules, ReLU, batch normalization	Initial Dataset Accuracy: 0.9935 Second Dataset Accuracy: 0.9963 Precision: 0.9950 Recall: 1.0 F1-Score: 0.9950
[96]	2023	Ultrasound dataset collected from three institutions	224 × 224	integrated DLR model	Developed an integrated DLR model with clinical–pathological variables	AUC: 0.962 (training) And 0.939 (validation)
[97]	2023	GDPH& SYSUCC	844 × 627	HoVer-Trans	An ROI-free model with HoVer-Transformer to capture tumor spatial relationships for breast cancer diagnosis	AUC: 0.924 Accuracy: 0.893 Specificity: 0.836 Sensitivity: 0.926
[99]	2024	Breast Ultrasound Images Dataset	500 × 500	VGG16-LSTM	VGG16 with LSTM for feature extraction, SMOTETomek for feature balancing, and Grad-CAM for visualization and LIME for interpretability.	F1-score: 0.99 MCC & Kappa: 0.989 AUC: 1.0
[100]	2024	ultrasound breast cancer (USBC) image	500 × 500	Inception V3 + Stacking	Inception V3 transfer learning + ensemble stacking (MLP, SVM with RBF and polynomial kernels)	Accuracy: 0.858 AUC: 0.947 Precision: 0.8579 Recall: 0.858 F1-score: 0.857
[101]	2024	NYU dataset	665 × 603	deep learning classifier	Deep learning with transfer learning for breast cancer detection using ultrasound images on mobile devices.	Accuracy: 78%

Figure 16 presents the highest accuracy rates achieved by various models on ultrasound datasets in 2023 and 2024, including UBC Benchmark [100], BUSI [102], BUS2 [103], UDAIT [104], multiple hospital datasets, and others [95,104,105,106,107,108,109,110,111,112,113,114], as outlined in the referenced studies, with the corresponding correlation figure using color intensity to depict accuracy levels.

Figure 16. Highest detection accuracy across multiple ultrasound datasets for breast cancer.

8.2.1. Model Performance on Ultrasound Image: Pros, Cons, and Future Directions

Table 12 summarizes the advantages, disadvantages, and future work recommendations for top-performing models in ultrasound-based breast cancer detection, facilitating direct comparison of their strengths and development needs.

Table 12. Comprehensive overview of top-performing models on ultrasound breast cancer datasets.

Ref.	Advantages	Limitations	Future Work Recommendations
[94]	The model proposed in the reference article applies to the Breast Ultrasound Image dataset, achieves 100% accuracy in breast cancer detection using efficient pre-trained CNNs and provides interpretable results through Grad-CAM visualizations.	Its small dataset of 250 images limits generalizability, while high computational needs may restrict clinical deployment.	Expanding datasets, testing multi-class classification, and conducting clinical validations will enhance real-world applicability.
[95]	DeepBreastCancerNet applies to the second dataset [93], achieves 99.63% accuracy using an efficient 24-layer CNN with specialized activation functions, outperforming existing models while maintaining computational efficiency.	Limited by small dataset size (1030 images) and lack of clinical validation, with potential deployment challenges in resource-constrained environments.	Expand datasets through multi-center collaborations and optimize the model for clinical deployment, followed by rigorous real-world testing.
[107]	The proposed model applies to the dataset obtained from Assosa General Hospital, integrating a feature enhancement technique (LPP) into the Google Inception network, achieving high accuracy (99.80%) for breast cancer detection and classification. It also employs transfer learning to address dataset limitations, improving performance on ultrasound images.	The model was evaluated on a limited dataset of 1578 images, which may affect generalizability to diverse populations. Additionally, the computational complexity of LPP and the Inception network could pose challenges for real-time deployment in clinical settings.	Future work should validate the model on larger and more diverse datasets to enhance robustness. Incorporating domain knowledge from radiologists could further improve interpretability and clinical relevance.
[109]	The proposed DL-AL hybrid model significantly outperformed 14 state-of-the-art algorithms on 1264 ultrasound images, achieving up to 96% Dice coefficient and 0.26% H3 distance on simple cases. It also maintained strong generalization, achieving an accurate segmentation ratio (ASR) of 98.40% even for the most complex cases ($H\left|H\right|H$), and handled 68% unseen data with a Dice score of 0.80.	The model relies heavily on manually engineered artificial life rules and genetic algorithms, which can increase design complexity and tuning effort. Moreover, the computational cost is relatively high, requiring up to 1.5 min and 0.5–1 million agents for segmentation on standard PCs, which may not be practical in time-constrained clinical workflows.	Future work should focus on optimizing the algorithm’s computational efficiency to ensure real-time deployment, especially by improving agent dynamics and parallelization. Expanding the model to include multi-modal imaging and evaluating it on larger, more diverse datasets could further enhance robustness and clinical applicability.

8.2.2. Summary

Ultrasound imaging is a widely used modality for breast cancer (BC) detection, supported by publicly available datasets such as BUSI, BUS-BRA, BUS-UCLM, and Breast-Lesions-USG. Recent studies have employed advanced deep learning (DL) models, including CNNs, transfer learning, and hybrid approaches, to achieve high accuracy and AUC scores. Notable models like DeepBreastCancerNet, HoVer-Transformer, and VGG16-LSTM have demonstrated exceptional performance. Additionally, interpretability techniques such as Grad-CAM and LIME have been integrated to enhance clinical applicability. Despite these advancements, challenges remain in standardizing image preprocessing, improving dataset diversity, and enabling real-world deployment. Table 13 presents a summary of key findings in ultrasound-based breast cancer detection.

Table 13. Summary of key findings in ultrasound-based BC detection.

Aspect	Details
Datasets	BUSI, BUS-BRA, BUS-UCLM, Breast-Lesions-USG, NYU, GDPH&SYSUCC, etc.
Models	DeepBreastCancerNet, HoVer-Transformer, VGG16-LSTM, ResNet50, DenseNet201
Techniques	Transfer learning, ensemble stacking, interpretability (Grad-CAM, LIME)
Highest Accuracy	100% (DenseNet201, ResNet50, VGG16 on Breast Ultrasound Image dataset)
Key Challenges	Dataset diversity, real-world deployment, interpretability, overfitting

8.2.3. Analysis

The studies reviewed demonstrate the effectiveness of DL models, particularly those leveraging transfer learning and ensemble techniques, in classifying BC from ultrasound images. Models like ResNet50, DenseNet201, and VGG16 achieve near-perfect accuracy when fine-tuned, with some studies reporting 100% accuracy. The integration of clinical–pathologic variables further enhances predictive performance, as seen in the integrated DLR model, which achieved an AUC of 0.962. However, dataset limitations, such as small sample sizes and inconsistent imaging protocols, hinder model generalizability. Additionally, the focus on real-time deployment in resource-limited settings highlights the need for computationally efficient models. Table 14 provides an analysis of techniques and their impact based on ultrasound-based breast cancer detection.

Table 14. Analysis of techniques and their impact in ultrasound-based BC detection.

Technique	Impact
Deep Learning Models	High accuracy and AUC scores; effective for complex pattern recognition.
Transfer Learning	Reduces overfitting and improves performance on small datasets.
Ensemble Models	Combines strengths of multiple models, improving robustness and accuracy.
Interpretability	Techniques like Grad-CAM and LIME enhance clinical trust and applicability.
Clinical Integration	Combining imaging data with clinical variables improves predictive power.

8.2.4. Critique

While the reported accuracies are impressive, several limitations warrant attention. First, the lack of cross-dataset validation raises concerns about model overfitting and generalizability. Many studies rely on a single dataset, which may not represent diverse populations or imaging conditions. Second, although interpretability techniques like Grad-CAM provide insights into model decisions, they are often insufficient for full clinical trust. Third, models achieving near-100% accuracy may indicate potential biases or dataset-specific overfitting, necessitating independent validation. Finally, the absence of standardized imaging protocols and the limited diversity of datasets hinder the development of robust models. Table 15 presents a critique of current limitations in ultrasound-based breast cancer detection.

Table 15. Critique of current limitations in ultrasound-based breast cancer detection.

Limitation	Details
Cross-Dataset Validation	Lack of testing on diverse datasets limits generalizability.
Interpretability	Grad-CAM and LIME provide insights but are insufficient for clinical trust.
Dataset Diversity	Limited diversity in datasets may lead to overfitting and reduced robustness.
Real-World Deployment	Challenges in deploying models in resource-limited settings remain unresolved.
Standardization	Inconsistent imaging protocols affect the model performance and applicability.

8.2.5. Comparison

The reviewed studies demonstrate varying approaches and outcomes, as summarized in Table 16. DeepBreastCancerNet [95] achieved the highest accuracy (99.63%) on a secondary dataset, followed by HoVer-Transformer (AUC 0.924) and VGG16-LSTM (AUC 1.0, F1-score 99%). The integrated DLR model outperformed image-only approaches by incorporating clinical features, achieving an AUC of 0.962. In contrast, simpler models designed for mobile devices, such as the one proposed by Appiah et al., achieved lower accuracy (78%), highlighting the trade-off between computational efficiency and predictive performance [100]. These comparisons underscore the importance of dataset quality, preprocessing, and algorithm choice in achieving high detection accuracy.

Table 16. Comparison of key studies in ultrasound-based breast cancer detection.

Study	Dataset	Model/Technique	Accuracy	Key Findings
Masud et al. (2021) [94]	Breast Ultrasound	DenseNet201, ResNet50, VGG16	100%	Fine-tuning with Adam, RMSprop, and SGD.
Raza et al. (2023) [95]	BUSI, Secondary	DeepBreastCancerNet	99.63%	Robust performance with ReLU and batch normalization.
Yu et al. (2023) [96]	Multi-Institution	Integrated DLR Model	AUC 0.962	Combined imaging and clinical variables.
Mo et al. (2023) [97]	GDPH&SYSUCC	HoVer-Transformer	AUC 0.924	ROI-free model with spatial relationship capture.
Lanjewar et al. (2024) [99]	BUSI	VGG16-LSTM	F1-score 99%	Feature balancing with SMOTETomek.
Rao et al. (2024) [100]	USBC	Inception V3 + Stacking	AUC 0.947	Ensemble stacking with MLP and SVM.
Appiah et al. (2024) [101]	NYU	Mobile DL Classifier	78%	Focus on low-resource settings.

Ultrasound-based BC detection has seen significant advancements through the application of deep learning models, transfer learning, and hybrid approaches. While models like DeepBreastCancerNet and HoVer-Transformer demonstrate exceptional performance, challenges related to dataset diversity, interpretability, and real-world deployment remain. Future research should focus on cross-dataset validation, standardized imaging protocols, and the development of computationally efficient models for resource-limited settings. Addressing these challenges will enhance the clinical applicability and robustness of ultrasound-based BC detection systems.

Current ultrasound DL models exhibit two key limitations: (1) overreliance on small, single-center datasets (Table 11) leading to poor generalization, (2) lack of standardized evaluation metrics for clinically crucial tasks like tumor boundary delineation.

8.3. Breast Cancer Detection Based on Magnetic Resonance Imaging (MRI) Images

Magnetic resonance imaging (MRI) is a safe, non-invasive technique that uses magnetic fields and radio waves to produce detailed internal body images. MRI is employed in the diagnosis and monitoring of conditions such as soft tissue abnormalities, tumors, and brain and spinal disorders. This section highlights AI-based methods for BC detection using MRI images [115]. Figure 17 displays a sample MRI image with a highlighted area, marked by the red arrow, indicating a potential abnormality for further examination.

Figure 17. A sample MRI image [92,116].

Zhang et al. proposed an approach that involved creating radiomics models for diagnosing BC using various MRI modalities (T1WI, T2WI, DWI, ADC, and DCE), with radiomic features extracted from plain, enhanced, and diffuse MRI scans and analyzed through DL for automatic extraction and classification. The diagnostic accuracy was significantly boosted by combining these modalities, with the highest diagnostic performance reflected in an AUC of 0.927 from the combined use of plain scan, enhanced, and diffuse sequences [117]. Liu et al. conducted an IRB-approved study utilizing a weakly supervised ResNet-101-based network for classifying malignant and benign images from a dataset of 278,685 image slices from 438 patients. The slices were grouped into 92,895 three-channel images, with 85% used for training and validation. The Adam optimizer with a SoftMax score threshold of 0.5 was used to train the model. The model achieved an AUC of 0.92 (±0.03), an accuracy of 94.2% (±3.4), a sensitivity of 74.4% (±8.5), and a specificity of 95.3% (±3.3). These results highlight the model’s ability to distinguish among various cancer types [118].

Yunan et al. developed a CNN-based computer-aided diagnostic tool for BC using DCE-MRI images, incorporating tumor geometric and pharmacokinetic features for classification. In a study of 130 patients, 71 were malignant cases and 59 were benign cases. The model achieved an accuracy, precision, sensitivity, and AUC of 87.7%, 91.2%, 86.1%, and 91.2% (±4.0%), respectively, across five-fold testing. Confidence in classification was reinforced by prediction probability, feature heatmaps, and dynamic scan time points, minimizing misclassification [119]. In a study by Yue et al., 516 BC patients were analyzed using a 3D UNet-based automatic segmentation model to extract 1316 radiomics features from regions of interest. Eighteen radiomics methods combining feature selection and classifiers were employed for model selection, and the model achieved an average Dice similarity coefficient of 0.89. The radiomics models effectively predicted four molecular subtypes, with the best performance showing an AUC of 0.8623, an accuracy of 0.6596, a sensitivity of 0.6383, and a specificity of 0.8775. For specific subtypes, the AUC values ranged from 0.8676 to 0.9335, with specificity reaching 0.9865 [120]. Qian et al. used CIBERSORT to analyze immune cell infiltration in BC (BRCA) patients from the TCGA database, applying univariate and multivariate Cox regression to identify M2 macrophages as an independent prognostic factor (HR = 32.288) while also obtaining imaging data from the TCIA database to construct an MRI-based radiomics model by selecting key features via LASSO. The models developed included intratumoral, peritumoral, and combined types, with the peritumoral model showing the highest performance, achieving an accuracy of 0.773, a sensitivity of 0.727, and a specificity of 0.818 [121].

Guo et al. proposed an MRI-based deep learning radiomics (DLR) model to predict HER2-low-positive status in BC patients and evaluate its prognostic value. The model utilized features from traditional radiomics and deep semantic segmentation and achieved AUCs of 0.868 and 0.763 for distinguishing HER2-negative from HER2-overexpressing patients, and 0.855 and 0.750 for differentiating HER2-low-positive from HER2-zero patients in the training and validation cohorts, respectively [122]. The study by Chao et al. used 569 local cases and 125 external cases for lesion classification, employing T1-weighted, DCE-MRI, T2WI, and diffusion-weighted imaging. A CNN and LSTM cascaded network achieved AUCs of 0.98/0.91 and sensitivities of 0.96/0.83 in internal/external cohorts. The DL model outperformed radiologists without DCE-MRI (AUC 0.96 vs. 0.90), and lesion localization had sensitivities of 0.97/0.93 with DCE-MRI/T2WI [123]. Liang et al. proposed a DL model (MLP-radiomic) using MRI radiomics and clinical radiological features, achieving a high AUC of 0.896 in predicting lymphovascular invasion in BC patients and outperforming ML models [124]. Table 17 provides an overview of studies conducted using MRI image-based BC detection.

Table 17. Comparison of recent MRI image-based techniques for breast cancer detection.

Ref.	Year	Dataset	Proposed Model	Proposed Approach	Results
[117]	2024	Private	Radiomics	Used multimodal MRI (T1WI, T2WI, DWI, ADC, DCE), feature extraction (plain, enhanced, diffuse), deep learning for classification	AUC: 0.927 Accuracy: 0.9015 Sensitivity: 0.9631 Sensitivity: 0.9548
[118]	2022	Inter MRI dataset + I-SPY TRIAL	Weakly supervised network based on the ResNet-101 architecture	ResNet-101-based weakly supervised framework, Adam optimizer, SoftMax threshold (0.5)	AUC: 0.92 (±0.03) Accuracy: 94.2% (±3.4) Sensitivity: 74.4% (±8.5) Specificity: 95.3% (±3.3)
[119]	2022	Private	CNN	Uses a CNN with DCE-MRI images, combining tumor features and confidence analysis via prediction probability, heatmaps, and scan timing	Accuracy: 0.877 Precision: 0.912 Sensitivity: 0.861 AUC: 0.912 (±4.0%)
[120]	2023	Private	3D UNet-based CNN	3D UNet-based CNN for automatic segmentation cancer regions, extracting radiomics features, applying cross-combination approaches with feature optimization and then classification	AUC: 86.23% Accuracy: 65.96% Sensitivity: 63.83% Specificity: 87.75%
[121]	2024	TCGA	MRI-based radiomics model	integrates CIBERSORT analysis and an MRI-based radiomics model with LASSO, identifying M2 macrophages as a key prognostic factor for BC	Accuracy: 0.773 Sensitivity: 0.727 Specificity: 0.818
[122]	2024	Private	Deep learning radiomics (DLR) model	Radiomics and deep semantic segmentation features combined for HER2 status prediction	HER2-negative vs. HER2-overexpressing Training AUC: 0.868 Validation AUC: 0.763 HER2-low-positive vs. HER2-zero Training AUC: 0.855 Validation AUC: 0.750
[123]	2024	Private	CNN and LSTM Cascaded Network	lesion classification and localization using mpMRI sequences, with histopathology as the ground truth	Lesion classification Internal cohort: AUC = 0.98, Sensitivity = 0.96 External cohort: AUC = 0.91, Sensitivity = 0.83 DL method vs. radiologists (without DCE-MRI): AUC = 0.96 vs. 0.90
[124]	2024	Private	MLP-radiomics	Combines radiomics and clinicoradiological features via MLP for lymphovascular invasion (LVI) prediction in BC	AUC: 0.896

Figure 18 illustrates the highest accuracy rates obtained by various models on MRI datasets in 2023 and 2024, including datasets such as RIDER [125], DCE-MRI [126], and Rider Breast MRI Public Dataset [127], as detailed in the referenced studies, accompanied by a correlation figure that represents accuracy levels through color intensity.

Figure 18. Highest detection accuracy across multiple MRI datasets for breast cancer.

8.3.1. Model Performance on MRI Image: Pros, Cons, and Future Directions

Table 18 provides an overview of the key benefits, limitations, and future research directions for leading MRI-based breast cancer detection models, enabling easy comparison of their capabilities and areas for improvement.

Table 18. Comprehensive overview of top-performing models on MRI breast cancer datasets.

Ref.	Advantages	Limitations	Future Work Recommendations
[117]	The study developed a radiomics model using multimodal MRI data (T1WI, T2WI, DWI, ADC, DCE) from 95 patients, achieving high diagnostic performance with an AUC of 0.927, accuracy of 90.15%, and sensitivity of 96.31% for the combined (plain scan + enhanced + diffuse) model.	The dataset was relatively small (n = 95), and only mass lesions were analyzed, excluding non-mass enhancements. Additionally, only intratumoral radiomics features were extracted, omitting valuable peritumoral information.	Future studies should include larger datasets, incorporate peritumoral features, and integrate clinical and genetic data for more comprehensive modeling. Enhancing AI algorithms and fostering interdisciplinary research are also recommended to improve diagnostic precision and individualized treatment.
[126]	Using a dataset of 1359 DCE-MRI images, the proposed CNN-SVM hybrid model achieved 99.80% accuracy and 100% AUC in classifying molecular subtypes of breast cancer with low computational complexity.	The study relied on a manually annotated dataset, which may limit scalability, and it used only the first post-contrast DCE-MRI, potentially excluding valuable dynamic information.	Future research should explore combining pre- and post-contrast MRI dynamics to enhance classification performance and validate results on larger, multi-center datasets for broader clinical applicability.
[127]	The modified YOLACT model, trained on the Rider Breast MRI dataset with 6000 preprocessed images (4400 normal and 1600 difficult), achieved a classification accuracy of 98.5% for common samples and 93.6% for difficult samples, outperforming standard YOLACT and offering faster performance than Mask R-CNN.	The model was trained and validated using data only from the Rider Breast MRI dataset, which may limit its generalizability. Additionally, comparisons were limited to YOLACT and Mask R-CNN, without evaluation against broader advanced models.	Future research should involve multi-source MRI datasets, compare the model with other advanced segmentation algorithms, and evaluate its diagnostic performance against histopathological results for real-world clinical applicability.

8.3.2. Summary

Magnetic resonance imaging (MRI) has emerged as a powerful tool for breast cancer (BC) detection, particularly in cases involving dense breast tissue where other imaging modalities may be less effective. Recent advancements in machine learning (ML) and deep learning (DL) have significantly enhanced the accuracy of MRI-based BC detection. Studies have utilized various MRI modalities, including T1-weighted, T2-weighted, diffusion-weighted imaging (DWI), apparent diffusion coefficient (ADC), and dynamic contrast-enhanced (DCE) MRI, to extract relevant features and detect cancerous lesions. These approaches have achieved impressive results, with models evaluated based on metrics such as accuracy, area under the curve (AUC), and Dice similarity coefficient. However, challenges related to dataset variability, model generalizability, and computational complexity remain. Table 19 presents a summary of key findings in MRI-based breast cancer detection.

Table 19. Summary of key findings in MRI-based breast cancer detection.

Aspect	Details
Datasets	Private datasets, TCGA, RIDER, DCE-MRI, etc.
MRI Modalities	T1WI, T2WI, DWI, ADC, DCE-MRI
Models	ResNet-101, CNN, 3D UNet, CNN-LSTM, MLP-radiomics
Techniques	Multimodal MRI, weakly supervised learning, radiomics, deep semantic segmentation
Highest AUC	0.98 (CNN-LSTM cascaded network)
Key Challenges	Generalizability, dataset variability, computational complexity

8.3.3. Analysis

The integration of multimodal MRI approaches with deep learning techniques has significantly improved BC detection accuracy. Studies like Zhang et al. (2024) [112] demonstrate the advantage of combining multiple MRI modalities (T1WI, T2WI, DWI, ADC, and DCE) to capture complementary features, achieving an AUC of 0.927. Weakly supervised learning, as employed by Liu et al. (2022) [118], has shown promise in scenarios with limited annotated data, achieving an AUC of 0.92. CNN-based architectures, such as ResNet-101 and 3D UNet, have been widely used for classification and segmentation tasks, with notable performance improvements. Additionally, advanced models incorporating immune composition data, such as Qian et al. (2024) [121], have opened new avenues for predicting patient prognosis and cancer status. Table 20 provides an analysis of techniques and their impact on MRI-based breast cancer detection.

Table 20. Analysis of techniques and their impact in MRI-based breast cancer detection.

Technique	Impact
Multimodal MRI	Combines complementary features from different MRI sequences for improved accuracy.
Weakly Supervised Learning	Effective in scenarios with limited annotated data, reducing labeling costs.
CNN Architectures	High accuracy in classification and segmentation tasks.
Radiomics	Extracts quantitative features from medical images, enhancing predictive power.
Immune Composition Integration	Enables prognosis prediction and personalized treatment strategies.

8.3.4. Critique

Despite the impressive results, several limitations need to be addressed. First, the variability in performance across different datasets and cohorts raises concerns about model generalizability. Many models perform well on internal datasets but struggle with external validation, as seen in Chao et al. (2024) [123]. Second, the reliance on large annotated datasets remains a significant challenge, as acquiring such data is costly and time-consuming. Third, the computational complexity of advanced models, such as CNN-LSTM cascaded networks, may limit their feasibility in real-world clinical settings. Finally, while models achieve high-performance metrics, their lack of interpretability hinders clinical trust and adoption. Table 21 presents a critique of current limitations in MRI-based breast cancer detection.

Table 21. Critique of current limitations in MRI-based breast cancer detection.

Limitation	Details
Generalizability	Models often fail to perform consistently across external datasets.
Dataset Dependency	Reliance on large annotated datasets limits scalability.
Computational Complexity	High resource requirements hinder real-world deployment.
Interpretability	Lack of transparency in model decisions reduces clinical trust.
Clinical Integration	Challenges in integrating AI-driven decisions into clinical workflows.

8.3.5. Comparison

The reviewed studies demonstrate varying approaches and outcomes, as summarized in Table 22. Chao et al. (2024) [123] achieved the highest AUC of 0.98 using a CNN-LSTM cascaded network, outperforming radiologists in lesion classification. Liu et al. (2022) [118] and Yue et al. (2023) [120] also reported strong performance, with AUC values of 0.92 and 0.86, respectively. The integration of radiomics features, as seen in Qian et al. (2024) [121], further enhanced predictive accuracy, achieving an AUC of 0.927. However, performance discrepancies in external validation datasets highlight the need for further research in model generalization. The table also highlights variations in model complexity, with simpler models like MLP-radiomics offering a balance between performance and computational efficiency.

Table 22. Comparison of key studies in MRI-based breast cancer detection.

Study	Dataset	Model/Technique	AUC	Key Findings
Zhang et al. (2024) [117]	Private	Multimodal MRI + DL	0.927	Combines T1WI, T2WI, DWI, ADC, and DCE-MRI.
Liu et al. (2022) [118]	Inter MRI + I-SPY	ResNet-101 (weakly supervised)	0.92	Effective with limited annotated data.
Wu et al. (2022) [119]	Private	CNN + DCE-MRI	0.912	Combines tumor features and confidence analysis.
Yue et al. (2023) [120]	Private	3D UNet + Radiomics	0.8623	Predicts molecular subtypes with high specificity.
Qian et al. (2024) [121]	TCGA	MRI-based Radiomics + CIBERSORT	0.773	Identifies M2 macrophages as a prognostic factor.
Guo et al. (2024) [122]	Private	DLR (Radiomics + Deep Semantic Segmentation)	0.868	Predicts HER2 status with high accuracy.
Chao et al. (2024) [123]	Private	CNN-LSTM Cascaded Network	0.98	Outperforms radiologists in lesion classification.
Liang et al. (2024) [124]	Private	MLP-Radiomics	0.896	Predicts lymphovascular invasion effectively.

MRI-based breast cancer detection has seen significant advancements through the integration of deep learning and radiomics techniques. While models like CNN-LSTM and ResNet-101 demonstrate exceptional performance, challenges related to generalizability, dataset dependency, and computational complexity remain. Future research should focus on improving model interpretability, expanding dataset diversity, and developing computationally efficient models for real-world clinical deployment. Addressing these challenges will enhance the clinical applicability and robustness of MRI-based BC detection systems.

MRI-based models face unresolved challenges: (1) multi-institutional data variability degrades performance (Table 17), (2) most studies ignore computational costs prohibitive for clinical deployment, (3) temporal dynamics in DCE-MRI remain underutilized.

8.4. Breast Cancer Detection Using Computed Tomography (CT) Scan Images

Computed tomography (CT) imaging, a diagnostic method using X-ray technology to produce detailed cross-sectional body images, facilitates the detection and monitoring of abnormalities such as tumors; this section examines AI-driven techniques for BC detection using CT images, with Figure 19 highlighting a region of interest for further analysis.

Figure 19. A sample CT image [116].

Bargalló et al. [128] proposed a comprehensive CT imaging pipeline for BC that employs a trained CNN-Unet model to auto-contour standard tissue structures, thereby enabling accurate classification of patients by laterality and surgical procedures, such as lumpectomy or mastectomy, with a Random Forest classifier achieving 93.75% accuracy and an AUC of 0.993 in a secondary sample of 16 patients. While this method reduced target delineation errors and highlighted the need to consider anatomical and procedural differences, there remains scope for enhancing workflow efficiency. Koh et al. developed a DL model for detecting BC from chest CT scans using augmented axial images with bounding boxes and the RetinaNet algorithm. The model achieved 96.5% sensitivity in the internal test set and 96.1% in the external test set, with sensitivities ranging from 88.5% to 93.0% depending on the candidate probability threshold [129].

Yasak et al. conducted a study on BC detection using DL models applied to contrast-enhanced chest CT images. They used a dataset of 201 training, 26 validation, and 30 test cases and achieved an AUC of 0.976, thus improving the performance of radiologists through model assistance [130]. Table 23 presents a summary of research focused on BC detection using computed tomography (CT) images.

Table 23. A comparison of recent techniques for breast cancer detection using computed tomography.

Ref.	Year	Dataset	Image Size	Proposed Model	Approach	Results
[128]	2023	Private	-	CNN-Unet with Random Forest	Automatic segmentation, normal tissues, laterality classification, target volume delineation, procedural classification.	Accuracy: 0.9375 AUC: 0.993
[129]	2022	Private	512 × 512	RetinaNet	Augmented chest CT images with bounding boxes and RetinaNet for breast cancer detection	Internal Test Set Sensitivity: 96.5% with 13.5 false positives per case (FPs/case) External Test Set Sensitivity: 96.1% with 15.6 FPs/case
[130]	2024	Private	512 × 512	Deep Learning	Enhanced contrast of chest CT images and the applied deep learning model	AUC on Validation set: 0.976 and Test set: 0.967
[128]	2023	Private	-	Cnn-U-Net with Stochastic Gradient Descent	Automatic contouring of normal tissues, classification of laterality, automatic delineation of target volumes, and classification	Accuracy: 0.9375 AUC: 0.93

Figure 20 displays the highest accuracy rates achieved by various models on computed tomography datasets in 2023 and 2024, including the Breast Wellness Centre [131], Bone Metastases: Breast vs. Non-Breast Cancer [132], GE Healthcare [133], and datasets from hospitals in Pakistan [134], as outlined in the referenced studies, along with a correlation figure illustrating accuracy levels using varying color intensities.

Figure 20. Highest detection accuracy across multiple computed tomography datasets for breast cancer.

8.4.1. Model Performance on CT Image: Pros, Cons, and Future Directions

The comparative analysis in Table 24 highlights the main strengths, weaknesses, and research opportunities for current CT imaging approaches in breast cancer detection.

Table 24. Comprehensive overview of top-performing models on CT breast cancer datasets.

Ref.	Advantages	Limitations	Future Work Recommendations
[133]	Chest CT achieved high diagnostic performance with 84.21% sensitivity, 99.3% specificity, and 98.70% accuracy, outperforming mammography (78.95%, 93.78%, and 93.16%, respectively). It also reduced unnecessary biopsies and eliminated the need for patient recall.	The study was retrospective and used experienced radiologists, so results may not generalize to less experienced readers. Breast imaging was incomplete in 132 cases, and standard-dose CT was used rather than optimized low-dose protocols.	Prospective studies with low-dose CT and standardized imaging protocols are needed. Integration of AI algorithms and the use of skin markers for consistent breast positioning could further improve diagnostic accuracy and reproducibility.
[131]	The proposed CAD system achieved high performance metrics with 95.2% accuracy, 92.4% sensitivity, and 98.1% specificity on CTLM images, demonstrating robust detection of breast cancer lesions. The 3D Fuzzy C-means segmentation and MLPNN classifier effectively analyzed angiogenesis patterns, with an AUROC of 0.98, outperforming traditional methods.	The study was limited by a small dataset (180 patients) and lacked diverse angiogenesis shapes, potentially reducing generalizability. Additionally, the reliance on TIFF instead of DICOM format may have impacted image quality and processing efficiency.	Expanding the dataset to include more angiogenesis variants (e.g., ringed or polypoid shapes) could improve model robustness. Integrating DICOM-formatted images and testing in multi-center clinical trials would enhance clinical applicability.

8.4.2. Summary

Computed tomography (CT) imaging is a widely used diagnostic tool for breast cancer (BC) detection, offering detailed cross-sectional images of the body. Recent advancements in artificial intelligence (AI) and deep learning (DL) have significantly enhanced the accuracy of BC detection using CT images. Studies have employed various DL models, such as CNN-Unet and RetinaNet, to segment and classify breast cancer lesions with high accuracy. For example, Bargalló et al. (2023) [128] achieved 93.75% accuracy using a CNN-Unet model combined with a Random Forest classifier, while Koh et al. (2022) [129] reported 96.5% sensitivity using RetinaNet on augmented CT images. Despite these advancements, challenges such as dataset dependency, computational complexity, and generalizability remain. Table 25 presents a summary of key findings in CT-based breast cancer detection.

Table 25. Summary of key findings in CT-based breast cancer detection.

Aspect	Details
Datasets	Private datasets, Breast Wellness Centre, GE Healthcare, etc.
Models	CNN-Unet, RetinaNet, Random Forest, Deep Learning
Techniques	Automatic segmentation, image augmentation, contrast enhancement
Highest Accuracy	93.75% (CNN-Unet + Random Forest)
Key Challenges	Dataset dependency, computational complexity, generalizability

8.4.3. Analysis

The integration of deep learning models with CT imaging has significantly improved BC detection accuracy. Bargalló et al. (2023) [128] proposed a comprehensive CT imaging pipeline using CNN-Unet for automatic segmentation and Random Forest for classification, achieving 93.75% accuracy and an AUC of 0.993. This approach reduces target delineation errors and highlights the importance of considering anatomical and procedural differences. Koh et al. (2022) [129] employed RetinaNet with augmented axial images, achieving 96.5% sensitivity in internal test sets and 96.1% in external test sets, demonstrating robust performance across different datasets. Yasak et al. (2024) [130] utilized deep learning on contrast-enhanced CT images, achieving an AUC of 0.976, which improved radiologists’ performance. These studies underscore the effectiveness of combining advanced CT imaging techniques with deep learning models for accurate BC detection. Table 26 provides an analysis of techniques and their impact in CT-based breast cancer detection.

Table 26. Analysis of techniques and their impact in CT-based breast cancer detection.

Technique	Impact
Automatic Segmentation	Reduces human error and improves workflow efficiency.
Image Augmentation	Enhances model performance by increasing dataset variability.
Contrast Enhancement	Improves lesion visibility and detection accuracy.
Ensemble Models	Combines strengths of multiple models, improving robustness and accuracy.
Weakly Supervised Learning	Effective in scenarios with limited annotated data, reducing labeling costs.

8.4.4. Critique

Despite the promising results, several limitations need to be addressed. First, the reliance on high-quality, annotated CT image datasets limits scalability, as such datasets are not always available in clinical settings. Second, the performance of deep learning models can vary when applied to real-world data, as seen in Koh et al. (2022) [129], where sensitivity varied across test sets. Third, the computational complexity of advanced models, such as CNN-Unet and RetinaNet, may hinder their practical implementation in resource-limited environments. Finally, while automatic segmentation and procedural classification reduce human error, they require further refinement to improve workflow efficiency. The lack of generalizability across diverse datasets and clinical practices remains a significant challenge. Table 27 presents a critique of current limitations in CT-based breast cancer detection.

Table 27. Critique of current limitations in CT-based breast cancer detection.

Limitation	Details
Dataset Dependency	Reliance on high-quality, annotated datasets limits scalability.
Generalizability	Models often fail to perform consistently across external datasets.
Computational Complexity	High resource requirements hinder real-world deployment.
Workflow Efficiency	Automatic segmentation methods require refinement for clinical integration.
Clinical Integration	Challenges in integrating AI-driven decisions into clinical workflows.

8.4.5. Comparison

The reviewed studies demonstrate varying approaches and outcomes, as summarized in Table 28. Bargalló et al. (2023) [128] achieved 93.75% accuracy using CNN-Unet and Random Forest, while Koh et al. (2022) [129] reported 96.5% sensitivity using RetinaNet. Yasak et al. (2024) [130] utilized deep learning on contrast-enhanced CT images, achieving an AUC of 0.976. These studies highlight the diversity of approaches in CT-based BC detection, with each method offering unique advantages. For instance, CNN-Unet excels in automatic segmentation, while RetinaNet performs well in lesion detection. However, performance discrepancies across datasets underscore the need for further research in model generalization.

Table 28. Comparison of key studies in CT-based breast cancer detection.

Study	Dataset	Model/Technique	Accuracy /AUC	Key Findings
Bargalló et al. (2023) [128]	Private	CNN-Unet + Random Forest	93.75%	Reduces target delineation errors.
Koh et al. (2022) [129]	Private	RetinaNet	96.5% (internal)	Robust performance across datasets.
Yasak et al. (2024) [130]	Private	Deep Learning + Contrast Enhancement	AUC 0.976	Improves radiologists’ performance.

CT-based breast cancer detection has seen significant advancements through the integration of deep learning models and advanced imaging techniques. While models like CNN-Unet and RetinaNet demonstrate exceptional performance, challenges related to dataset dependency, computational complexity, and generalizability remain. Future research should focus on improving model interpretability, expanding dataset diversity, and developing computationally efficient models for real-world clinical deployment. Addressing these challenges will enhance the clinical applicability and robustness of CT-based BC detection systems.

While CT-based models show promise (Table 28), two fundamental gaps limit clinical translation: (1) Radiation concerns—Most studies use standard-dose CT [133], ignoring the trade-offs between detection accuracy and patient safety in low-dose protocols; (2) Anatomical specificity—Models like CNN-Unet [128] focus on lesion detection but fail to incorporate breast density patterns that critically impact diagnostic interpretation.

8.5. Histopathological Image-Based Breast Cancer Detection

This subsection explores the application of histopathological and microscopic imaging, key diagnostic tools for BC detection that involve examining tissue samples at the cellular level to identify malignancies and support accurate diagnosis and treatment planning. A simple histopathological image is shown in Figure 21.

Figure 21. A sample histopathological image [135].

Muntean and Chowkkar compared CNN and DenseNet121 models for BC detection using 7909 histopathological images. The CNN achieved 90.9% accuracy with 100× magnification, while DenseNet121 reached 86.6% accuracy, with a 16.4% accuracy improvement through transfer learning. Both models demonstrated effective performance [136]. Bhowal et al. [137] proposed a BC histology classification method that combined multiple DL models—VGG16, VGG19, Xception, Inception V3, and InceptionResnet V2—using Choquet Integral, Coalition Game Theory, and Information Theory for synthesis. The method incorporated image preprocessing, feature extraction, and classification, achieving improved accuracy, precision, and recall over traditional techniques. The dataset used for evaluation was from the BACH dataset, with an accuracy of 96% for the two-class problem and 95% for the four-class problem, outperforming state-of-the-art methods.

Yang et al. proposed a multimodal DL model to predict the prognosis of HER2-positive BC patients by integrating whole slide H&E images and clinical data. The model utilized a deep convolutional neural network (CNN) to extract features from 512 × 512-pixel H&E images that were combined with clinical information to assess the risks of relapse and metastasis. The model achieved an AUC of 0.76 in two-fold cross-validation and 0.72 in an independent TCGA testing set, demonstrating its effectiveness in predicting the prognosis of HER2-positive BC patients across different datasets [138].

Maleki et al. proposed a method for histopathological image classification using DL and transfer learning. The approach utilized pre-trained networks for feature extraction and then applied the XGBoost classifier. Evaluated on the BreakHis dataset, the proposed method achieved accuracies of 93.6%, 91.3%, 93.8%, and 89.1% for magnifications of 40×, 100×, 200×, and 400×, respectively. The DenseNet201 model was used for feature extraction, and XGBoost was employed as the final classifier [139]. Majumdar et al. proposed a rank-based ensemble method for BC detection from histopathological images, utilizing three CNN models: GoogleNet, VGG11, and MobileNetV3_Small. The decision scores from these models were combined using the Gamma function for a two-class classification problem. Their approach was evaluated on the BreakHis and ICIAR-2018 datasets, achieving classification accuracies of 99.16%, 98.24%, 98.67%, and 96.16% for 40×, 100×, 200×, and 400× magnification levels on the BreakHis dataset and 96.95% on the ICIAR-2018 dataset [140]. Ray et al. proposed the use of pre-trained deep transfer learning models, including ResNet50, ResNet101, VGG16, and VGG19, for the detection of BC using histopathological images. The study utilized a dataset of 2453 images, divided into invasive ductal carcinoma (IDC) and non-IDC categories. Among the models analyzed, ResNet50 showed superior performance, achieving an accuracy of 92.2%, an AUC of 91.0%, and a recall of 95.7%, with a minimal loss of 3.5% [141].

Rajkumar et al. proposed a method for BC detection using the Darknet-53 convolutional neural network (CNN) that enhanced the accuracy of image classification. They utilized the contrast-limited adaptive histogram equalization (CLAHE) technique for image preprocessing and the Haralick grey-level co-occurrence matrix (HGLCM) for feature extraction. The model achieved an accuracy of 95.6% [142]. Alhassan et al. proposed a comprehensive classification framework for differentiating types of lung, colon, and breast cancers by analyzing histopathological images using AI, machine learning, and digital image processing techniques. They utilized the BreakHis and LC25000 datasets and applied various preprocessing steps, including quadruple clipped adaptive histogram equalization for noise reduction and stain color adaptive normalization (SCAN) to enhance image contrast. The VGG16 model was used for feature extraction, followed by an ensemble max-voting classifier for classification. The results demonstrated strong performance, with 89.03% accuracy, 88.56% precision, 88.09% sensitivity, and an F1-score of 88.7%, highlighting the effectiveness of their approach for faster and more accurate cancer diagnoses [143]. The study used the BreaKHis and BACH datasets for histopathological BC image classification. The image sizes varied according to the magnification level, with images from 40×, 100×, 200×, and 400× magnifications used for evaluation. The proposed model was BCHI-CovNet, a lightweight AI model employing multiscale depth-wise separable convolution, an additional pooling module, and a multi-head self-attention mechanism to reduce computational complexity while maintaining high accuracy. The model achieved impressive results, with accuracies of 99.15% at 40×, 99.08% at 100×, 99.22% at 200×, and 98.87% at 400× magnification on the BreaKHis dataset, and 99.38% on the BACH dataset [144]. A summary of BC classification using histopathological images is provided in Table 29.

Table 29. A review of recent approaches for detecting breast cancer through histopathological images.

Ref.	Year	Dataset	Image Size	Proposed Model	Approach	Results
[136]	2022	BreaKHis	128 × 128 64 × 64 32 × 32	CNN and DenseNet121	The CNN used tuning, and DenseNet121 applied transfer learning, with magnification, scaling, and rotation effects	CNN: Accuracy: 90.9% DenseNet121: Accuracy: 86.6%
[137]	2022	ICIAR 2018	512 × 512	Choquet Integral-based Fusion Model	DCNN feature extraction, and model fusion (VGG16, Xception, Inception, InceptionResNet V2) using Choquet Integral and Coalition Game Theory	Two-Class (Accuracy: 96%, Precision and Recall: 0.96) Four-Class (Accuracy: 95%, Precision & Recall: 0.95)
[138]	2022	H&E images from Cancer Hospital of the Chinese Academy of Medical Sciences, and TCGA	512 × 512	Multimodal deep learning model	Integrates whole slide H&E images with clinical data, DCNN for feature extraction, classification	AUC: 0.76 CV: 0.72 in the TCGA testing set
[139]	2023	BreakHis	128 × 128 64 × 64 32 × 32	Extreme Gradient Boosting	DenseNet201 networks to extract features, classification using XGBoost	Accuracy: 93.6% (40×) 91.3% (200×) 93.8% (400×)
[140]	2023	BreakHis, ICIAR-2018	128 × 128 512 × 512	Ensemble technique	Rank-based ensemble of GoogleNet, VGG11, and MobileNetV3_Small	Accuracy: 99.16% (40X) on BreakHis, and 96.95% on ICIAR-2018
[141]	2024	Breast Cancer Histopathology	-	ResNet50	Transfer learning models (ResNet50, ResNet101, VGG16, VGG19)	Accuracy: 0.922 AUC: 0.91 Recall: 0.957
[142]	2024	Private	-	Darknet-53 CNN	CLAHE for image enhancement and HGLCM for feature extraction. The Darknet-53 CNN for classification	Accuracy: 95.6%
[143]	2024	BreakHis and LC25000	128 × 128	Ensemble max-voting	QCAHE reduced noise, SCAN optimized stain preservation, and Arithmetic Optimization improved SCAN parameters.	Accuracy: 89.03% Precision: 88.56% Sensitivity: 88.09% F1-score: 88.7% Specificity: 89.08%
[144]	2024	BreaKHis and BACH	128 × 128 512 × 512	BCHI-CovNet	Uses multiscale depth-wise separable convolution, pooling, and self-attention	99.38% accuracy on BACH and 99.22% on BreaKHis

Figure 22 presents the highest accuracy rates achieved by various models on Histopathological datasets in 2023 and 2024, including the ImageNet [145], BreakHis [146], Pennsylvania University Hospital (HUP) and the New Jersey Cancer Institute [147], BACH [144], LC25000 [143], MITOS-ATYPIA-14 [148], PatchCamelyon (PCam) Benchmark [149], IDC (Invasive Ductal Carcinoma) [150], and ICIAR-2018 [140] datasets, as detailed in the referenced studies, paired with a correlation figure that conveys accuracy levels through color intensity variation.

Figure 22. Highest detection accuracy across multiple histopathological datasets for breast cancer.

8.5.1. Model Performance on Histopathological Image: Pros, Cons, and Future Directions

The comparative analysis in Table 30 highlights the main strengths, weaknesses, and research opportunities for current histopathological imaging approaches in breast cancer detection.

Table 30. Comprehensive overview of top-performing models on histopathological breast cancer datasets.

Ref.	Advantages	Limitations	Future Work Recommendations
[144]	The proposed BCHI-CovNet model integrates depth-wise separable convolution, second-order pooling, and multi-head self-attention, achieving high classification accuracy with fewer parameters and low FLOPs. It attained 99.15% (40×), 99.08% (100×), 99.22% (200×), and 98.87% (400×) accuracy on the BreaKHis dataset, and 99.38% on the BACH dataset, demonstrating superior performance and lightweight design.	The model’s performance is affected by color variations in histopathological images across BreaKHis and BACH datasets. This color inconsistency limits feature extraction quality and impacts generalizability.	Future work should include advanced color normalization during preprocessing to handle image variability. Additionally, expanding the model evaluation to more diverse and multi-center datasets is suggested to enhance robustness.
[145]	The proposed Hybrid CNN model, integrating the Sine Cosine Algorithm (SCA) and transfer learning, achieves a high accuracy of 96.9%, outperforming traditional models like K-NN, SVM, and standard CNN. The model utilizes VGG16 pre-trained on ImageNet and fine-tunes the last three convolutional layers, enhancing performance while preventing overfitting.	The study lacks evaluation on publicly available breast cancer datasets such as BreaKHis or BACH, limiting comparison with existing works. Also, performance across varied imaging modalities and patient demographics remains unexplored.	Future work should validate the model on benchmark datasets like BreaKHis and BACH to assess generalizability. Incorporating multi-modal data and real-time blockchain integration for secure deployment in clinical settings is also recommended.
[137]	The proposed Choquet Integral-based ensemble achieved 96% (2-class) and 95% (4-class) accuracy on the BACH dataset, outperforming all single models. It effectively combines features from five DCNNs using a novel fuzzy measure method.	The fuzzy measure calculation is computationally expensive and complex. The method was tested only on the BACH dataset, limiting generalizability.	Future work should reduce computational cost and test on larger datasets. Adding more models and improving stain robustness is also recommended.

8.5.2. Summary

Histopathological imaging is a critical tool for breast cancer (BC) detection, enabling the examination of tissue samples at the cellular level to identify malignancies. Recent advancements in deep learning (DL) have significantly improved the accuracy of BC detection using histopathological images. Studies have employed various DL models, such as CNNs, DenseNet, ResNet, and ensemble methods, to achieve high accuracy and AUC scores. For example, Muntean and Chowkkar (2022) [136] achieved 90.9% accuracy using a CNN, while Bhowal et al. (2022) [137] reported 96.0% accuracy using a fusion model. Despite these advancements, challenges such as dataset dependency, computational complexity, and generalizability remain. Table 31 presents a summary of key findings in histopathology-based breast cancer detection.

Table 31. Summary of key findings in histopathological-based breast cancer detection.

Aspect	Details
Datasets	BreakHis, BACH, ICIAR-2018, TCGA, LC25000, etc.
Models	CNN, DenseNet, ResNet, XGBoost, ensemble methods
Techniques	Transfer learning, feature fusion, multimodal analysis, preprocessing
Highest Accuracy	99.38% (BCHI-CovNet on BACH dataset)
Key Challenges	Dataset dependency, computational complexity, generalizability

8.5.3. Analysis

The integration of deep learning models with histopathological imaging has significantly improved BC detection accuracy. Muntean and Chowkkar (2022) [136] demonstrated the effectiveness of CNNs and DenseNet121, achieving 90.9% and 86.6% accuracy, respectively, with transfer learning further improving performance. Bhowal et al. [137] proposed a fusion model combining VGG16, Xception, and InceptionResNet V2, achieving 96.0% accuracy on the ICIAR 2018 dataset. Yang et al. (2022) [138] integrated whole slide H&E images with clinical data, achieving an AUC of 0.76, showcasing the potential of multimodal approaches. Ensemble methods, such as those proposed by Majumdar et al. (2023) [140], further enhance robustness, achieving 99.16% accuracy on the BreakHis dataset. Table 32 provides an analysis of techniques and their impact on histopathological-based breast cancer detection.

Table 32. Analysis of techniques and their impact in histopathological-based breast cancer detection.

Technique	Impact
CNN Architectures	High accuracy in classification tasks, effective feature extraction.
Transfer Learning	Improves performance on small datasets, reduces training time.
Ensemble Methods	Combines strengths of multiple models, improving robustness and accuracy.
Multimodal Approaches	Integrates imaging and clinical data for enhanced predictive power.
Preprocessing	Techniques like CLAHE improve image quality and model performance.

8.5.4. Critique

Despite the promising results, several limitations need to be addressed. First, the reliance on high-quality, annotated histopathological image datasets limits scalability, as such datasets are not always available in clinical settings. Second, the performance of deep learning models can vary when applied to real-world data, as seen in Yang et al. (2022) [138], where the AUC dropped to 0.76 in external validation. Third, the computational complexity of advanced models, such as ensemble methods and multimodal approaches, may hinder their practical implementation in resource-limited environments. Finally, while models achieve high-performance metrics, their lack of interpretability hinders clinical trust and adoption. Table 33 presents a critique of current limitations in histopathological-based breast cancer detection.

Table 33. Critique of current limitations in histopathological-based breast cancer detection.

Limitation	Details
Dataset Dependency	Reliance on high-quality, annotated datasets limits scalability.
Generalizability	Models often fail to perform consistently across external datasets.
Computational Complexity	High resource requirements hinder real-world deployment.
Interpretability	Lack of transparency in model decisions reduces clinical trust.
Clinical Integration	Challenges in integrating AI-driven decisions into clinical workflows.

8.5.5. Comparison

The reviewed studies demonstrate varying approaches and outcomes, as summarized in Table 34. Muntean and Chowkkar (2022) [136] achieved 90.9% accuracy using a CNN, while Bhowal et al. (2022) [137] reported 95.6% accuracy using a fusion model. Yang et al. (2022) [138] integrated clinical data with imaging, achieving an AUC of 0.76. Majumdar et al. (2023) [140] employed an ensemble method, achieving 99.16% accuracy on the BreakHis dataset. Addo et al. (2024) [144] proposed BCHI-CovNet, achieving 99.38% accuracy on the BACH dataset, showcasing the effectiveness of lightweight models. These comparisons highlight the diversity of approaches in histopathology-based BC detection, with each method offering unique advantages.

Table 34. Comparison of key studies in histopathological-based breast cancer detection.

Study	Dataset	Model/Technique	Accuracy/ AUC	Key Findings
Muntean & Chowkkar (2022) [136]	BreakHis	CNN, DenseNet121	90.9%	Transfer learning improves performance.
Bhowal et al. (2022) [137]	BACH	Fusion Model (VGG16, Xception, InceptionResNet V2)	95.6%	Combines multiple models for high accuracy.
Yang et al. (2022) [138]	TCGA	Multimodal DL model	AUC 0.76	Integrates imaging and clinical data.
Majumdar et al. (2023) [140]	BreakHis, ICIAR-2018	Ensemble method (GoogleNet, VGG11, MobileNetV3)	99.16%	Robust performance across datasets.
Addo et al. (2024) [144]	BACH, BreakHis	BCHI-CovNet	99.38%	Lightweight model with high accuracy.

Histopathological imaging combined with deep learning models has significantly advanced breast cancer detection, achieving high accuracy and AUC scores. While models like CNNs, DenseNet, and ensemble methods demonstrate exceptional performance, challenges related to dataset dependency, computational complexity, and generalizability remain. Future research should focus on improving model interpretability, expanding dataset diversity, and developing computationally efficient models for real-world clinical deployment. Addressing these challenges will enhance the clinical applicability and robustness of histopathological-based BC detection systems.

Four fundamental gaps emerge in histopathology analysis: (1) stain variation robustness is rarely tested, (2) whole-slide analysis remains computationally expensive, (3) few models incorporate pathologist feedback loops, (4) clinical significance of extracted features is often unverified.

8.6. Breast Cancer Detection Using Thermal Imaging

The thermal image shown in Figure 23 captures the heat emitted by objects, visualizing patterns of temperature variation. This study presents a CADx system for BC detection using DL, integrating multi-view thermograms (frontal and lateral) with patient clinical data. The system employs transfer learning with pre-trained models and focuses on regions of interest (ROIs) for targeted analysis. By addressing the limitations of single-view thermograms and incorporating critical clinical data, the proposed approach achieved 90.48% accuracy, 93.33% sensitivity, and an AUROC of 0.94, offering a cost-effective, less hazardous screening option [151]. Chatterjee et al. proposed a two-stage model for BC detection using thermographic images. The model utilizes VGG16 for feature extraction and a memory-based version of the Dragonfly Algorithm (DA) enhanced by the Grunwald–Letnikov (GL) method for optimal feature selection. Evaluated on the DMR-IR dataset, the framework achieved 100% diagnostic accuracy while reducing features by 82%, demonstrating its efficiency and potential for accurate early detection [152].

Figure 23. A sample thermal image [153].

Civilibal et al. proposed a Mask R-CNN model with ResNet-50 and ResNet-101 backbones for breast tumor detection, segmentation, and classification using thermal images. The ResNet-50 model pre-trained on COCO achieved superior performance, with 97.1% accuracy, an mAP of 0.921, and an overlap score of 0.868, outperforming models used in prior studies. This single DL model effectively delineates tumors and adjacent tissues for accurate diagnosis [154]. The study utilized the DMR database for thermal breast image classification, employing a custom CNN model trained on augmented, standardized, and enhanced images with histogram of oriented gradients (HOG) feature extraction. The model achieved impressive accuracy rates ranging from 95.7% to 98.5%, outperforming other classifiers and demonstrating a significant potential for real-time BC diagnosis [155]. Mahoro et al. proposed a BC detection system using DL techniques. They employed TransUNet for breast region segmentation and four models—ResNet-50, EfficientNet-B7, VGG-16, and DenseNet-201—for classifying images into healthy, sick, or unknown categories. The best performance was achieved with ResNet-50 and demonstrated an accuracy of 97.26%, a sensitivity of 97.26%, and a specificity of 100%, 96.94%, and 99.72% for healthy, sick, and unknown classes, respectively [156].

Husaini et al. investigated the impact of various types of noise on DL models for early BC detection using thermal images. They evaluated the performance of several models, particularly Inception MV4, under different noise conditions, showing that noise significantly reduced classification accuracy without proper preprocessing. The proposed method achieved impressive results, with 99.748% accuracy, 0.996 sensitivity, 1.0 specificity, and 0.998 AUC. Their findings highlight the importance of noise reduction strategies and preprocessing techniques to improve the reliability and accuracy of thermal imaging for BC diagnosis [157]. Table 35 provides a summary of the results of BC detection using thermal imaging.

Table 35. A review of recent methods for breast cancer detection using thermal imaging.

Ref.	Year	Dataset	Image Size	Proposed Model	Approach	Results
[151]	2023	MIAS Thermogram	224 × 224	Transfer learning	Multi-view thermograms, clinical data integration, ROI extraction, transfer learning	Accuracy: 90.48% Sensitivity: 93.33% AUC: 0.94
[152]	2022	DMR-IR	640 × 480	VGG16 + Memory-Based Dragonfly Algorithm	Two-stage framework: (1) VGG16 for feature extraction, (2) GL-enhanced DA for feature selection	Accuracy: 100% with 82% fewer features than VGG16
[154]	2023	COCO	-	Mask R-CNN with ResNet-50 and ResNet-101 backbones	DL model detects, segments, and classifies breast tissues, assigning bounding boxes and tumor borders.	ResNet-50: 97.1% accuracy, mAP of 0.921, overlap score of 0.868
[155]	2024	DMR	640 × 480	Custom CNN model	Augmentation, resizing, filtering, and HOG feature extraction; custom CNN for classification	Accuracy rates of 95.7% to 98.5%
[156]	2024	Not mentioned	224 × 224	TransUNet	TransUNet for segmentation, followed by classification	Accuracy: 97.26% Sensitivity: 97.26%
[157]	2024	Not mentioned	224 × 224	Inception MV4	Noise mitigation and preprocessing techniques, followed by classification.	Accuracy: 99.748% Sensitivity: 0.996 Specificity: 1.0 Precision: 1.0 AUC: 0.998

Figure 24 illustrates the highest accuracy rates achieved by various models on thermograph datasets in 2023 and 2024, including the Real-Time Thermography Video Streaming [158], thermal images from Benha University Hospital [159], RGC-IR [42], DMR-IR [160], and infrared images [161], as outlined in the referenced studies, accompanied by a correlation figure where accuracy levels are represented by different color intensities.

Figure 24. Highest detection accuracy across multiple thermograph datasets for breast cancer.

8.6.1. Model Performance on Thermal Image: Pros, Cons, and Future Directions

Table 36 presents a comparative analysis of the top performing models, outlining their key strengths, limitations, and potential avenues for future research in thermal imaging-based breast cancer detection.

Table 36. Comprehensive overview of top-performing models on thermal breast cancer datasets.

Ref.	Advantages	Limitations	Future Work Recommendations
[160]	The proposed hybrid model using Genetic Algorithm (GA) and Grey Wolf Optimizer (GWO) achieves 100% accuracy on the DMR-IR dataset, utilizing only 3% of the 1000 deep features extracted by the lightweight SqueezeNet 1.1 model. This significantly reduces computational time, storage, and processing overhead while maintaining state-of-the-art performance.	Despite the promising results, the model is evaluated on a single primary dataset (DMR-IR), which limits the generalizability across diverse clinical settings. Moreover, the study does not consider real-time deployment scenarios or the variability in thermal imaging conditions and patient demographics.	Future work should focus on validating the approach across multiple and more diverse datasets, including real-world clinical environments, to ensure robustness and scalability. Additionally, integrating multi-modal data such as clinical reports or other imaging types (e.g., mammography, ultrasound) may enhance diagnostic accuracy and applicability.
[158]	The proposed real-time thermography framework using deep learning models—Inception v3, Inception v4, and modified Inception Mv4—achieved high detection accuracy, with Inception Mv4 reaching 99.748%, outperforming Inception v4 (99.712%) and Inception v3 (96.8%) on a dataset of 1000 thermal images (700 normal, 300 abnormal) captured via FLIR One Pro thermal camera. Incorporating in situ cooling enhanced detection accuracy by up to 11%, particularly for small and deep tumors.	The detection performance is sensitive to external conditions like room temperature and patient movement, which can affect thermal image quality and classification results. Moreover, smaller or deeper tumors (e.g., <2 V and depth T9) remained difficult to detect even with cooling, and patient discomfort during cooling procedures may limit widespread applicability.	Future work should integrate thermal video clips into the dataset and explore higher-resolution thermal cameras to improve small tumor detection. Additionally, using alternative or advanced deep learning models (e.g., spiking neural networks) and standardizing situ-cooling protocols can further enhance real-time breast cancer diagnostic precision and reproducibility.
[161]	The proposed intelligent method, using infrared breast thermograms from the DMR-IR dataset (200 subjects), achieved high accuracy in malignancy detection—98.8% with SVM, outperforming kNN and D-Tree classifiers. Features extracted using Local Binary Pattern (LBP) and Gray Level Co-occurrence Matrix (GLCM), followed by Firefly feature selection, enabled effective classification of healthy vs. abnormal breast tissue.	Despite strong performance with SVM, kNN (81.5%) and D-Tree (94.5%) yielded lower accuracy, showing model performance is sensitive to classifier choice and may not generalize equally. Additionally, the small dataset size (200 IR images) limits the robustness of the model across larger, real-world clinical settings.	Future studies should expand testing on larger and more diverse IR thermogram datasets to improve model reliability. Integrating deep learning approaches or hybrid models could further enhance feature extraction and classification accuracy from infrared breast images.

8.6.2. Summary

Thermal imaging has emerged as a promising, cost-effective, and radiation-free modality for breast cancer (BC) detection. Recent advancements in deep learning (DL) have significantly improved the accuracy of BC detection using thermal images. Studies have employed techniques such as multi-view thermograms, transfer learning, feature selection, and deep segmentation models. Notable results include Chatterjee et al.’s (2022) [152] 100% accuracy using a two-stage model with VGG16 and the Dragonfly Algorithm, Civilibal et al.’s (2023) 97.1% accuracy with Mask R-CNN and ResNet [154], and Mahoro et al.’s (2024) 97.26% accuracy using TransUNet [156]. Additionally, Husaini et al. (2024) [157] achieved 99.82% accuracy with noise mitigation strategies. These advancements demonstrate the potential of DL-based thermal imaging for accurate and non-invasive BC screening. Table 37 presents a summary of key findings in thermal imaging-based breast cancer detection.

Table 37. Summary of key findings in thermal imaging-based breast cancer detection.

Aspect	Details
Datasets	MIAS Thermogram, DMR-IR, COCO, DMR, etc.
Models	VGG16, Mask R-CNN, ResNet, TransUNet, Inception MV4
Techniques	Multi-view thermograms, transfer learning, feature selection, noise mitigation
Highest Accuracy	100% (Chatterjee et al., 2022 [152]
Key Challenges	Dataset variability, computational complexity, noise reduction

8.6.3. Analysis

The integration of multi-view thermograms and clinical data, along with the use of transfer learning and advanced feature extraction techniques, has led to improved accuracy rates. Segmentation models, like TransUNet, and feature selection algorithms, such as the Dragonfly Algorithm, enhance model performance. Noise reduction techniques have proven essential, highlighting the importance of preprocessing in thermal imaging. Despite these advancements, challenges in dataset variability, model architectures, and preprocessing techniques remain, hindering standardization. Table 38 provides an analysis of techniques and their impact on thermal imaging-based breast cancer detection.

Table 38. Analysis of techniques and their impact on thermal imaging-based breast cancer detection.

Technique	Impact
Multi-View Thermograms	Provides a more comprehensive analysis compared to single-view approaches.
Transfer Learning	Improves performance on small datasets, reduces training time.
Feature Selection	Reduces feature dimensionality while maintaining accuracy.
Noise Mitigation	Enhances image quality and model reliability.
Segmentation Models	Improves precision in tumor localization and classification.

8.6.4. Critique

While accuracy rates are impressive, several limitations need to be addressed. First, the reliance on small or proprietary datasets limits generalizability. Second, some claims, such as 100% accuracy, may indicate overfitting. Third, the lack of standardized datasets and preprocessing techniques complicates comparisons across studies. Fourth, models like EfficientNet-B7 and TransUNet have high computational demands, limiting their real-time application. Finally, more comprehensive studies are needed on noise and thermal image quality to ensure robust and reliable BC detection. Table 39 presents a critique of current limitations in thermal imaging-based breast cancer detection.

Table 39. Critique of current limitations in thermal imaging-based breast cancer detection.

Limitation	Details
Dataset Dependency	Reliance on small or proprietary datasets limits generalizability.
Overfitting	High accuracy claims may indicate overfitting to specific datasets.
Standardization	Lack of standardized datasets and preprocessing techniques.
Computational Complexity	High resource requirements hinder real-world deployment.
Noise Reduction	Insufficient focus on noise and image quality in some studies.

8.6.5. Comparison

The reviewed studies demonstrate varying approaches and outcomes, as summarized in Table 40. Chatterjee et al. (2022) [152] achieved 100% accuracy using a two-stage model with VGG16 and the Dragonfly Algorithm, significantly reducing feature dimensionality. Civilibal et al. (2023) [154] employed Mask R-CNN with ResNet-50, achieving 97.1% accuracy and precise tumor localization. Mahoro et al. (2024) [156] utilized TransUNet for segmentation, achieving 97.26% accuracy. Husaini et al. (2024) [157] emphasized noise mitigation, achieving 99.82% accuracy with Inception MV4. These comparisons highlight the diversity of approaches in thermal imaging-based BC detection, with each method offering unique advantages.

Table 40. Comparison of key studies in thermal imaging-based breast cancer detection.

Study	Dataset	Model/Technique	Accuracy	Key Findings
Chatterjee et al. (2022) [152]	DMR-IR	VGG16 + Dragonfly Algorithm	100%	Reduces feature dimensionality by 82%.
Civilibal et al. (2023) [154]	COCO	Mask R-CNN with ResNet-50	97.1%	Precise tumor localization and classification.
Mahoro et al. (2024) [156]	Not mentioned	TransUNet	97.26%	Effective segmentation and classification.
Husaini et al. (2024) [157]	Not mentioned	Inception MV4	99.82%	Emphasizes noise mitigation and preprocessing.

Thermal imaging combined with deep learning models has significantly advanced breast cancer detection, achieving high accuracy and precision. While models like VGG16, Mask R-CNN, and TransUNet demonstrate exceptional performance, challenges related to dataset dependency, computational complexity, and noise reduction remain. Future research should focus on improving model interpretability, expanding dataset diversity, and developing computationally efficient models for real-world clinical deployment. Addressing these challenges will enhance the clinical applicability and robustness of thermal imaging-based BC detection systems.

Two critical gaps persist in thermal imaging-based breast cancer detection: (1) generalizability remains limited due to reliance on small or proprietary datasets, (2) real-time deployment is hindered by the high computational costs of models like TransUNet and EfficientNet.

8.7. CSV Dataset-Based Breast Cancer Detection

The proposed method using the WDBC dataset applies preprocessing and employs 13 significant features with a gradient boosting regressor and Bonferroni correction. The eXtreme Gradient Boosting model achieved 99.12% accuracy, 0.9767 precision, 1.0 recall, 0.9861 specificity, and a 0.9882 F1-score, demonstrating superior performance and reliability [162]. Manikandan et al. utilized the SEER dataset to classify BC patients’ survival status using a two-step feature selection method (Variance Threshold and PCA) and various classifiers, including Decision Tree, AdaBoosting, XGBoosting, Gradient Boosting, and Naive Bayes. The Decision Tree model achieved 98% accuracy when evaluated with train–test split and k-fold cross-validation, outperforming other approaches [163]. Albadr et al. proposed the Online Sequential Extreme Learning Machine (OSELM) approach to enhance diagnostic accuracy. This method was applied to the WDBC (Wisconsin Diagnostic Breast Cancer) dataset and achieved impressive results, with a precision of 94.09%, a recall of 95.57%, an accuracy of 96.13%, a G-Mean of 94.82%, an F-Measure of 94.80%, a specificity of 96.51%, and an MCC of 91.76%. The results indicate that OSELM is a reliable and efficient technique for BC diagnosis [164]. Table 41 provides a summary of BC detection methods using CSV-based data.

Table 41. An overview of contemporary approaches for BC detection utilizing CSV datasets.

Ref.	Year	Dataset	Instances	Proposed Model	Approach	Results
[162]	2024	WDBC	569	XGBoost	Preprocessing, feature engineering, feature selection using a gradient boosting regressor with Bonferroni correction, and classification.	Accuracy: 99.12% Precision: 0.9767 Recall:1.0 Specificity: 0.9861 F1-score: 0.9882.
[163]	2023	SEER	712319	Decision Tree	Two-step feature selection (Variance Threshold, PCA) and classification using Decision Tree	Accuracy: 98% Precision, Recall, Specificity and F1-score: 0.98
[164]	2024	WDBC	569	OSELM	Data preprocessing, selection and classificaiton	Accuracy: 96.13% Precision: 94.09% Recall: 95.57%

Figure 25 displays the highest accuracy rates attained by various models on CSV datasets in 2023 and 2024, including WDBC [162], SEER [163], BC-UCI [165], Breast Cancer Coimbra [165], and Cuban [166], as detailed in the referenced studies, along with a correlation figure that visualizes accuracy levels through changes in color intensity.

Figure 25. Highest detection accuracy across multiple CSV datasets for breast cancer.

8.7.1. Model Performance on Tabular Data: Pros, Cons, and Future Directions

Table 42 provides a comparison of the top-performing models, highlighting their primary advantages, drawbacks, and prospective research directions in breast cancer detection using tabular data.

Table 42. Detailed review of high-performing models applied to tabular breast cancer datasets.

Ref.	Advantages	Limitations	Future Work Recommendations
[165]	The proposed models—CRO+SVM and GA+RF—demonstrated superior performance across all three datasets: WBC (569 samples), BC-UCI (286 samples), and BCC (116 samples). Notably, CRO+SVM achieved 99.64% accuracy on the WBC dataset, while GA+RF reached 98.00% on BCC and CRO+XGBoost scored 99.12% on BC-UCI, outperforming existing models in accuracy, F1 score, precision, recall, and specificity.	The model’s performance may be influenced by class imbalance and preprocessing techniques in the datasets, potentially limiting real-world applicability. Additionally, while the study focuses on improving predictive performance, it lacks the integration of explainable AI techniques for interpretability in clinical settings.	Future work should integrate explainable AI methods to enhance trust and transparency in clinical decision-making. Collaboration with domain experts and clinicians is encouraged to develop models that are both technically robust and clinically relevant.
[166]	The proposed model using the Cuban dataset (n = 1697) achieved high performance with a Random Forest predictor showing 0.996 training accuracy and 0.997 AUC, surpassing traditional models. It is also cost-effective, requiring only standard hardware and easily integrable into clinical systems.	The study is limited by the size and scope of the dataset (1697 instances from one Cuban hospital), which may not fully represent the wider Cuban population. Additionally, due to the lack of historical follow-up data, the model only supports current diagnosis prediction, not long-term risk estimation.	Future work will focus on expanding data collection across multiple hospitals to enhance the generalizability and accuracy of predictions. Additionally, the integration of new risk factors such as breast density and type of occupation will be explored in collaboration with healthcare professionals.
[162]	Using the WDBC dataset (569 samples), the proposed eXtreme Gradient Boosting (XGBoost) model achieved 99.12% accuracy, 1.0 recall, 0.9861 specificity, and 0.9882 F1-score, outperforming state-of-the-art techniques. The model also demonstrated superior computational efficiency, being 2.11 × faster than Extra Trees in the original feature space.	The model was evaluated solely on the WDBC dataset, which may limit its generalizability to other datasets with different feature distributions or imaging data. Additionally, the dataset’s binary classification nature restricts the method’s applicability in multi-class or real-time clinical scenarios.	Future work should focus on parameter optimization to further enhance performance and adapt the model for more diverse datasets. Moreover, expanding the approach to multi-class classification and integrating real-world clinical data could significantly improve its practical utility.

8.7.2. Summary

Recent advancements in breast cancer (BC) detection using CSV datasets have demonstrated high accuracy and reliability. The eXtreme Gradient Boosting (XGBoost) model applied to the WDBC dataset achieved 99.12% accuracy, with excellent precision, recall, specificity, and F1-score [162]. Manikandan et al. (2023) [163] utilized the SEER dataset and a two-step feature selection method, achieving 98% accuracy with a Decision Tree model. Albadr et al. (2024) [164] proposed the Online Sequential Extreme Learning Machine (OSELM) approach, achieving 96.13% accuracy on the WDBC dataset. These results highlight the effectiveness of feature selection, preprocessing, and advanced machine learning models in improving BC detection accuracy. Table 43 presents a summary of key findings in CSV-based breast cancer detection.

Table 43. Summary of key findings in CSV-based breast cancer detection.

Aspect	Details
Datasets	WDBC, SEER, BC-UCI, Breast Cancer Coimbra, Cuban
Models	XGBoost, Decision Tree, OSELM
Techniques	Feature selection, preprocessing, gradient boosting, ensemble learning
Highest Accuracy	99.12% (XGBoost on WDBC)
Key Challenges	Dataset dependency, computational complexity, generalizability

8.7.3. Analysis

The integration of feature selection and preprocessing techniques has significantly improved the accuracy of BC detection using CSV datasets. Rahman et al. (2024) [162] employed gradient boosting with Bonferroni correction, achieving 99.12% accuracy on the WDBC dataset. Manikandan et al. (2023) [163] utilized a two-step feature selection method (Variance Threshold and PCA) with a Decision Tree model, achieving 98% accuracy on the SEER dataset. Albadr et al. (2024) [164] demonstrated the effectiveness of OSELM, achieving 96.13% accuracy on the WDBC dataset. These approaches underscore the importance of feature engineering and model optimization in achieving high classification accuracy. Table 44 provides an analysis of techniques and their impact on CSV-based breast cancer detection.

Table 44. Analysis of techniques and their impact on CSV-based Bbeast cancer detection.

Technique	Impact
Feature Selection	Reduces dimensionality and improves model performance.
Preprocessing	Enhances data quality and model accuracy.
Gradient Boosting	High accuracy and robustness in classification tasks.
Ensemble Learning	Combines strengths of multiple models, improving robustness and accuracy.
Online Learning	Efficient for real-time applications and large datasets.

8.7.4. Critique

Despite the impressive results, several limitations need to be addressed. First, the reliance on limited datasets, such as WDBC and SEER, may affect the generalizability of the models. Second, the computational complexity of methods like XGBoost and OSELM may hinder their real-time application. Third, the lack of standardized preprocessing techniques complicates comparisons across studies. Finally, more comprehensive validation on larger and more diverse datasets is needed to confirm the robustness of these models across different populations. Table 45 presents a critique of current limitations in CSV-based breast cancer detection.

Table 45. Critique of current limitations in CSV-based breast cancer detection.

Limitation	Details
Dataset Dependency	Reliance on limited datasets may affect generalizability.
Computational Complexity	High resource requirements hinder real-world deployment.
Standardization	Lack of standardized preprocessing techniques.
Generalizability	Further validation on diverse datasets is needed.
Real-Time Application	Computational complexity may limit real-time use.

8.7.5. Comparison

The reviewed studies demonstrate varying approaches and outcomes, as summarized in Table 46. Rahman et al. (2024) [162] achieved the highest accuracy (99.12%) using XGBoost on the WDBC dataset. Manikandan et al. (2023) [163] employed a Decision Tree model on the SEER dataset, achieving 98% accuracy. Albadr et al. (2024) [164] proposed OSELM, achieving 96.13% accuracy on the WDBC dataset. While the Decision Tree model has a simpler architecture, XGBoost and OSELM offer better performance, particularly in terms of precision and recall. These comparisons highlight the diversity of approaches in CSV-based BC detection, with each method offering unique advantages.

Table 46. Comparison of key studies in CSV-based breast cancer detection.

Study	Dataset	Model/Technique	Accuracy	Key Findings
Rahman et al. (2024) [162]	WDBC	XGBoost	99.12%	High accuracy and robustness.
Manikandan et al. (2023) [163]	SEER	Decision Tree	98%	Simple architecture with competitive accuracy.
Albadr et al. (2024) [164]	WDBC	OSELM	96.13%	Efficient for real-time applications.

CSV-based breast cancer detection has seen significant advancements through the integration of feature selection, preprocessing, and advanced machine learning models. While models like XGBoost and Decision Tree demonstrate exceptional performance, challenges related to dataset dependency, computational complexity, and generalizability remain. Future research should focus on improving model interpretability, expanding dataset diversity, and developing computationally efficient models for real-world clinical deployment. Addressing these challenges will enhance the clinical applicability and robustness of CSV-based BC detection systems.

Four persistent challenges emerge in CSV-based breast cancer detection: (1) limited generalizability due to small or homogeneous datasets like WDBC and SEER, (2) computational inefficiency of high-performing models (e.g., XGBoost) for real-time clinical use, (3) lack of standardized feature selection and preprocessing pipelines hindering reproducibility, (4) minimal clinical integration, with few models validated against multi-center or prospective data.

9. Imaging Techniques and Data in Breast Cancer Detection: Applications, Strengths, and Weaknesses

Table 47 provides an overview of the various medical imaging modalities used in breast cancer detection, along with corresponding CSV data that highlight their advantages, disadvantages, and applications. Understanding these factors is crucial for effective BC detection, as selecting the appropriate imaging technique and data can significantly impact early diagnosis, treatment planning, and patient outcomes.

Table 47. Imaging modalities—applications, strengths, and weaknesses.

Image Modality and Tabular Data	Advantages	Disadvantages	Applications
Mammography [167,168,169]	Mammography offers high accuracy, detailed breast tissue views, and detects cancers in dense breast tissue. Advanced methods, such as tomosynthesis, improve sensitivity, with mammography machines widely accessible.	Mammography is expensive, involves higher amounts of radiation, and can be uncomfortable. False positives/negatives are a risk, and thus it is not recommended for women with dense breasts or younger patients.	Mammography image-based breast cancer detection is used to produce 2D and 3D images of the breasts, detect early signs of cancer, and monitor changes over time.
Ultrasound [169,170]	Safe, accurate, low-cost, and free from radiation; suitable for young patients and those with dense breast tissue. The necessary equipment is readily available in most hospitals.	Requires expert radiologists, as accuracy depends on their expertise. May miss malignant tumors or calcifications, with high rates of false positives and negatives.	Suitable for assessing both soft and dense breast tissues. Identifies malignant and benign areas to reduce the need for unnecessary biopsies.
MRI [23,170]	Offers greater sensitivity than ultrasound and is effective in identifying suspicious masses in high-risk individuals. Capable of detecting multifocal lesions, and breast tissue density does not affect its diagnostic performance.	MRI is expensive, has limited specificity, and requires a skilled radiologist. MRI equipment is only available in specialized hospitals, and long scans may raise body temperature.	Detects abnormalities to guide biopsy procedures and screens for issues in dense breast tissue. Used to monitor treatment progress and track changes in breast tissue over time.
CT [168,171]	The technique has minimal motion artifacts and is simple to standardize for consistent results.	CT scans can produce false negatives, struggle to differentiate between malignant and fibrous tissue, and may not accurately locate all tumors in the body.	CT scans are not commonly used for breast assessment, but can evaluate cancer spread to the chest wall and determine the suitability of a mastectomy to remove malignancies.
Histopathological [23,170]	Whole slide imaging allows for complete tissue analysis, minimizes the risk of missing early-stage cancer by evaluating multiple regions, and provides detailed and accurate visual insights.	Manual analysis is time-consuming, demands high proficiency, and is subject to misdiagnosis due to color variation and inconsistent staining methods.	Multi-color imaging is used to diagnose a variety of cancer types, not just malignancies.
Thermal imaging [155,172]	Thermal imaging offers a safe, radiation-free approach that detects subtle temperature variations that may indicate early-stage breast cancer.	Thermal imaging struggles with limited specificity and can be impacted by external factors, including physical activity, stress, or room temperature, which may compromise its diagnostic accuracy.	Thermal imaging detects subtle changes in temperature linked to early-stage breast cancer, allowing for timely diagnosis and treatment.
Tabular Data [173,174]	Easy to process and user-friendly for data analysis	Missing or incorrect data can affect outcomes	Tabular data helps predict breast cancer risk by analyzing patient patterns.

Model Interpretability in Clinical Practice

Despite achieving high accuracy, deep learning (DL) models often encounter skepticism due to their opaque, black-box nature of decision-making. In imaging modalities such as ultrasound and MRI, models like HoVer-Transformer and DLR have integrated attention mechanisms to generate visual explanations, but still rely heavily on radiologist validation to ensure reliability. Similarly, in histopathological analysis, architectures such as BCHI-CovNet incorporate self-attention to enhance interpretability, though their real-world clinical adoption remains limited. A persistent trade-off exists—simpler models like SVM applied to structured CSV data offer greater interpretability but typically lag behind in accuracy. Moving forward, it is essential to strike a balance between performance and transparency by leveraging explainable AI (XAI) frameworks.

10. Trends in Breast Cancer Detection: ML, Statistical, and DL Techniques Across Journals

Breast cancer detection is currently the leading type of cancer worldwide, garnering significant attention from researchers aiming to develop efficient detection systems. Figure 26 illustrates the number of papers published from 2022 to 2024 across different journals, with a focus on the keyword “Breast Cancer Detection” in the article title. As shown in the figure, the IEEE journal published the highest number of articles with this keyword. The year 2023 saw a peak, with 138 publications. Moreover, research in BC detection has increased over the years, with 187 publications in 2022, 284 in 2023, and 303 in 2024, an increase of 38.28% in 2024 compared to 2022 and 6.27% compared to 2023.

Figure 26. Publications on breast cancer detection in different journals from 2022 to 2024.

11. Conclusions

This review highlights the remarkable advancements in breast cancer (BC) detection achieved through the integration of machine learning (ML) and deep learning (DL) technologies. Recent studies from 2023 to 2024 have demonstrated the transformative impact of DL models, particularly convolutional neural networks (CNNs), in improving diagnostic accuracy across various imaging modalities, including mammography, ultrasound, MRI, CT scans, histopathology, and thermal imaging. These advancements have enabled earlier and more reliable detection, significantly contributing to improved patient outcomes and survival rates.

However, the clinical adoption of these models hinges on overcoming their ’black-box’ nature. While DL achieves high accuracy, interpretability remains a barrier to trust among physicians. Future work must prioritize explainable AI (XAI) techniques—such as attention maps, Grad-CAM, and clinician feedback loops—to align model decisions with medical reasoning and ensure transparent diagnostics.

Beyond technological innovations, this review underscores the critical role of lifestyle interventions and risk factor management in BC prevention and control. Reducing red meat consumption, maintaining a healthy body mass index (BMI), addressing high fasting plasma glucose, limiting alcohol use, increasing physical activity, and avoiding smoking and secondhand smoke exposure are essential strategies for reducing BC risk. A comprehensive understanding of genetic, hormonal, and environmental risk factors further enhances early detection and prevention efforts.

While technological advancements have revolutionized BC detection, a holistic approach that combines cutting-edge diagnostic tools with lifestyle modifications and early detection strategies remains vital for effective BC management. This review provides valuable insights into the current state of BC detection technologies and emphasizes the need for continued research to refine these tools, improve diagnostic accuracy, and reduce the global burden of breast cancer. By integrating technological innovation with preventive healthcare, we can move closer to a future where breast cancer is detected early, treated effectively, and, ultimately, prevented.

12. Future Research Directions

To advance deep learning applications in breast cancer detection and management, future research should focus on the following key areas:

• Exploring advanced data augmentation techniques remains essential to address the challenge of limited medical imaging data in breast cancer detection. While traditional augmentation methods can enhance dataset diversity, they may compromise data integrity, potentially affecting model reliability. Future research should focus on developing and refining augmentation strategies, such as implicit data augmentation and novel generative approaches, to ensure both diversity and authenticity in the augmented data. Additionally, integrating these methods with deep learning models, including CNNs and transfer learning architectures, can further improve diagnostic accuracy and robustness, ultimately aiding in more reliable breast cancer detection [175,176].
• Develop deep learning models that are trained on diverse, high-quality datasets to mitigate biases and improve generalizability across various clinical settings. This will ensure that models perform consistently across different populations and healthcare systems.
• Prioritize XAI techniques (e.g., counterfactual explanations, concept activation vectors) to align model decisions with clinical reasoning. For example, integrating radiologist feedback loops during training could improve trust in CNN-based mammography systems [78].
• Design user-friendly tools and interfaces that seamlessly integrate deep learning models into existing healthcare practices. Adequate training for clinicians on AI-driven tools will be essential to ensure their adoption and practical utility.
• Focus on early breast cancer diagnosis by integrating longitudinal and multi-modal data to predict risk trajectories and enable preventive measures, thereby allowing for timely interventions.
• Foster collaboration among researchers, healthcare institutions, and data scientists with the goal of sharing datasets. Data should be collected from multiple sources to avoid biases and be properly formatted for ease of analysis, ensuring that deep learning models are trained on diverse and high-quality data.
• Develop advanced image preprocessing techniques for noise reduction, contrast enhancement, and image normalization. These techniques will improve image quality, leading to more accurate detection and classification.
• Innovate new segmentation architectures to improve the extraction of the region of interest (ROI) in medical images. Better segmentation will enhance the overall system performance and accuracy in identifying cancerous regions.
• Integrate various medical image modalities such as mammography, ultrasound, MRI, CT scans, histopathology, and thermal imaging to improve the accuracy and reliability of breast cancer detection. The synthesis of these modalities can provide a more comprehensive view of the disease.
• Incorporate non-imaging data such as cancer history, age, and other health issues alongside imaging data. This combination will allow for earlier and more accurate diagnosis, particularly for high-risk individuals.
• Conduct rigorous validation of deep learning models in real-world clinical settings, with the involvement of medical professionals. This ensures that the models are relevant, applicable, and safe for clinical use.
• Future studies should (1) develop fault-tolerant cloud architectures for reliable deployment, (2) integrate blockchain/IoT for secure health data workflows, and (3) validate AI monitoring systems in real-world clinical settings to enhance diagnostic pipelines [177,178,179].

By addressing these key areas, future research will be able to overcome current limitations, ensuring that deep learning can achieve its full potential in transforming breast cancer diagnosis, treatment, and patient care.