Skip to Content
MDPIMDPI

Advanced

CancersCancers
PDF
  • Review
  • Open Access

13 March 2026

31 Pages

Clinical AI in Radiology: Foundations, Trends, Applications, and Emerging Directions

,
,
,
,
,
,
,
,
and
1
Department of Machine Learning, H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL 33612, USA
2
Department of Electrical Engineering, University of South Florida, Tampa, FL 33620, USA
3
Department of Health Outcomes and Behavior, H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL 33612, USA
4
Department of Cancer Epidemiology, H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL 33612, USA
Cancers2026, 18(6), 942;https://doi.org/10.3390/cancers18060942
This article belongs to the Special Issue Advances in Medical Imaging for Cancer Detection and Diagnosis
Version Notes
Order Reprints

Simple Summary

Artificial intelligence (AI) is increasingly being explored in radiology to assist with image interpretation, organize clinical information, and improve workflow efficiency. This review summarizes the foundations of clinical AI in radiology and highlights current trends shaping its development and use in medical imaging. Several representative examples are presented, including locally deployed language models that help structure radiology reports, multimodal AI approaches for early detection of cachexia, federated learning systems that enable collaboration across institutions without sharing patient data, and automated methods for removing identifying information from radiology images and reports. Emerging AI-driven directions are also discussed, including tumor board decision support, clinical trial matching, radiology report quality assurance, and the development of an imaging complexity index. Together, these developments illustrate how AI is gradually being integrated into radiology practice while maintaining clinician oversight and protecting patient privacy.

Abstract

Artificial intelligence (AI) is at the vanguard of transforming radiology in several ways, including augmenting diagnoses, improving workflows, and increasing operational efficiency. Several integration challenges, including concerns over privacy, clinical usability, and workflow compatibility, still remain. This review discusses the foundations and current trends of clinical AI in radiology to provide essential context for ongoing developments. To illustrate translational potential, we describe representative applications, including: (1) local deployment of large language models (LLMs) for restructuring and streamlining radiology reports, improving clarity and consistency without relying on external resources; (2) multimodal AI frameworks combining CT images, clinical data, laboratory biomarkers, and LLM-extracted features from clinical notes for early detection of cachexia in pancreatic cancer; (3) privacy-preserving federated learning (FL) infrastructure enabling collaborative AI model development across institutions without sharing raw patient data; and (4) an uncertainty-aware de-identification pipeline for removing Protected Health Information (PHI) from radiology images and clinical reports to support secure data analysis and sharing. We further discuss emerging opportunities for tumor board decision support, clinical trial matching, radiology report quality assurance, and the development of an imaging complexity index. Collectively, these applications highlight the importance of local deployment, multimodal reasoning, privacy preservation, and human-in-the-loop oversight in translating AI models from research to oncology radiology practice.

1. Introduction

Radiology plays a central role in oncology, including early detection, diagnosis, staging, treatment monitoring, and post-therapy surveillance. As medical imaging volumes grow and case complexity increases, radiologists face mounting demands for accurate and rapid interpretation, efficient reporting, and multidisciplinary communication [1]. Artificial intelligence (AI) offers a promising solution to support radiologists across these workflows [2]. Large language models (LLMs), such as GPT-4/5, Vicuna, and Rad-Phi2, have been applied to radiology-specific tasks, including structured reporting, automated impression filling, extraction of key diagnostic information from unstructured text, and question answering [3,4,5]. Vision-based models, like U-Net, YOLO, and ResNet, have been increasingly used in radiology image segmentation, detection, and classification tasks [6,7,8]. Advances in multimodal learning and vision language models (VLMs) have introduced new capabilities for combining imaging information with clinical text to enhance diagnostic reasoning and support informed decision-making. Models such as Flamingo CXR (2024) [9], RadVLM (2025) [10], Med3D VLM (2025) [11], RadAlign (2025) [12], and RadZero (2025) [13] enable tasks including automated report generation, visual question answering (VQA), and prognosis prediction, highlighting their potential in oncology and beyond [14,15]. Alongside these systems, several radiology foundation models (FMs) have begun to emerge. RadFM (2023) [16], CURIA (2025) [17], and ONCOPILOT (2025) [18] provide broader, transferable representations that can support a wide range of downstream radiology and oncology tasks, reflecting a shift toward more general purpose radiology AI. As multimodal models continue to grow in scale, standardized preprocessing pipelines are essential for consistent image quality and reproducible model performance. Frameworks such as Pillar-0 provide modular tools for large-scale preprocessing, intensity normalization, and dataset curation, which support the development of both unimodal and multimodal radiology FMs [19].
Despite rapid technical advances, many radiology teams remain cautious about integrating AI tools into daily practice [2]. AI development often occurs in research environments and may not always align with the realities of clinical radiology workflows [20]. Concerns over opaque “black-box” decision-making remain a major barrier to clinical trust [21]. As a result, models may lack the usability, reliability, generalizability, and transparency required at the point-of-care, with diagnostic accuracy issues and the risk of automation bias further limiting adoption. In oncology imaging specifically, where complex, multi-modality assessments and nuanced interpretations are routine, the absence of cancer-focused, radiologist-centered AI tools exacerbates these challenges [22]. Privacy regulations and institutional data governance frameworks may restrict the use of cloud-based solutions, while some models remain proprietary and unsuitable for customization or local deployment. Consequently, AI adoption in radiology remains measured and fragmented, despite radiologists’ interest in tools that may help reduce reporting burden, support diagnostic confidence, and improve workflow efficiency.
Comprehensive cancer centers are placing increasing emphasis on AI solutions that address radiology needs through secure local deployment and clinician-centered design. At institutions such as the H. Lee Moffitt Cancer Center & Research Institute, these considerations have informed the development and evaluation of AI tools tailored to radiology workflows in oncologic practice. Rather than relying on generic commercial platforms, these efforts prioritize adapting state-of-the-art methods to the specific challenges of cancer imaging, including the high prevalence of multi-modality studies, the importance of longitudinal follow-up in oncology care, and the need for clear communication with oncologists. Local deployment ensures that model inference occurs entirely within institutional firewalls, thereby reducing privacy risks and strengthening compliance with regulatory standards. Equally important, clinician input is embedded at every stage of development, from prompt design for report restructuring to validation of multimodal prognostic models, so that tools align with established reading patterns and reporting standards. This emphasis on workflow integration and privacy preservation supports the development of AI systems that are both technically advanced and operationally feasible.
While local deployment and institution-centered development offer advantages for privacy, customization, and workflow integration, they may raise concerns about feasibility for smaller hospitals or institutions without dedicated machine learning teams. In practice, many radiology AI applications can operate on modest infrastructure, such as a workstation with a single GPU or CPU-based inference for smaller models. Hardware demands can be further reduced using lightweight approaches such as quantized or distilled models. Institutions may also adopt AI gradually, beginning with simple tools such as locally deployed inference models or reporting assistants, while components such as model training or federated learning coordination can be shared across institutions.
This manuscript presents a thematic narrative review of key domains shaping clinical AI development and deployment in radiology, drawing on the contemporary literature and selected translational examples to synthesize current trends and implementation considerations. We first review the foundations and current trends in clinical AI for radiology, highlighting advances in language models, computer vision, and multimodal vision–language systems, followed by developments in federated learning, considerations for clinical readiness and translation, and practical deployment pathways (Section 2). We then describe illustrative applications demonstrating practical integration across structured reporting, multimodal cachexia prediction, federated learning, and PHI/PII-redaction workflows (Section 3). Finally, we discuss emerging directions in radiology AI, including tumor board decision support, clinical trial matching, radiology report quality assurance, and imaging complexity index development (Section 4). Our aim is to offer actionable insights for radiology departments seeking to responsibly adopt and integrate AI technologies to enhance diagnostic quality, workflow efficiency, and patient care. The overall organization of the paper is summarized in Figure 1.
Figure 1. Overview of the paper. The manuscript is structured around three primary components: (1) Foundations and current trends in radiology AI, (2) illustrative applications, and (3) emerging directions in radiology AI.

3. Illustrative Applications

The following applications are presented as illustrative case studies demonstrating how AI systems can be developed and deployed within real-world clinical environments. While these examples reflect implementations within a specific institutional context, they highlight transferable design principles for clinical AI development and deployment. For each use case we discuss practical considerations such as compute requirements, data scale, evaluation strategies, and deployment constraints to support adaptation across institutions with different infrastructure and regulatory environments.

3.1. Structured Radiology Reporting Using Local LLMs

Radiology reporting plays an essential role in cancer care by guiding diagnostic interpretation, staging, and treatment planning. However, radiology reports can often be lengthy, formatted in various ways, and variable between individual radiologists [139]. This lack of standardization poses challenges for referring oncologists, who may be interested in quickly extracting key diagnostic information to inform patient management. Furthermore, equally important is the fact that these radiology reports will serve as a data point for future AI model training. Consistently structured reports therefore become an important asset for future AI development. We identified the need for a scalable and privacy-preserving solution to improve the clarity, conciseness, and structural consistency of radiology reports, particularly in the context of high-volume oncology imaging.
To address this need, we developed and implemented an LLM-based radiology report processing pipeline aimed at restructuring and streamlining radiology reports [33]. A central design priority was safeguarding patient privacy by ensuring that all model inference occurred locally, behind our institutional firewall. This local deployment strategy allowed us to avoid reliance on external cloud-based APIs, thus eliminating the risks associated with transmitting PHI outside institutional control.
Our technical workflow involved evaluating several open-weight state-of-the-art LLMs, such as Mixtral [140], Mistral [141], and Llama 3 [142]. All models were deployed on institution-owned workstations equipped with standard hardware configurations, including RTX 3060 GPUs with 12 GB of VRAM. The Ollama framework facilitated efficient local inference, and LangChain was used to implement flexible prompt engineering workflows tailored to radiology-specific tasks.
A key component of our development process involved optimizing prompt strategies to balance two primary goals: reducing report verbosity and ensuring its structural consistency. In collaboration with board-certified body radiologists, we designed and tested five distinct prompting strategies. Among these, the two-step Conciseness then Structure (C >> S) strategy consistently produced the most reliable results. This approach first condensed the original report to eliminate redundancy and unnecessary phrases, followed by a second step that reorganized the content into a standardized, organ-based template. We also incorporated an automated output validation step using an OutputFixingParser, which prompted the LLM to correct its formatting if initial outputs did not follow the desired structure.
In a retrospective quality improvement study, we applied the LLM pipeline to 814 de-identified radiology reports covering CT scans of the chest, abdomen, and pelvis from cancer patients. These reports were authored by seven board-certified radiologists at our institution. The results demonstrated a reduction in report length, with an average word count decrease exceeding 53% across all reports. The LLM-processed reports also exhibited greater structural uniformity, potentially improving the accessibility of key findings for referring oncologists. Preliminary feedback from radiologists underscored the perceived benefits of the tool in enhancing report readability and communication with clinical teams. However, this work was not designed as a formal clinical validation study; therefore, we did not perform blinded reader assessments, structured semantic error analysis, or quantitative evaluation of clinical decision-making impact.
Building on these initial findings, we have recently expanded our LLM-based radiology report pipeline to incorporate reasoning-focused LLM capabilities. A recognized limitation of current LLM outputs is the absence of transparent, interpretable reasoning behind the generated content. To address this, we now employ reasoning-enabled models, including DeepSeek [101] and QwQ [143]. These LLMs extract and present intermediate reasoning steps during report transformation to the user for transparency and validation. These reasoning traces include justifications for why certain findings were categorized under specific organ systems. By exposing this internal decision-making process, we aim to establish a more interactive AI–human feedback loop. While reducing redundancy remains beneficial, in subsequent iterations the emphasis shifted away from conciseness after recognizing that excessive compression may risk loss of important clinical information, particularly in complex oncology imaging. As a result, preserving critical clinical detail and minimizing the risk of unintended information loss became central design priorities. Currently, three board-certified body radiologists are conducting a review of both the final structured reports and the underlying reasoning pathways, to provide feedback for informing further model refinement and prompt optimization.
Through this iterative approach, we are advancing toward a more explainable and collaborative paradigm of AI deployment in radiology. The lessons learned from this application, particularly around local deployment feasibility, prompt strategy design, and clinician–AI interaction, are informing broader efforts toward responsible AI development in radiology reporting. This experience highlights the feasibility of developing and refining locally deployed LLM-based tools for radiology reporting workflows in oncology settings.
For institutions interested in adopting a similar approach, the pipeline relies on prompt engineering with open-weight LLMs in inference mode and does not require model training or fine-tuning. Adoption therefore involves defining a report structure appropriate for the specific CT study type, as different CT examinations (e.g., chest, abdomen/pelvis, angiography) require distinct organizational frameworks. Institutions may begin with the structure presented in [33] for chest, abdomen, and pelvis CT and refine or redesign templates and prompts to accommodate other study types. Smaller or quantized models can run on a single GPU workstation (e.g., 12 GB VRAM), and in some cases CPU-based inference may be feasible depending on model size and performance expectations. Because no model training is required, large-scale datasets are not necessary; instead, institutions should validate the pipeline on a representative sample of reports for each CT type to assess structural consistency and information preservation. Before routine use, local testing with radiologists and institutional IT teams is recommended to confirm that outputs align with existing reporting practices and data governance requirements, with ongoing periodic review after deployment.

3.2. Imaging-Informed AI for Early Cancer Cachexia Diagnosis

Cancer cachexia, a multifactorial metabolic syndrome characterized by substantial weight loss, skeletal muscle wasting, and systemic inflammation, remains one of the underdiagnosed oncology conditions despite its profound impact on patient survival and quality of life. Up to 80% of cancer patients are affected by cachexia with high prevalence in certain cancers, such as gastroesophageal, pancreatic, colorectal, lung, and hematological [144,145,146]. Existing diagnostic mechanisms are often limited by manual processes, fixed thresholds, and poor integration into clinical workflows [147,148,149]. To address this gap, we have developed clinically deployable AI tools to bridge this gap, focusing on reliability, interpretability, and workflow compatibility to bring imaging intelligence directly into patient care.
Skeletal Muscle Assessment–Automated and Reliable Tool based on AI (SMAART-AI) introduces an uncertainty-aware imaging data processing framework for automated skeletal muscle segmentation from routine CT scans. Rather than optimizing for accuracy alone, SMAART-AI prioritizes reliability and trustworthiness, a critical step in clinical deployment [150,151]. Unlike existing commercial segmentation tools that may silently degrade under input data domain shifts (e.g., differences in scanner type, protocol, image quality, population demographics, or events like a pandemic), SMAART-AI quantifies prediction uncertainty and flags potentially unreliable outputs, enabling HITL validation. Training strategies and reliability-based design make the tool robust in heterogeneous cancer cohorts. SMAART-AI also supports longitudinal tracking of skeletal muscle area (SMA) and skeletal muscle index (SMI), providing insight into changes throughout the treatment journey. The imaging-derived metrics are further integrated with structured and unstructured patient data for downstream tasks such as cachexia risk and survival prediction.
SMAART-AI implements a standardized CT preprocessing and segmentation workflow in which axial CT series are identified, the L3 vertebral level is localized, and a 2D nnU-Net-based model segments skeletal muscle on the selected L3 slice to compute SMA and height-normalized SMI. To improve robustness and quantify epistemic uncertainty, segmentation uses a deep ensemble trained with 5-fold cross-validation across two independently trained architectures. Specifically, for each fold, both architectures are trained from random initialization, yielding a 10-model ensemble (5 folds × 2 architectures), which improves prediction stability under limited labeled training data and heterogeneous imaging conditions. The models were trained on gastroesophageal and pancreatic L3 CT images and evaluated on held-out test images not used during training. At inference, the final muscle mask is obtained by pixel-wise averaging across all ensemble members, and uncertainty is computed from the variance across ensemble outputs, enabling reliability triage and expert review for flagged cases. The pipeline was evaluated across four heterogeneous cohorts, including gastroesophageal, colorectal, pancreatic, and ovarian cancer, and benchmarked against available reference standards, primarily expert manual SliceOmatic annotations and, where available, clinical/commercial tool measurements, ensuring that downstream multimodal analyses use quality-controlled SMA/SMI features under real-world imaging variability.
Building on SMAART-AI, we developed a “Multimodal AI Biomarker” framework that extends radiologic insight into a broader, data-driven pipeline for early cachexia diagnosis (see Figure 3). The Multimodal AI Biomarker fuses CT-derived skeletal muscle metrics with structured clinical variables (e.g., age, sex, race, ethnicity, weight, height, BMI, and stage of disease), blood-based laboratory biomarkers (e.g., serum albumin, neutrophil, lymphocyte count, blood, urea, nitrogen, creatinine, and derived ratios such as neutrophil-to-lymphocyte and blood-urea-to-creatinine), and cachexia-related symptoms extracted from unstructured clinical notes using locally deployed LLMs. The framework adapts dynamically to patient-specific factors and missing data modalities, ensuring compatibility with real-world clinical data and readiness for potential clinical deployment.
Figure 3. Multimodal AI workflow for cachexia prediction. The framework integrates imaging-derived skeletal muscle metrics generated by SMAART-AI with structured clinical variables, laboratory biomarkers, and cachexia-related symptoms extracted from unstructured physician notes using locally deployed LLMs. Each data modality is processed through dedicated feature extraction modules and fused into a unified multimodal representation for cachexia classification and survival prediction.
A retrospective dataset of 236 patients with pancreatic cancer from the Florida Pancreas Collaborative cohort was used to build the multi-modal model. To reduce overfitting given the modest cohort size and high-dimensional multimodal embeddings, we used a strict evaluation protocol with an independent held-out test set that was not used during model selection. Cross-validation and hyperparameter selection were performed only within the training split (Cohort I: 10-fold CV on n = 210; Cohort II: 7-fold CV on n = 105), and final performance was reported on the held-out test set (n = 26 for each cohort). Hyperparameter optimization was restricted to training folds only. The fusion MLP includes a learned bottleneck (first hidden layer) that compresses the concatenated embeddings and functions as a data-driven dimensionality reduction step. Regularization was applied via dropout and L2 weight decay, and early stopping was used based on validation loss. Final patient-level predictions were obtained by averaging across folds and across five MLP configurations to reduce variance. Future work will evaluate the approach on external multi-institution and multi-cancer datasets for independent validation.
The results show that the step-wise addition of each data modality (imaging, clinical, lab, and physician’s notes) improved model performance, achieving 77% accuracy using clinical and imaging data, and 85% accuracy when laboratory and LLM-extracted symptom features were combined [152]. In a separate embedding-based implementation, CT image embeddings corresponding to the L3 level generated using the RadImageNet FM [153] were fused with embeddings from GatorTron [36] for clinical, laboratory, and free-test medical note data, enabling joint multimodal representation learning. Incorporating these radiology-derived image embeddings further enhanced cachexia prediction accuracy to 92%, underscoring the additive value of radiologic information in multimodal fusion [152]. Similarly, survival prediction improved progressively, with concordance index (C-index) gains from 0.64 using clinical data alone to 0.72 when skeletal muscle metrics, laboratory indicators, and clinical text features were added. Removing skeletal muscle metrics led to considerable declines in both cachexia diagnosis accuracy and survival prediction performance, highlighting the indispensable prognostic role of radiology-derived features [152].
From the clinical deployment standpoint, all model inferences and patient data processing, including LLM-based extraction from unstructured medical notes, are performed locally within secure computing infrastructure, avoiding reliance on external APIs and ensuring compliance with institutional data governance policies. The multimodal cachexia prediction model is intended to be portable across clinical sites because it uses routinely available inputs at diagnosis, such as CT-derived muscle biomarkers (SMA/SMI) and standard demographic/anthropometric and staging variables. Practical transfer requires consistent definition of cachexia labels, harmonization of baseline timepoints, and standardized CT series/slice selection. Inference can be performed with modest compute (CPU feasible), while model retraining benefits from GPU resources primarily for skeletal muscle segmentation. To support comparability across sites, we report performance using standardized classification metrics with confidence intervals and recommend calibration assessment and threshold selection aligned to the target site’s dataset distribution. External validation on independent multi-institution cohorts is an important next step to quantify generalizability under differing protocols and population shifts.
Together, SMAART-AI and the Multimodal AI Biomarker exemplify how imaging-derived information serves as a cornerstone for multimodal clinical AI, advancing early diagnosis, personalized intervention, and scalable cachexia management. Ongoing work is focused on expanding this framework to a multi-institutional settings through FL, supporting privacy-preserving model development across oncology centers. These efforts reflect a commitment to developing patient-centered, trustworthy AI systems designed for safe, transparent, and clinically integrated oncology practice.

3.3. Privacy-Preserving and Multi-Institutional AI Collaboration Using FL

We have implemented and validated a privacy-preserving FL framework to address the challenges of collaborative AI development between institutions while maintaining strict data governance standards. We implemented the NVIDIA Federated Learning Application Runtime Environment (FLARE) Software Development Kit (SDK) to enable distributed training of AI models without the need to share raw patient data. This approach is particularly well-suited for multi-institutional efforts focused on early cancer detection and the development of imaging-based biomarkers.
Initial FL simulations focused on lung cancer screening, using low-dose CT (LDCT) images from the National Lung Cancer Screening Trial (NLST) [154] and a cohort of real-world patients from our institution. This cohort varied significantly in population and characteristics of the two NLST cohorts, providing a real-world testbed to evaluate model generalization under heterogeneous conditions. To simplify onboarding across collaborating institutions, we developed a “One-Touch FL” framework in which each client site connects to the central FL server using a single Linux terminal command. This approach enables collaborators who may have limited FL experience to participate in FL runs by reducing the setup and configuration burden and mitigating the steep learning curve associated with establishing FL infrastructure. All model development orchestration, including local training, secure model updates, and aggregation, is handled centrally, reducing technical complexity at the partner sites.
Using a standard FedAvg strategy within NVIDIA FLARE, both centralized and federated models were trained on the NLST dataset to enable a direct comparison between centralized and federated training approaches. The FL global model achieved performance comparable to its centralized counterpart, demonstrating that federated training can achieve similar predictive performance to centralized model development while preserving data locality.
Next, the patient cohort from our institution was added as a third site, and a new FL model was trained across all three datasets. This model was then evaluated on a withheld test set of patients from our institution and directly compared against a centralized model trained solely on the local cohort. The task was binary cancer classification, with the model outputting either “yes” or “no”. Initially, the global FL model performed significantly worse than centralized model trained exclusively on the local dataset, underscoring a clear domain gap between the NLST data and the local cohort. To address this, the global model was fine-tuned on the local training data to better align it with the local distribution, a process known as personalized FL. These findings suggest that features learned from the NLST cohorts can be effectively leveraged in the local domain when the global model is properly adapted to site-specific data.
To evaluate NV-FLARE server–client communication and readiness for model training, we conducted cross-institutional experiments using the CIFAR-10 dataset. Training was performed between our institution and other FL sites, using a centrally hosted NV-FLARE server within a locally hosted AWS Virtual Private Cloud (VPC). This successful deployment confirmed the technical deployment of NV-FLARE for secure asynchronous federated training between geographically distributed sites. Currently, we incorporate multiple privacy preservation strategies, including secure server aggregation and optional integration of differential privacy or homomorphic encryption for enhanced protection (see Section 2.4 for discussion of additional privacy risks, attack vectors, and mitigation strategies). We are integrating several “uncertainty quantification” methods, including ensemble learning and conformal prediction, to enable model confidence estimation in distributed settings. This work is needed to enable HITL workflows in which uncertain predictions could be flagged for clinical review. Current efforts focus on scaling this the FL infrastructure to support the development of multimodal models between institutions that are part of the National Cancer Institute’s Early Detection Research Network (NCI’s EDRN) and the Privacy-Preserving Federated Learning (PPFL) consortium.
NVIDIA FLARE is an open-source software development kit designed to facilitate the implementation of FL infrastructure. The framework allows institutions to integrate existing machine learning pipelines into a federated training environment while maintaining local control of patient data. Multi-institutional FL collaborations still require appropriate data governance agreements between participating sites, which may vary depending on institutional policies. From a computational perspective, FL does not substantially increase local training requirements compared with centralized approaches. Each participating site performs standard model training, while a central aggregation server combines the model updates. The aggregation server does not require GPU resources and primarily performs parameter aggregation across sites. FL can operate with as few as two participating institutions and scale to large international collaborations. The Federated Tumor Segmentation Initiative (FeTS), for example, included 72 medical institutions across six continents, representing the largest medical deployment of federated learning to date [86]. By enabling models to learn from distributed datasets without transferring raw data, FL can particularly benefit institutions with smaller local datasets by leveraging the diversity and scale of multi-institutional data.

3.4. PHI/PII Redaction from Radiology Images and Reports

A critical barrier to secondary use of medical imaging data for AI model development and clinical research is the presence of PHI and Personally Identifiable Information (PII) in both DICOM metadata and pixel data. In collaboration with Impact Business Information Solutions (IBIS), we have developed a novel PHI/PII de-identification framework that addresses these challenges using a hybrid AI-driven and rule-based approach with integrated uncertainty quantification.
The framework employs a two-tiered pipeline for metadata and pixel data de-identification:
Metadata De-Identification: A rule-based system removes explicit PHI from DICOM headers. This is supplemented by a fine-tuned LLM-based Named Entity Recognition (NER) pipeline trained on synthetic clinical data. By simulating PHI instances using synthetic data, we ensure robust identification of sensitive fields without exposing real patient data. Additionally, fuzzy string matching techniques are applied to catch near-variants of detected PHI.
Pixel Data De-Identification: For image pixel data, an uncertainty-aware Faster R-CNN model detects and localizes burned-in text regions, a common but under-addressed privacy risk in medical imaging. Detected text regions undergo Optical Character Recognition (OCR) step, and the extracted text is processed through the same NER pipeline for PHI identification and removal. The model integrates uncertainty quantification, enabling the system to assess the confidence of each detection and flag uncertain cases for HITL review.
The AI-assisted hybrid workflow enhances both scalability and reliability. The use of uncertainty metrics provides transparency and allows for HITL verification when the model confidence is low. The de-identification solution has been benchmarked against regulatory standards including HIPAA, GDPR, and TCIA’s best practice guidelines. In external evaluation through the Medical Image Deidentification Benchmark Challenge (MIDI-B), across over 580,000 data elements evaluated, the system achieved over 99.8% de-identification accuracy, including successful removal of embedded PHI from diverse imaging modalities (CT, MRI, and X-ray) [155]. Quality assurance remains essential in PHI redaction workflows, as the primary concern is the occurrence of false negatives where PHI is not successfully removed. To mitigate this risk, the framework incorporates uncertainty-aware detection that flags low-confidence cases for HITL review prior to dataset release. The combination of automation and uncertainty-aware HITL review supports scalable, secure sharing of medical imaging datasets for AI development without compromising patient privacy.
A key innovation of this framework is its explicit risk calibration. Uncertain predictions are transparently quantified, enabling operational teams to set confidence thresholds, quarantine ambiguous data, and meet specific institutional or jurisdictional privacy requirements (such as “Expert Determination” under HIPAA).
The framework was designed to be institution-agnostic and deployable within local computing environments. It is currently implemented at several institutions, including Nemours, Harvard, Massachusetts General Hospital, and Emory, where it supports PHI redaction and dataset indexing workflows. Importantly, the system operates within institutional firewalls, allowing de-identification to occur without transferring data to external systems and ensuring that institutions retain full control of PHI under their existing governance frameworks. From a computational perspective, the system can operate across a range of infrastructure environments. GPU acceleration improves performance but is not required, and the pipeline can run on CPU-based systems with parallel processing to increase throughput.

4. Emerging Directions

4.1. AI-Enhanced Tumor Board Support

Medical-focused AI is gaining momentum as a decision-support tool for multidisciplinary tumor boards, where radiology provides a central foundation for oncologic care planning. By synthesizing imaging findings with pathology, genomics, and clinical records, AI has the potential to streamline case preparation, reduce administrative workload, and strengthen collaboration across specialties.
A recent study assessed 102 complex oncology cases, providing ChatGPT-4o with detailed summaries that included imaging, pathology, and prior treatment history [156]. While the model produced plausible recommendations, concordance with tumor board decisions remained modest, highlighting both the promise and the current limitations of LLM-based tumor board support. Similarly, one study tested GPT-3.5 Turbo on the data from 52 patients with non-small cell lung cancer. Each case summary included patient history, imaging reports, and guideline context. The LLM was tasked with recommending one management strategy. The agreement with the decisions of the tumor board reached 76% and the model showed a particularly high consistency in the surgical recommendations [157].
These studies demonstrate the early value of LLMs in handling clinical narratives, but they also reveal an important gap: neither incorporated radiology or pathology images together with text. So far, no published work has combined both imaging and text data in the tumor board setting. Advancing toward true multimodal integration of imaging and clinical data represents a key opportunity for medical AI. In parallel, contemporary oncology increasingly depends on molecular characterization. Radiogenomics links imaging phenotypes with underlying genomic and molecular alterations, providing insight into tumor biology beyond visual assessment alone. As next-generation sequencing becomes routine in cancer care, integrating radiologic features with genomic alterations, transcriptomic profiles, and other molecular markers within the clinical context will be essential. Future systems that can simultaneously interpret radiology, pathology, genomic, and clinical information will better capture the multidisciplinary character of tumor boards and provide more effective support for oncology decision making.

4.2. AI for Clinical Trial Matching

Clinical trial enrollment depends on accurately identifying patients who meet, often complex and cumbersome, eligibility rules, but this process is time-consuming is routinely done manually. AI models are now being developed to support this task by analyzing patient records along with the clinical trial enrollment criteria. A recent study showed that LLMs can effectively match patients to clinical trials by interpreting free-text eligibility criteria and comparing them with the patient’s EHR data, achieving better performance than traditional rule-based methods [158].
A promising direction for future work is to extend such approaches beyond textual records to include multimodal data sources such as diagnostic images and radiology reports. Since many eligibility criteria depend on imaging findings (e.g., tumor size, metastasis status, disease progression), VLMs that jointly reason over text and images could enable more comprehensive trial matching. Including imaging is important because reports may omit quantitative details, such as exact lesion measurements, vary by reporting style, or overlook subtle findings that are relevant to eligibility. Direct access to images allows AI systems to extract standardized biomarkers and verify reported impressions, resulting in more accurate and robust clinical trial-patient matching. Integrating these multimodal AI systems into clinical workflows would enable a major advancement in clinical trials.

4.3. AI-Driven Radiology Report Quality Assurance and Error Detection

Quality assurance and error detection are essential components in radiology workflows to maintain the accuracy of the radiology reports and protect patient safety. However, these tasks are both time-consuming and susceptible to human variability. LLMs are increasingly being applied to support this process by automatically detecting inconsistencies and potential errors in imaging findings. For example, a recent large-scale validation study demonstrated that GPT-4 could proofread head CT reports with high sensitivity to both factual and interpretive errors, achieving near-radiologist performance while greatly reducing review time [159].
More recently, multimodal approaches have been explored to cross-check reports directly against the underlying images [160]. A study evaluated VLMs for radiology report error detection by introducing synthetic changes, such as inserted or omitted findings, and found that models with access to chest radiographs and reports outperformed text-only systems. These findings suggest that introducing VLMs in radiology reporting workflows could move quality assurance from retrospective error detection to proactive, image-based validation that strengthens both accuracy and patient safety.

4.4. AI-Driven Complexity Indexing for Radiological Imaging

The U.S. radiology reimbursement system is built around CPT codes and RVUs, but these tools do not reflect the wide variation in interpretive difficulty that often exists among studies assigned the same code [161,162]. In practice, radiologists receive the same payment whether a case is straightforward or requires substantial cognitive effort [161]. Work that demands close comparison with multiple prior studies, careful reconstruction of treatment histories, or nuanced assessment of treatment response typically takes far longer and adds to the growing problem of professional fatigue and burnout [163,164]. With imaging volumes continuing to rise and staffing pressures becoming more acute, the absence of a way to represent case complexity has created a persistent mismatch between the work performed and the reimbursement assigned.
Recent advances in multimodal AI offer a way to address this gap by generating transparent, quantitative measures of interpretive complexity. These models can integrate patient comorbidities, the number and timing of prior examinations, treatment timelines, the specificity of the clinical question, and image-derived characteristics such as lesion burden, expected treatment effects, and post-operative changes. These elements can then be combined into a single complexity score that reflects the level of interpretive effort likely required. As decision-support tools become more integrated into clinical practice, a validated complexity index has the potential to give departments a clearer picture of workload, provide payers with a more appropriate basis for recognizing high-effort interpretations, and support more consistent, high-quality imaging assessment.

5. Conclusions

AI in radiology is steadily moving from experimental promise to practical application. Experience to date suggests that meaningful progress requires more than technical performance: local deployment, multimodal reasoning, data security, and human oversight are equally critical for building trust and ensuring clinical value. Emerging applications including multimodal AI for tumor board support, clinical trial–patient matching, automated detection of inconsistencies or errors in radiology reports, and the development of imaging complexity indices illustrate how AI can augment decision-making in ways that directly influence patient care. Future success will depend on transparent development, cross-disciplinary collaboration, and alignment with real-world radiology and oncology workflows, ensuring that AI strengthens both diagnostic quality and patient safety.

Author Contributions

Conceptualization, I.H., C.L., C.A. and G.R.; writing—original draft preparation, I.H., N.K., S.A. and G.R.; writing—review and editing, all authors; funding acquisition, G.R. and R.A.G. All authors have read and agreed to the published version of the manuscript.

Funding

This study was partly funded by the Department of Diagnostic Imaging and Interventional Radiology, Moffitt Cancer Center, Florida Biomedical Research Award (21B12), NIH/NCI award (U01-CA200464), and NSF Awards 2234836 and 2234468.

Data Availability Statement

No new data were created or analyzed in this study. Data sharing is not applicable to this article.

Acknowledgments

The authors gratefully acknowledge the Data Services Core at the Moffitt Cancer Center & Research Institute, an NCI-designated Comprehensive Cancer Center (P30-CA076292), for its collaborative support. We further acknowledge our ongoing collaboration with IBIS, Inc., which is funded by three NIH/NCI SBIR awards [75N91021C00042, 75N91023C00041 and 75N91024C00069], addressing the de-identification of PHI and PII, and advancing multimodal machine learning research.

Conflicts of Interest

The authors declare no conflicts of interest.

Abbreviations

The following abbreviations are used in this manuscript:
AIArtificial Intelligence
NLPNatural Language Processing
CNNConvolutional Neural Network
ViTVisual Transformer
LLMLarge Language Model
VLMVision Language Model
PHIProtected Health Information
EHRElectronic Health Record
VQAVisual Question Answering
FLFederated Learning
HITLHuman-in-the-loop

References

  1. Siewert, B.; Bruno, M.A.; Bourland, J.D.; Slanetz, P.J.; Guillerman, P.; Schwartz, E.S.; Paltiel, H.J.; Hublall, R.; Brook, O.R.; Scanlon, M.H.; et al. Seven Challenges in Radiology Practice: From Declining Reimbursement to Inadequate Labor Force: Summary of the 2023 ACR Intersociety Meeting. J. Am. Coll. Radiol. 2025, 22, 129–138. [Google Scholar] [CrossRef] [Scilit]
  2. Korfiatis, P.; Kline, T.L.; Meyer, H.M.; Khalid, S.; Leiner, T.; Loufek, B.T.; Blezek, D.; Vidal, D.E.; Hartman, R.P.; Joppa, L.J.; et al. Implementing Artificial Intelligence Algorithms in the Radiology Workflow: Challenges and Opportunities. Mayo Clin. Proc. Digit. Health 2025, 3, 100188. [Google Scholar] [CrossRef] [Scilit]
  3. Adams, L.C.; Truhn, D.; Busch, F.; Kader, A.; Niehues, S.M.; Makowski, M.R.; Bressem, K.K. Leveraging GPT-4 for Post Hoc Transformation of Free-Text Radiology Reports into Structured Reporting: A Multilingual Feasibility Study. Radiology 2023, 307, e230725. [Google Scholar] [CrossRef] [Scilit]
  4. Guellec, B.L.; Lefèvre, A.; Geay, C.; Shorten, L.; Bruge, C.; Hacein-Bey, L.; Amouyel, P.; Pruvo, J.P.; Kuchcinski, G.; Hamroun, A. Performance of an Open-Source Large Language Model in Extracting Information from Free-Text Radiology Reports. Radiol. Artif. Intell. 2024, 6, e230364. [Google Scholar] [CrossRef] [Scilit]
  5. Ranjit, M.; Ganapathy, G.; Srivastav, S.; Ganu, T.; Oruganti, S. RAD-PHI2: Instruction Tuning PHI-2 for Radiology. arXiv 2024, arXiv:2403.09725. [Google Scholar] [CrossRef] [Scilit]
  6. Ronneberger, O.; Fischer, P.; Brox, T. U-Net: Convolutional Networks for Biomedical Image Segmentation. In Medical Image Computing and Computer-Assisted Intervention—MICCAI; Springer: Berlin/Heidelberg, Germany, 2015. [Google Scholar] [CrossRef] [Scilit]
  7. Ragab, M.G.; Abdulkadir, S.J.; Muneer, A.; Alqushaibi, A.; Sumiea, E.H.; Qureshi, R.; Al-Selwi, S.M.; Alhussian, H. A Comprehensive Systematic Review of YOLO for Medical Object Detection (2018 to 2023). IEEE Access 2024, 12, 57815–57836. [Google Scholar] [CrossRef] [Scilit]
  8. Chen, C.; Isa, N.A.M.; Liu, X. A review of convolutional neural network based methods for medical image classification. Comput. Biol. Med. 2025, 185, 109507. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  9. Tanno, R.; Barrett, D.G.T.; Sellergren, A.; Ghaisas, S.; Dathathri, S.; See, A.; Welbl, J.; Lau, C.; Tu, T.; Azizi, S.; et al. Collaboration between clinicians and vision–language models in radiology report generation. Nat. Med. 2025, 31, 599–608. [Google Scholar] [CrossRef] [Scilit]
  10. Deperrois, N.; Matsuo, H.; Ruipérez-Campillo, S.; Vandenhirtz, M.; Laguna, S.; Ryser, A.; Fujimoto, K.; Nishio, M.; Sutter, T.M.; Vogt, J.E.; et al. RadVLM: A Multitask Conversational Vision-Language Model for Radiology. arXiv 2025, arXiv:2502.03333. [Google Scholar]
  11. Xin, Y.; Ates, G.C.; Gong, K.; Shao, W. Med3DVLM: An Efficient Vision-Language Model for 3D Medical Image Analysis. arXiv 2025, arXiv:2503.20047. [Google Scholar] [CrossRef] [Scilit]
  12. Gu, D.; Gao, Y.; Zhou, Y.; Zhou, M.; Metaxas, D. RadAlign: Advancing Radiology Report Generation with Vision-Language Concept Alignment. arXiv 2025, arXiv:2501.07525. [Google Scholar]
  13. Park, J.; Kim, S.; Yoon, B.; Choi, K. RadZero: Similarity-Based Cross-Attention for Explainable Vision-Language Alignment in Radiology with Zero-Shot Multi-Task Capability. arXiv 2025, arXiv:2504.07416. [Google Scholar]
  14. Hartsock, I.; Rasool, G. Vision-language models for medical report generation and visual question answering: A review. Front. Artif. Intell. 2024, 7, 1430984. [Google Scholar] [CrossRef] [Scilit]
  15. Yang, L.; Xu, S.; Sellergren, A.; Kohlberger, T.; Zhou, Y.; Ktena, I.; Kiraly, A.; Ahmed, F.; Hormozdiari, F.; Jaroensri, T.; et al. Advancing Multimodal Medical Capabilities of Gemini. arXiv 2024, arXiv:2405.03162. [Google Scholar] [CrossRef] [Scilit]
  16. Wu, C.; Zhang, X.; Zhang, Y.; Wang, Y.; Xie, W. Towards Generalist Foundation Model for Radiology by Leveraging Web-Scale 2D&3D Medical Data. arXiv 2023, arXiv:2308.02463. [Google Scholar]
  17. Dancette, C.; Khlaut, J.; Saporta, A.; Philippe, H.; Ferreres, E.; Callard, B.; Danielou, T.; Alberge, L.; Machado, L.; Tordjman, D.; et al. Curia: A Multi-Modal Foundation Model for Radiology. arXiv 2025, arXiv:2509.06830. [Google Scholar] [CrossRef] [Scilit]
  18. Machado, L.; Alberge, L.; Philippe, H.; Ferreres, E.; Khlaut, J.; Dupuis, J.; Le Floch, K.; Gatenyo, D.H.; Roux, P.; Grégory, J.; et al. A promptable CT foundation model for solid tumor evaluation. NPJ Precis. Oncol. 2025, 9, 121. [Google Scholar] [CrossRef] [Scilit]
  19. Agrawal, K.K.; Liu, L.; Lian, L.; Nercessian, M.; Harguindeguy, N.; Wu, Y.; Mikhael, P.; Lin, G.; Sequist, L.V.; Fintelmann, F.; et al. Pillar-0: A New Frontier for Radiology Foundation Models. arXiv 2025, arXiv:2511.17803. [Google Scholar]
  20. Gupta, V.; Erdal, B.; Ramirez, C.; Floca, R.; Genereaux, B.; Bryson, S.; Bridge, C.; Kleesiek, J.; Nensa, F.; Braren, R.; et al. Current State of Community-Driven Radiological AI Deployment in Medical Imaging. JMIR AI 2024, 3, e55833. [Google Scholar] [CrossRef] [Scilit]
  21. Marey, A.; Arjmand, P.; Alerab, A.D.S.; Eslami, M.J.; Saad, A.M.; Sanchez, N.; Umair, M. Explainability, transparency and black box challenges of AI in radiology: Impact on patient care in cardiovascular radiology. Egypt. J. Radiol. Nucl. Med. 2024, 55, 183. [Google Scholar] [CrossRef] [Scilit]
  22. Koh, D.; Papanikolaou, N.; Bick, U.; Illing, R.; Kahn, C.E., Jr.; Kalpathi-Cramer, J.; Matos, C.; Martí-Bonmatí, L.; Miles, A.; Mun, S.K.; et al. Artificial intelligence and machine learning in cancer imaging. Commun. Med. 2022, 2, 133. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  23. Khalifa, M.; Albadawy, M. AI in diagnostic imaging: Revolutionising accuracy and efficiency. Comput. Methods Programs Biomed. Update 2024, 5, 100146. [Google Scholar] [CrossRef] [Scilit]
  24. Pons, E.; Braun, L.M.M.; Hunink, M.G.M.; Kors, J.A. Natural Language Processing in Radiology: A Systematic Review. Radiology 2016, 279, 329–343. [Google Scholar] [CrossRef] [Scilit]
  25. Tariq, A.; Banerjee, I.; Trivedi, H.; Gichoya, J. Multimodal Artificial Intelligence Models for Radiology. BJR|Artif. Intell. 2025, 2, ubae017. [Google Scholar] [CrossRef] [Scilit]
  26. Sheller, M.; Edwards, B.; Reina, G.; Martin, J.; Pati, S.; Kotrotsou, A.; Milchenko, M.; Xu, W.; Marcus, D.; Colen, R.; et al. Federated learning in medicine: Facilitating multi-institutional collaborations without sharing patient data. Sci. Rep. 2020, 10, 12598. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  27. Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A.N.; Kaiser, L.; Polosukhin, I. Attention Is All You Need. In Advances in Neural Information Processing Systems; NeurIPS: San Diego, CA, USA, 2017; Volume 30, pp. 5998–6008. [Google Scholar]
  28. Devlin, J.; Chang, M.W.; Lee, K.; Toutanova, K. BERT: Pre-Training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics; Association for Computational Linguistics: Minneapolis, MN, USA, 2019; Volume 1, pp. 4171–4186. [Google Scholar] [CrossRef] [Scilit]
  29. D’Antonoli, T.A.; Stanzione, A.; Bluethgen, C.; Vernuccio, F.; Ugga, L.; Klontzas, M.E.; Cuocolo, R.; Cannella, R.; Koçak, B. Large language models in radiology: Fundamentals, applications, ethical considerations, risks, and future directions. Diagn. Interv. Radiol. 2024, 30, 80–90. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  30. Sahoo, P.; Singh, A.K.; Saha, S.; Jain, V.; Mondal, S.; Chadha, A. A Systematic Survey of Prompt Engineering in Large Language Models: Techniques and Applications. arXiv 2025, arXiv:2402.07927. [Google Scholar]
  31. Schwartz, L.H.; Panicek, D.M.; Berk, A.R.; Li, Y.; Hricak, H. Improving communication of diagnostic radiology findings through structured reporting. Radiology 2011, 260, 174–181. [Google Scholar] [CrossRef] [Scilit]
  32. Busch, F.; Hoffmann, L.; dos Santos, D.P.; Makowski, M.R.; Saba, L.; Prucker, P.; Hadamitzky, M.; Navab, N.; Kather, J.N.; Truhn, D.; et al. Large language models for structured reporting in radiology: Past, present, and future. Eur. Radiol. 2025, 35, 2589–2602. [Google Scholar] [CrossRef] [Scilit]
  33. Hartsock, I.; Araujo, C.; Folio, L.; Rasool, G. Improving Radiology Report Conciseness and Structure via Local Large Language Models. J. Imaging Inform. Med. 2026, 39, 1005–1016. [Google Scholar] [CrossRef] [Scilit]
  34. Reichenpfader, D.; Müller, H.; Denecke, K. A scoping review of large language model based approaches for information extraction from radiology reports. NPJ Digit. Med. 2024, 7, 222. [Google Scholar] [CrossRef] [Scilit]
  35. Soni, S.; Gudala, M.; Pajouhi, A.; Roberts, K. RadQA: A Question Answering Dataset to Improve Comprehension of Radiology Reports. In Proceedings of the Thirteenth Language Resources and Evaluation Conference; European Language Resources Association: Paris, France, 2022; pp. 6250–6259. [Google Scholar]
  36. Yang, X.; Chen, A.; PourNejatian, N.; Shin, H.C.; Smith, K.E.; Parisien, C.; Compas, C.; Martin, C.; Costa, A.B.; Flores, M.G.; et al. A large language model for electronic health records. npj Digit. Med. 2022, 5, 194. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  37. Kuckelman, I.J.; Yi, P.H.; Bui, M.; Onuh, I.; Anderson, J.A.; MD, A.B.R. Assessing AI-Powered Patient Education: A Case Study in Radiology. Acad. Radiol. 2024, 31, 338–342. [Google Scholar] [CrossRef] [Scilit]
  38. Ayers, J.W.; Poliak, A.; Dredze, M.; Leas, E.C.; Zhu, Z.; Kelley, J.B.; Faix, D.J.; Goodman, A.M.; Longhurst, C.A.; Hogarth, M.; et al. Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Forum. JAMA Intern. Med. 2023, 183, 589–596. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  39. Hussein, R.; Elkhateeb, A.; Soliman, M.; Yousry, O.; Khedr, H.; Elhawary, H. Evaluating the accuracy and reliability of AI chatbots in patient education on cardiovascular imaging. Egypt. J. Radiol. Nucl. Med. 2025, 56, 45. [Google Scholar] [CrossRef] [Scilit]
  40. Draelos, R.L.; Afreen, S.; Blasko, B.; Brazile, T.L.; Chase, N.; Desai, D.P.; Evert, J.; Gardner, H.L.; Herrmann, L.; House, A.V.; et al. Large language models provide unsafe answers to patient-posed medical questions. arXiv 2025, arXiv:2507.18905. [Google Scholar] [CrossRef] [Scilit]
  41. Çiçek, Ö.; Abdulkadir, A.; Lienkamp, S.S.; Brox, T.; Ronneberger, O. 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. arXiv 2016, arXiv:1606.06650. [Google Scholar] [CrossRef] [Scilit]
  42. Huang, X.; Deng, Z.; Li, D.; Yuan, X.; Fu, Y. MISSFormer: An Effective Transformer for 2D Medical Image Segmentation. IEEE Trans. Med. Imaging 2023, 42, 1484–1494. [Google Scholar] [CrossRef] [Scilit]
  43. Dai, Y.; Gao, Y.; Liu, F. TransMed: Transformers Advance Multi-Modal Medical Image Classification. Diagnostics 2021, 11, 1384. [Google Scholar] [CrossRef] [Scilit]
  44. Yamashita, R.; Nishio, M.; Do, R.; Togashi, K. Convolutional Neural Networks: An Overview and Application in Radiology. Insights Imaging 2018, 9, 611–629. [Google Scholar] [CrossRef] [Scilit]
  45. Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S.; et al. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. arXiv 2020, arXiv:2010.11929. [Google Scholar]
  46. Chen, J.; Mei, J.; Li, X.; Lu, Y.; Yu, Q.; Wei, Q.; Luo, X.; Xie, Y.; Adeli, E.; Wang, Y.; et al. TransUNet: Rethinking the U-Net architecture design for medical image segmentation through the lens of transformers. Med. Image Anal. 2024, 97, 103280. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  47. Hatamizadeh, A.; Tang, Y.; Nath, V.; Yang, D.; Myronenko, A.; Landman, B.; Roth, H.R.; Xu, D. UNETR: Transformers for 3D Medical Image Segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV); IEEE: New York, NY, USA, 2022; pp. 574–584. [Google Scholar]
  48. Litjens, G.; Kooi, T.; Bejnordi, B.E.; Setio, A.A.; Ciompi, F.; Ghafoorian, M.; Van Der Laak, J.A.; Van Ginneken, B.; Sánchez, C.I. A Survey on Deep Learning in Medical Image Analysis. Med. Image Anal. 2017, 42, 60–88. [Google Scholar] [CrossRef] [Scilit]
  49. Teng, Z.; Li, L.; Xin, Z.; Xiang, D.; Huang, J.; Zhou, H.; Shi, F.; Zhu, W.; Cai, J.; Peng, T.; et al. A literature review of artificial intelligence (AI) for medical image segmentation: From AI and explainable AI to trustworthy AI. Quant. Imaging Med. Surg. 2024, 14, 9620–9652. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  50. Abedalla, A.; Abdullah, M.; Al-Ayyoub, M.; Benkhelifa, E. Chest X-ray pneumothorax segmentation using U-Net with EfficientNet and ResNet architectures. PeerJ Comput. Sci. 2021, 7, e607. [Google Scholar] [CrossRef] [Scilit]
  51. Meshram, N.H.; Mitchell, C.C.; Wilbrand, S.; Dempsey, R.J.; Varghese, T. Deep Learning for Carotid Plaque Segmentation Using a Dilated U-Net Architecture. Ultrason. Imaging 2020, 42, 221–230. [Google Scholar] [CrossRef] [Scilit]
  52. Chiu, T.W.; Tsai, Y.L.; Su, S.F. Automatic detect lung node with deep learning in segmentation and imbalance data labeling. Sci. Rep. 2021, 11, 11174. [Google Scholar] [CrossRef] [Scilit]
  53. Zunair, H.; Hamza, A.B. Sharp U-Net: Depthwise convolutional network for biomedical image segmentation. Comput. Biol. Med. 2021, 136, 104699. [Google Scholar] [CrossRef] [Scilit]
  54. Hirsch, L.; Huang, Y.; Luo, S.; Saccarelli, C.R.; Gullo, R.L.; Naranjo, I.D.; Bitencourt, A.G.V.; Onishi, N.; Ko, E.S.; Leithner, D.; et al. Radiologist-Level Performance by Using Deep Learning for Segmentation of Breast Cancers on MRI Scans. Radiol. Artif. Intell. 2021, 4, e200231. [Google Scholar] [CrossRef] [Scilit]
  55. Azad, R.; Heidari, M.; Shariatnia, M.; Aghdam, E.K.; Karimijafarbigloo, S.; Adeli, E.; Merhof, D. TransDeepLab: Convolution-Free Transformer-based DeepLab v3+ for Medical Image Segmentation. arXiv 2022, arXiv:2208.00713. [Google Scholar]
  56. Li, S.; Sui, X.; Luo, X.; Xu, X.; Yong, L.; Goh, R.S.M. Medical Image Segmentation using Squeeze-and-Expansion Transformers. In Proceedings of the 30th International Joint Conference on Artificial Intelligence (IJCAI), Montreal, QC, Canada, 21–26 August 2021. [Google Scholar]
  57. Heidari, M.; Kazerouni, A.; Soltany, M.; Azad, R.; Aghdam, E.K.; Cohen-Adad, J.; Merhof, D. HiFormer: Hierarchical Multi-scale Representations Using Transformers for Medical Image Segmentation. In 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV); IEEE: New York, NY, USA, 2023; pp. 6191–6201. [Google Scholar] [CrossRef] [Scilit]
  58. Cheng, P.M.; Montagnon, E.; Yamashita, R.; Pan, I.; Cadrin-Chênevert, A.; Romero, F.P.; Chartrand, G.; Kadoury, S.; Tang, A. Deep Learning: An Update for Radiologists. RadioGraphics 2021, 41, 1427–1445. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  59. Tian, J.; Lee, S.; Kang, K. Faster R-CNN in Healthcare and Disease Detection: A Comprehensive Review. In 2025 International Conference on Electronics, Information, and Communication (ICEIC); IEEE: New York, NY, USA, 2025; pp. 1–6. [Google Scholar] [CrossRef] [Scilit]
  60. Lin, T.Y.; Goyal, P.; Girshick, R.; He, K.; Dollár, P. Focal Loss for Dense Object Detection. arXiv 2018, arXiv:1708.02002. [Google Scholar] [CrossRef] [Scilit]
  61. Gu, Y.; Lu, X.; Yang, L.; Zhang, B.; Yu, D.; Zhao, Y.; Gao, L.; Wu, L.; Zhou, T. Automatic lung nodule detection using a 3D deep convolutional neural network combined with a multi-scale prediction strategy in chest CTs. Comput. Biol. Med. 2018, 103, 220–231. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  62. Azad, R.; Heidari, M.; Shariatnia, M.; Aghdam, E.K.; Karimijafarbigloo, S.; Adeli, E.; Merhof, D. Advances in medical image analysis with vision Transformers: A comprehensive review. Med. Image Anal. 2024, 91, 103000. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  63. Rajpurkar, P.; Irvin, J.; Zhu, K.; Yang, B.; Mehta, H.; Duan, T.; Ding, D.; Bagul, A.; Langlotz, C.; Shpanskaya, K.; et al. CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning. arXiv 2017, arXiv:1711.05225. [Google Scholar]
  64. Gheflati, B.; Rivaz, H. Vision Transformers for Classification of Breast Ultrasound Images. In 2022 44th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC); IEEE: New York, NY, USA, 2022; pp. 480–483. [Google Scholar] [CrossRef] [Scilit]
  65. Gao, X.; Khan, M.H.M.; Hui, R.; Tian, Z.; Qian, Y.; Gao, A.; Baichoo, S. COVID-VIT: Classification of Covid-19 from 3D CT chest images based on vision transformer model. In 2022 3rd International Conference on Next Generation Computing Applications (NextComp); IEEE: New York, NY, USA, 2022; pp. 1–4. [Google Scholar] [CrossRef] [Scilit]
  66. Tanzi, L.; Audisio, A.; Cirrincione, G.; Aprato, A.; Vezzetti, E. Vision Transformer for femur fracture classification. Injury 2022, 53, 2625–2634. [Google Scholar] [CrossRef] [Scilit]
  67. Tajbakhsh, N.; Shin, J.Y.; Gurudu, S.R.; Hurst, R.T.; Kendall, C.B.; Gotway, M.B.; Liang, J. Convolutional Neural Networks for Medical Image Analysis: Full Training or Fine Tuning? IEEE Trans. Med. Imaging 2016, 35, 1299–1312. [Google Scholar] [CrossRef] [Scilit]
  68. Moor, M.; Huang, Q.; Wu, S.; Yasunaga, M.; Zakka, C.; Dalmia, Y.; Reis, E.P.; Rajpurkar, P.; Leskovec, J. Med-Flamingo: A Multimodal Medical Few-Shot Learner. arXiv 2023, arXiv:2307.15189. [Google Scholar]
  69. Pellegrini, C.; Özsoy, E.; Busam, B.; Navab, N.; Keicher, M. RaDialog: A Large Vision-Language Model for Radiology Report Generation and Conversational Assistance. arXiv 2023, arXiv:2311.18681. [Google Scholar]
  70. Li, P.; Liu, G.; He, J.; Zhao, Z.; Zhong, S. Masked Vision and Language Pre-Training with Unimodal and Multimodal Contrastive Losses for Medical Visual Question Answering. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2023; Springer: Berlin/Heidelberg, Germany, 2023; pp. 374–383. [Google Scholar] [CrossRef] [Scilit]
  71. Johnson, A.E.; Pollard, T.J.; Berkowitz, S.J.; Greenbaum, N.R.; Lungren, M.P.; Deng, C.y.; Mark, R.G.; Horng, S. MIMIC-CXR, a De-Identified Publicly Available Database of Chest Radiographs with Free-Text Reports. Sci. Data 2019, 6, 317. [Google Scholar] [CrossRef] [Scilit]
  72. Thawkar, O.; Shaker, A.; Mullappilly, S.S.; Cholakkal, H.; Anwer, R.M.; Khan, S.; Laaksonen, J.; Khan, F.S. XrayGPT: Chest Radiographs Summarization Using Medical Vision-Language Models. arXiv 2023, arXiv:2306.07971. [Google Scholar]
  73. Irvin, J.; Rajpurkar, P.; Ko, M.; Yu, Y.; Ciurea-Ilcus, S.; Chute, C.; Marklund, H.; Haghgoo, B.; Ball, R.; Shpanskaya, K.; et al. CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison. arXiv 2019, arXiv:1901.07031. [Google Scholar] [CrossRef] [Scilit]
  74. Vendrowa, E.; Schonfeldb, E. Understanding transfer learning for chest radiograph clinical report generation with modified transformer architectures. Heliyon 2023, 9, e17968. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  75. Ranjit, M.; Ganapathy, G.; Manuel, R.; Ganu, T. Retrieval Augmented Chest X-Ray Report Generation using OpenAI GPT models. PMLR 2023, 219, 650–666. [Google Scholar]
  76. Lau, J.J.; Gayen, S.; Ben Abacha, A.; Demner-Fushman, D. A Dataset of Clinically Generated Visual Questions and Answers about Radiology Images. Sci. Data 2018, 5, 180251. [Google Scholar] [CrossRef] [Scilit]
  77. Liu, B.; Zhan, L.M.; Xu, L.; Ma, L.; Yang, Y.F.; Wu, X.M. Slake: A Semantically-Labeled Knowledge-Enhanced Dataset for Medical Visual Question Answering. In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI); IEEE: New York, NY, USA, 2021; pp. 1650–1654. [Google Scholar] [CrossRef] [Scilit]
  78. Yuan, Z.; Jin, Q.; Tan, C.; Zhao, Z.; Yuan, H.; Huang, F.; Huang, S. RAMM: Retrieval-augmented Biomedical Visual Question Answering with Multi-Modal Pre-Training. In MM ’23: Proceedings of the 31st ACM International Conference on Multimedia; Association for Computing Machinery: New York, NY, USA, 2023. [Google Scholar]
  79. Zhong, Z.; Wang, Y.; Wu, J.; Hsu, W.C.; Somasundaram, V.; Bi, L.; Kulkarni, S.; Ma, Z.; Collins, S.; Baird, G.; et al. Vision-language model for report generation and outcome prediction in CT pulmonary angiogram. NPJ Digit. Med. 2025, 8, 432. [Google Scholar] [CrossRef] [Scilit]
  80. Blankemeier, L.; Cohen, J.P.; Kumar, A.; Van Veen, D.; Gardezi, S.J.S.; Paschali, M.; Chen, Z.; Delbrouck, J.B.; Reis, E.; Truyts, C.; et al. Merlin: A vision language foundation model for 3d computed tomography. Res. Sq. 2024. [Google Scholar] [CrossRef] [Scilit]
  81. Ankolekar, A.; Boie, S.; Abdollahyan, M.; Gadaleta, E.; Hasheminasab, S.A.; Yang, G.; Beauville, C.; Dikaios, N.; Kastis, G.A.; Bussmann, M.; et al. Advancing breast, lung and prostate cancer research with federated learning. A systematic review. NPJ Digit. Med. 2025, 8, 314. [Google Scholar] [CrossRef] [Scilit]
  82. Health Insurance Portability and Accountability Act of 1996. Pub. L. No. 104-191, 110 Stat. 1936. 1996. Available online: https://www.govinfo.gov/content/pkg/PLAW-104publ191/pdf/PLAW-104publ191.pdf (accessed on 10 January 2026).
  83. General Data Protection Regulation (GDPR). Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016. 2016. Available online: https://eur-lex.europa.eu/eli/reg/2016/679/oj (accessed on 10 January 2026).
  84. McMahan, B.; Moore, E.; Ramage, D.; Hampson, S.; Arcas, y.B.A. Communication-efficient learning of deep networks from decentralized data. PMLR 2017, 54, 1273–1282. [Google Scholar]
  85. Darzidehkalani, E.; Ghasemi-rad, M.; van Ooijen, P.V. Federated Learning in Medical Imaging: Part II: Methods, Challenges, and Considerations. J. Am. Coll. Radiol. 2022, 19, 755–765. [Google Scholar] [CrossRef] [Scilit]
  86. Pati, S.; Baid, U.; Zenk, M.; Edwards, B.; Sheller, M.; Reina, G.A.; Foley, P.; Gruzdev, A.; Martin, J.; Albarqouni, S.; et al. The Federated Tumor Segmentation (FeTS) Challenge. arXiv 2021, arXiv:2105.05874. [Google Scholar] [CrossRef] [Scilit]
  87. Che, L.; Zhang, W.; Zhou, M.; Wang, Y.; Ma, H. Multimodal Federated Learning: A Survey. Sensors 2023, 23, 6986. [Google Scholar] [CrossRef] [Scilit]
  88. Zhang, K.; Song, X.; Zhang, C.; Yu, C. Challenges and future directions of secure federated learning: A survey. Front. Comput. Sci. 2022, 16, 165817. [Google Scholar] [CrossRef] [Scilit]
  89. Wu, R.; Chen, X.; Guo, C.; Weinberger, K.Q. Learning to Invert: Simple Adaptive Attacks for Gradient Inversion in Federated Learning. arXiv 2022. [Google Scholar] [CrossRef] [Scilit]
  90. Jere, M.; Farnan, T.; Koushanfar, F. A Taxonomy of Attacks on Federated Learning. IEEE Secur. Priv. 2021, 19, 20–28. [Google Scholar] [CrossRef] [Scilit]
  91. Koutsoubis, N.; Waqas, A.; Yilmaz, Y.; Ramachandran, R.P.; Schabath, M.B.; Rasool, G. Privacy-preserving Federated Learning and Uncertainty Quantification in Medical Imaging. Radiol. Artif. Intell. 2025, 7, e240637. [Google Scholar] [CrossRef] [Scilit]
  92. Adnan, M.; Kalra, S.; Cresswell, J.C.; Taylor, G.W.; Tizhoosh, H. Federated learning and differential privacy for medical image analysis. Nature 2022, 12, 1953. [Google Scholar] [CrossRef] [Scilit]
  93. Wei, K.; Li, J.; Ding, M.; Ma, C.; Yang, H.H.; Farokhi, F.; Jin, S.; Quek, T.Q.S.; Vincent Poor, H. Federated Learning with Differential Privacy: Algorithms and Performance Analysis. IEEE Trans. Inf. Forensics Secur. 2020, 15, 3454–3469. [Google Scholar] [CrossRef] [Scilit]
  94. Nampalle, K.B.; Singh, P.; Narayan, U.; Raman, B. Vision Through the Veil: Differential Privacy in Federated Learning for Medical Image Classification. arXiv 2023, arXiv:2306.17794. [Google Scholar] [CrossRef] [Scilit]
  95. Acar, A.; Aksu, H.; Uluagac, A.S.; Conti, M. A Survey on Homomorphic Encryption Schemes: Theory and Implementation. arXiv 2017, arXiv:1704.03578. [Google Scholar] [CrossRef] [Scilit]
  96. Stripelis, D.; Saleem, H.; Ghai, T.; Dhinagar, N.J.; Gupta, U.; Anastasiou, C.; Steeg, G.V.; Ravi, S.; Naveed, M.; Thompson, P.M.; et al. Secure neuroimaging analysis using federated learning with homomorphic encryption. In SPIE 12088, 17th International Symposium on Medical Information Processing and Analysis; SPIE: Bellingham, WA, USA, 2021. [Google Scholar] [CrossRef] [Scilit]
  97. Jain, S.; Wallace, B.C. Attention is not Explanation. arXiv 2019, arXiv:1902.10186. [Google Scholar] [PubMed]
  98. Lundberg, S.M.; Lee, S.I. A Unified Approach to Interpreting Model Predictions. In Advances in Neural Information Processing Systems; Guyon, I., Luxburg, U.V., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., Garnett, R., Eds.; Curran Associates, Inc.: Red Hook, NY, USA, 2017; Volume 30. [Google Scholar]
  99. Goldshmidt, R.; Horovicz, M. TokenSHAP: Interpreting Large Language Models with Monte Carlo Shapley Value Estimation. arXiv 2024, arXiv:2407.10114. [Google Scholar] [CrossRef] [Scilit]
  100. Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Ichter, B.; Xia, F.; Chi, E.; Le, Q.; Zhou, D. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv 2023, arXiv:2201.11903. [Google Scholar]
  101. Guo, D.; Yang, D.; Zhang, H.; Song, J.; Zhang, R.; Xu, R.; Zhu, Q.; Ma, S.; Wang, P.; Bi, X.; et al. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv 2025, arXiv:2501.12948. [Google Scholar]
  102. Springenberg, J.T.; Dosovitskiy, A.; Brox, T.; Riedmiller, M. Striving for Simplicity: The All Convolutional Net. arXiv 2015, arXiv:1412.6806. [Google Scholar] [CrossRef] [Scilit]
  103. Zhou, B.; Khosla, A.; Lapedriza, A.; Oliva, A.; Torralba, A. Learning Deep Features for Discriminative Localization. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR); IEEE: New York, NY, USA, 2016; pp. 2921–2929. [Google Scholar] [CrossRef] [Scilit]
  104. Selvaraju, R.R.; Cogswell, M.; Das, A.; Vedantam, R.; Parikh, D.; Batra, D. Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. In 2017 IEEE International Conference on Computer Vision (ICCV); IEEE: New York, NY, USA, 2017; pp. 618–626. [Google Scholar] [CrossRef] [Scilit]
  105. Wang, H.; Wang, Z.; Du, M.; Yang, F.; Zhang, Z.; Ding, S.; Mardziel, P.P.; Hu, X. Score-CAM: Score-Weighted Visual Explanations for Convolutional Neural Networks. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW); IEEE: New York, NY, USA, 2019; pp. 111–119. [Google Scholar]
  106. Kim, B.; Wattenberg, M.; Gilmer, J.; Cai, C.J.; Wexler, J.; Viégas, F.B.; Sayres, R. Interpretability Beyond Feature Attribution: Quantitative Testing with Concept Activation Vectors (TCAV). PMLR 2018, 80, 2668–2677. [Google Scholar]
  107. Zhang, Y.; Hong, D.; McClement, D.; Oladosu, O.; Pridham, G.; Slaney, G. Grad-CAM helps interpret the deep learning models trained to classify multiple sclerosis types using clinical brain magnetic resonance imaging. J. Neurosci. Methods 2021, 353, 109098. [Google Scholar] [CrossRef] [Scilit]
  108. Alghamdi, A.; Alghamdi, K.; Alqahtani, M.; AboHejji, R. Enhance Breast Cancer Diagnosis Using Deep Learning Models on Mammogram Images in Saudi Arabia. In 2025 International Conference on Innovation in Artificial Intelligence and Internet of Things (AIIT); IEEE: New York, NY, USA, 2025; pp. 1–8. [Google Scholar] [CrossRef] [Scilit]
  109. Maksudov, B.; Curran, K.; Mileo, A. Towards generating more interpretable counterfactuals via concept vectors: A preliminary study on chest X-rays. arXiv 2025, arXiv:2506.04058. [Google Scholar] [CrossRef] [Scilit]
  110. Aflalo, E.; Du, M.; Tseng, S.Y.; Liu, Y.; Wu, C.; Duan, N.; Lal, V. VL-InterpreT: An Interactive Visualization Tool for Interpreting Vision-Language Transformers. arXiv 2022, arXiv:2203.17247. [Google Scholar]
  111. Parcalabescu, L.; Frank, A. MM-SHAP: A Performance-agnostic Metric for Measuring Multimodal Contributions in Vision and Language Models & Tasks. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers); Association for Computational Linguistics: Stroudsburg, PA, USA, 2023. [Google Scholar] [CrossRef] [Scilit]
  112. Goldshmidt, R. Attention, Please! PixelSHAP Reveals What Vision-Language Models Actually Focus On. arXiv 2025, arXiv:2503.06670. [Google Scholar]
  113. For trustworthy AI, keep the human in the loop. Nat. Med. 2025, 31, 3207. [CrossRef] [Scilit]
  114. Luo, M.; Yousefirizi, F.; Rouzrokh, P.; Jin, W.; Alberts, I.; Gowdy, C.; Bouchareb, Y.; Hamarneh, G.; Klyuzhin, I.; Rahmim, A. Physician-in-the-Loop Active Learning in Radiology Artificial Intelligence Workflows: Opportunities, Challenges, and Future Directions. Am. J. Roentgenol. 2025, 225, e2533364. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  115. Kaufmann, T.; Weng, P.; Bengs, V.; Hüllermeier, E. A Survey of Reinforcement Learning from Human Feedback. arXiv 2024, arXiv:2312.14925. [Google Scholar] [CrossRef] [Scilit]
  116. Acosta, J.N.; Dogra, S.; Adithan, S.; Wu, K.; Moritz, M.; Kwak, S.; Rajpurkar, P. The Impact of AI Assistance on Radiology Reporting: A Pilot Study Using Simulated AI Draft Reports. arXiv 2024, arXiv:2412.12042. [Google Scholar] [CrossRef] [Scilit]
  117. Awasthi, A.; Le, N.; Deng, Z.; Wu, C.C.; Nguyen, H.V. Enhancing Radiological Diagnosis: A Collaborative Approach Integrating AI and Human Expertise for Visual Miss Correction. arXiv 2024, arXiv:2406.19686. [Google Scholar] [CrossRef] [Scilit]
  118. Chow, J.; Lee, R.; Wu, H. How Do Radiologists Currently Monitor AI in Radiology and What Challenges Do They Face? An Interview Study and Qualitative Analysis. J. Imaging Inform. Med. 2026, 39, 6–19. [Google Scholar] [CrossRef] [Scilit]
  119. Blezek, D.J.; Olson-Williams, L.; Missert, A.; Korfiatis, P. AI Integration in the Clinical Workflow. J. Digit. Imaging 2021, 34, 1435–1446. [Google Scholar] [CrossRef] [Scilit]
  120. Kim, S.H.; Schramm, S.; Riedel, E.O.; Schmitzer, L.; Rosenkranz, E.; Kertels, O.; Bodden, J.; Paprottka, K.; Sepp, D.; Renz, M.; et al. Automation bias in AI-assisted detection of cerebral aneurysms on time-of-flight MR angiography. Radiol. Medica 2025, 130, 555–566. [Google Scholar] [CrossRef] [Scilit]
  121. Yu, A.C.; Mohajer, B.; Eng, J. External Validation of Deep Learning Algorithms for Radiologic Diagnosis: A Systematic Review. Radiol. Artif. Intell. 2022, 4, e210064. [Google Scholar] [CrossRef] [Scilit]
  122. Topff, L.; Groot Lipman, K.B.W.; Guffens, F.; Wittenberg, R.; Bartels-Rutten, A.; van Veenendaal, G.; Hess, M.; Lamerigts, K.; Wakkie, J.; Ranschaert, E.; et al. Is the generalizability of a developed artificial intelligence algorithm for COVID-19 on chest CT sufficient for clinical use? Results from the International Consortium for COVID-19 Imaging AI (ICOVAI). Eur. Radiol. 2023, 33, 4249–4258. [Google Scholar] [CrossRef] [Scilit]
  123. Wang, X.; Liang, G.; Zhang, Y.; Blanton, H.; Bessinger, Z.; Jacobs, N. Inconsistent Performance of Deep Learning Models on Mammogram Classification. J. Am. Coll. Radiol. 2020, 17, 796–803. [Google Scholar] [CrossRef] [Scilit]
  124. Zech, J.R.; Badgeley, M.A.; Liu, M.; Costa, A.B.; Titano, J.J.; Oermann, E.K. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: A cross-sectional study. PLoS Med. 2018, 15, e1002683. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  125. Larrazabal, A.J.; Nieto, N.; Peterson, V.; Milone, D.H.; Ferrante, E. Gender imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis. Proc. Natl. Acad. Sci. USA 2020, 117, 12592–12594. [Google Scholar] [CrossRef] [Scilit]
  126. Gichoya, J.W.; Banerjee, I.; Bhimireddy, A.R.; Burns, J.L.; Celi, L.A.; Chen, L.; Correa, R.; Dullerud, N.; Ghassemi, M.; Huang, S.; et al. AI recognition of patient race in medical imaging: A modelling study. Lancet Digit. Health 2022, 4, e406–e414. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  127. Linardos, A.; Kushibar, K.; Walsh, S.; Gkontra, P.; Lekadir, K. Federated learning for multi-center imaging diagnostics: A simulation study in cardiovascular disease. Sci. Rep. 2022, 12, 3551. [Google Scholar] [CrossRef] [Scilit]
  128. Sarma, K.V.; Harmon, S.; Sanford, T.; Roth, H.R.; Xu, Z.; Tetreault, J.; Xu, D.; Flores, M.G.; Raman, A.G.; Kulkarni, R.; et al. Federated learning improves site performance in multicenter deep learning without data sharing. J. Am. Med. Inform. Assoc. 2021, 28, 1259–1264. [Google Scholar] [CrossRef] [Scilit]
  129. Guan, H.; Liu, Y.; Yang, E.; Yap, P.; Shen, D.; Liu, M. Multi-site MRI harmonization via attention-guided deep domain adaptation for brain disorder identification. Med. Image Anal. 2021, 71, 102076. [Google Scholar] [CrossRef] [Scilit]
  130. Liu, S.; Yap, P. Learning multi-site harmonization of magnetic resonance images without traveling human phantoms. Commun. Eng. 2024, 3, 6. [Google Scholar] [CrossRef] [Scilit]
  131. Mallardi, G.; Calefato, F.; Lanubile, F.; Logroscino, G.; Tafuri, B. Diffusion Models for Neuroimaging Data Augmentation: Assessing Realism and Clinical Relevance. J. Med. Syst. 2025, 49, 161. [Google Scholar] [CrossRef] [Scilit]
  132. Li, Y.; Vasconcelos, N. REPAIR: Removing Representation Bias by Dataset Resampling. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); IEEE: New York, NY, USA, 2019; pp. 9564–9573. [Google Scholar] [CrossRef] [Scilit]
  133. Zhang, B.H.; Lemoine, B.; Mitchell, M. Mitigating Unwanted Biases with Adversarial Learning. In Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society; Association for Computing Machinery: New York, NY, USA, 2018; pp. 335–340. [Google Scholar] [CrossRef] [Scilit]
  134. Eichelberg, M.; Kleber, K.; Kämmerer, M. Cybersecurity in PACS and Medical Imaging: An Overview. J. Digit. Imaging 2020, 33, 1527–1542. [Google Scholar] [CrossRef] [Scilit]
  135. Nguyen, X.V.; Petscavage-Thomas, J.M.; Straus, C.M.; Ikuta, I. Cybersecurity in radiology: Cautionary Tales, Proactive Prevention, and What to do When You Get Hacked. Curr. Probl. Diagn. Radiol. 2025, 54, 245–250. [Google Scholar] [CrossRef] [Scilit]
  136. Pianykh, O.S.; Langs, G.; Dewey, M.; Enzmann, D.R.; Herold, C.J.; Schoenberg, S.O.; Brink, J.A. Continuous Learning AI in Radiology: Implementation Principles and Early Applications. Radiology 2020, 297, 6–14. [Google Scholar] [CrossRef] [Scilit]
  137. Strohm, L.; Hehakaya, C.; Ranschaert, E.R.; Boon, W.P.C.; Moors, E.H.M. Implementation of artificial intelligence (AI) applications in radiology: Hindering and facilitating factors. Eur. Radiol. 2020, 30, 5525–5532. [Google Scholar] [CrossRef] [Scilit]
  138. Kim, B.; Romeijn, S.; van Buchem, M.; Rezazade Mehrizi, M.H.; Grootjans, W. A holistic approach to implementing artificial intelligence in radiology. Insights Imaging 2024, 15, 22. [Google Scholar] [CrossRef] [Scilit]
  139. Wallis, A.; McCoubrie, P. The radiology report—Are we getting the message across? Clin. Radiol. 2011, 66, 1015–1022. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  140. Jiang, A.Q.; Sablayrolles, A.; Roux, A.; Mensch, A.; Savary, B.; Bamford, C.; Chaplot, D.S.; de las Casas, D.; Hanna, E.B.; Bressand, F.; et al. Mixtral of Experts. arXiv 2024, arXiv:2401.04088. [Google Scholar] [CrossRef] [Scilit]
  141. Jiang, A.Q.; Sablayrolles, A.; Mensch, A.; Bamford, C.; Chaplot, D.S.; de las Casas, D.; Bressand, F.; Lengyel, G.; Lample, G.; Saulnier, L.; et al. Mistral 7B. arXiv 2023, arXiv:2310.06825. [Google Scholar] [CrossRef] [Scilit]
  142. Grattafiori, A.; Dubey, A.; Jauhri, A.; Pandey, A.; Kadian, A.; Al-Dahle, A.; Letman, A.; Mathur, A.; Schelten, A.; Vaughan, A.; et al. The Llama 3 Herd of Models. arXiv 2024, arXiv:2407.21783. [Google Scholar] [CrossRef] [Scilit]
  143. Qwen Team. QwQ-32B: Embracing the Power of Reinforcement Learning; Qwen Team: Hangzhou, China, 2025. [Google Scholar]
  144. Park, M.A.; Whelan, C.J.; Ahmed, S.; Boeringer, T.; Brown, J.; Carson, T.L.; Crowder, S.L.; Gage, K.; Gregg, C.; Jeong, D.K.; et al. Defining and addressing research priorities in cancer cachexia through transdisciplinary collaboration. Cancers 2024, 16, 2364. [Google Scholar] [CrossRef] [Scilit]
  145. Han, J.; Harrison, L.; Patzelt, L.; Wu, M.; Junker, D.; Herzig, S.; Berriel Diaz, M.; Karampinos, D.C. Imaging modalities for diagnosis and monitoring of cancer cachexia. EJNMMI Res. 2021, 11, 94. [Google Scholar] [CrossRef] [Scilit]
  146. Babic, A.; Rosenthal, M.H.; Sundaresan, T.K.; Khalaf, N.; Lee, V.; Brais, L.K.; Loftus, M.; Caplan, L.; Denning, S.; Gurung, A.; et al. Adipose tissue and skeletal muscle wasting precede clinical diagnosis of pancreatic cancer. Nat. Commun. 2023, 14, 4317. [Google Scholar] [CrossRef] [Scilit]
  147. Go, S.I.; Park, M.J.; Park, S.; Kang, M.H.; Kim, H.G.; Kang, J.H.; Kim, J.H.; Lee, G.W. Cachexia index as a potential biomarker for cancer cachexia and a prognostic indicator in diffuse large B-cell lymphoma. J. Cachexia Sarcopenia Muscle 2021, 12, 2211–2219. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  148. Tan, S.; Xu, J.; Wang, J.; Zhang, Z.; Li, S.; Yan, M.; Tang, M.; Liu, H.; Zhuang, Q.; Xi, Q.; et al. Development and validation of a cancer cachexia risk score for digestive tract cancer patients before abdominal surgery. J. Cachexia Sarcopenia Muscle 2023, 14, 891–902. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  149. Argilés, J.M.; López-Soriano, F.J.; Toledo, M.; Betancourt, A.; Serpe, R.; Busquets, S. The cachexia score (CASCO): A new tool for staging cachectic cancer patients. J. Cachexia Sarcopenia Muscle 2011, 2, 87–93. [Google Scholar] [CrossRef] [Scilit]
  150. Ahmed, S.; Dera, D.; Hassan, S.U.; Bouaynaya, N.; Rasool, G. Failure detection in deep neural networks for medical imaging. Front. Med. Technol. 2022, 4, 919046. [Google Scholar] [CrossRef] [Scilit]
  151. Nowak, S.; Theis, M.; Wichtmann, B.D.; Faron, A.; Froelich, M.F.; Tollens, F.; Geißler, H.L.; Block, W.; Luetkens, J.A.; Attenberger, U.I.; et al. End-to-end automated body composition analyses with integrated quality control for opportunistic assessment of sarcopenia in CT. Eur. Radiol. 2022, 32, 3142–3151. [Google Scholar] [CrossRef] [Scilit]
  152. Ahmed, S.; Parker, N.; Park, M.; Davis, E.W.; Permuth, J.B.; Schabath, M.B.; Yilmaz, Y.; Rasool, G. Multimodal AI-driven Biomarker for Early Detection of Cancer Cachexia. arXiv 2025, arXiv:2503.06797. [Google Scholar] [CrossRef] [Scilit]
  153. Mei, X.; Liu, Z.; Robson, P.M.; Marinelli, B.; Huang, M.; Doshi, A.; Jacobi, A.; Cao, C.; Link, K.E.; Yang, T.; et al. RadImageNet: An Open Radiologic Deep Learning Research Dataset for Effective Transfer Learning. Radiol. Artif. Intell. 2022, 4, e210315. [Google Scholar] [CrossRef] [Scilit]
  154. National Lung Screening Trial Research Team. The National Lung Screening Trial: Overview and Study Design. Radiology 2011, 258, 243–253. [Google Scholar] [CrossRef] [Scilit]
  155. Pei, L.; Sutton, G.; Rutherford, M.; Wagner, U.; Nolan, T.; Smith, K.; Farmer, P.; Gu, P.; Rana, A.; Chen, K.; et al. Medical Image De-Identification Benchmark Challenge. arXiv 2025. [Google Scholar] [CrossRef] [Scilit]
  156. Karabuğa, B.; Karaçin, C.; Büyükkör, M.; Bayram, D.; Aydemir, E.; Kaya, O.B.; Yılmaz, M.E.; Çamöz, E.S.; Ergün, Y. The Role of Artificial Intelligence (ChatGPT-4o) in Supporting Tumor Board Decisions. J. Clin. Med. 2025, 14, 3535. [Google Scholar] [CrossRef] [Scilit]
  157. Zabaleta, J.; Aguinagalde, B.; Lopez, I.; Fernandez-Monge, A.; Lizarbe, J.A.; Mainer, M.; Ferrer-Bonsoms, J.A.; de Assas, M. Utility of Artificial Intelligence for Decision Making in Thoracic Multidisciplinary Tumor Boards. J. Clin. Med. 2025, 14, 399. [Google Scholar] [CrossRef] [Scilit]
  158. Jin, Q.; Wang, Z.; Floudas, C.S.; Chen, F.; Gong, C.; Bracken-Clarke, D.; Xue, E.; Yang, Y.; Sun, J.; Lu, Z. Matching patients to clinical trials with large language models. Nat. Commun. 2024, 15, 9074. [Google Scholar] [CrossRef] [Scilit] [PubMed]
  159. Kim, S.; Kim, D.; Shin, H.J.; Lee, S.H.; Kang, Y.; Jeong, S.; Kim, J.; Han, M.; Lee, S.; Kim, J.; et al. Large-Scale Validation of the Feasibility of GPT-4 as a Proofreading Tool for Head CT Reports. Radiology 2025, 314, e240701. [Google Scholar] [CrossRef] [Scilit]
  160. Wu, J.; Kim, Y.; Keller, E.C.; Chow, J.; Levine, A.P.; Pontikos, N.; Ibrahim, Z.; Taylor, P.; Williams, M.C.; Wu, H. Exploring Multimodal Large Language Models for Radiology Report Error-Checking. arXiv 2024, arXiv:2312.13103. [Google Scholar] [CrossRef] [Scilit]
  161. Chen, M.M.; Hirsch, J.A.; Lee, R.K.; Hughes, D.R.; Nicola, G.N.; Rosenkrantz, A.B. Determining the Patient Complexity of Head CT Examinations: Implications for Proper Valuation of a Critical Imaging Service. Curr. Probl. Diagn. Radiol. 2020, 49, 177–181. [Google Scholar] [CrossRef] [Scilit]
  162. Forsberg, D.; Rosipko, B.; Sunshine, J.L. Radiologists’ Variation of Time to Read Across Different Procedure Types. J. Digit. Imaging 2017, 30, 86–94. [Google Scholar] [CrossRef] [Scilit]
  163. McDonald, R.J.; Schwartz, K.M.; Eckel, L.J.; Diehn, F.E.; Hunt, C.H.; Bartholmai, B.J.; Erickson, B.J.; Kallmes, D.F. The Effects of Changes in Utilization and Technological Advancements of Cross-Sectional Imaging on Radiologist Workload. Acad. Radiol. 2015, 22, 1191–1198. [Google Scholar] [CrossRef] [Scilit]
  164. Chetlen, A.L.; Chan, T.L.; Ballard, D.H.; Frigini, L.A.; Hildebrand, A.; Kim, S.; Brian, J.M.; Krupinski, E.A.; Ganeshan, D. Addressing Burnout in Radiologists. Acad. Radiol. 2019, 26, 526–533. [Google Scholar] [CrossRef] [Scilit]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Article Metrics

Citations

Article Access Statistics

Journal Statistics
Multiple requests from the same IP address are counted as one view.
Spatial Immune Profiling and AI-Based Classifiers Identify Predictors of BCG Therapy Outcomes in High-Risk Non-Muscle-Invasive Bladder Cancer
Previous
Correction: Bourefis et al. HSP27/Menin Expression as New Prognostic Serum Biomarkers of Prostate Cancer Aggressiveness Independent of PSA. Cancers 2022, 14, 4773
Next