**1. Introduction**

According to International Agency for Research on Cancer, an estimated 2.3 million new cases of breast cancer has overtaken lung cancer as the most prevalent cancer diagnosed, with cancer death rates significantly higher in transitioning nations [1]. Breast screening programs are a way to detect early signs of breast cancer and are dominated by utilizing digital mammography as the primary tool for cancer detection [2]. Additional modalities such as ultrasound are used in conjunction with mammography for denser breasts, whereas magnetic resonance imaging (MRI) is used for more progressive breast analysis for repeated and high-risk patients [3].

Breast density, as defined by the American College of Radiology (ACR), is used during clinical diagnosis that classifies the breast into four categories with increasing density: almost entirely fatty, scattered fibroglandular, heterogenous, and finally, extremely dense breast [4].

**Citation:** Razali, N.F.; Isa, I.S.; Sulaiman, S.N.; Abdul Karim, N.K.; Osman, M.K.; Che Soh, Z.H. Enhancement Technique Based on the Breast Density Level for Mammogram for Computer-Aided Diagnosis. *Bioengineering* **2023**, *10*, 153. https://doi.org/10.3390/ bioengineering10020153

Academic Editors: Paolo Zaffino and Maria Francesca Spadea

Received: 28 November 2022 Revised: 4 January 2023 Accepted: 16 January 2023 Published: 23 January 2023

**Copyright:** © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

The heterogeneous dense breast as depicted in Figure 1A and the overlapped mass (red region) (in Figure 1B) on the dense region (blue region) is visually harder to distinguish compared to a non-dense breast that only contains mostly fatty (orange region) tissue. Diverse breast tissue structures cause mixed-intensity variations and limited visibility of breast features [5]. Due to this factor, the processed images may result in less acceptable breast tissue segmentation and inconsistent diagnosis by compromising the system's sensitivity and specificity to detect abnormalities [6,7]. Past studies concluded that mass detection decreased with increased density, due to the mass itself being similar to the surrounding dense tissue of the breast [8–10]. Additionally, image quality conditions also make it difficult to detect the lesion in dense breasts [11,12]. Specifying the edge of the mass from its surrounding dense tissue requires image processing that enhances the textural element of the image as one of the defining mass descriptors to assess a mammogram visually [13]. The textural analysis identifies distinctive descriptors in the form of a changing pattern or pixel intensity with various spatial arrangements. Its refinement aims to go beyond human-eye perception by defining semantic descriptors to extract quantitative radiological data [14].

**\$ %** 

**Figure 1.** (**A**) Original mammogram image example. (**B**) Mapped tissue region for the image on (**A**)—red: mass, green: dense tissue, orange: non-dense tissue.

To accommodate the analysis of mammographic mass, Computer-Aided Diagnosis (CAD) systems are introduced to breast cancer diagnosis stages, from improving the image quality [11,15], breast lesion detection, and segmentation [16], as well as benign or malignant classification [16–19]. Moreover, CAD implementation in mammography diagnostic could reduce the human rater's false-positive rate by 5.7% and false negative by 9.4%, as shown in a USA-based dataset [19], and an increase rate of 3% recall rate for a radiologist's mammogram analysis with CAD assistance for an expert radiologist [20]. CAD systems proved to aid radiologists in making a better diagnosis with the area under the curve (AUC) of 0.896 from 0.850 without affecting diagnosis timing [21]. Since deeplearning CAD systems performed best when trained using large datasets [22], it is harder to apply suitable image quality improvements individually on the images, leading to a need for special enhancement procedures and careful pre-processing for the images before they can be trained on a deep-learning architecture.

Most Convolutional Neural Network (CNN) applications for CAD systems have focused on direct mammogram images for detection and classification rather than the need for specific enhancement based on breast density level and the quality of the input images. This could unintentionally lead to reduced sensitivity for mass detection in dense mammograms, resulting from higher training weightage on non-dense breasts because of dataset class imbalance [23]. Enhancement techniques based on histogram manipulation, such as adaptive/histogram equalization (HE/AHE) and contrast-limited adaptive histogram equalization (CLAHE), have been extensively used to enhance the images before training. Nevertheless, the method's adaptability for different densities of the breast images and their effects needs to be paid attention. Several studies have included the analysis of the impact of breast density on the post-training level rather than countering its effect on the pre-processing level [10,18,24–26]. However, pre-processing analysis of the mass surrounded by dense tissue is essential to verify that the established CAD system is robust to dense breast images for accurate mass detection.

Based on this motivation, we proposed an enhancement technique that adapts nondense and dense breast categories by subtly changing the non-dense region appearance within a mammogram image through textural refinement, mimicking the radiologist's manual contrast adjustment on individual images while maintaining the visual perceptual of the original image. The textural refinement on the mass edges boosts its feature vector representability during the convolutional process for detection and segmentation algorithm for better classification performance.

In summary, this work's contributions are focused on:


#### **2. Past Literature**

Image enhancement is required to optimize the image's overall quality in preparation for subsequent stages. Enhancements using histogram-based techniques have been proven to enhance mammogram images, such as through histogram equalization [26,27] and the widely used contrast-limited adaptive histogram equalization (CLAHE) [10,18,24,28,29]. Histogram-based image enhancement increases the contrast and dynamic range of the grayscale image by adjusting an image's contrast using its histogram and increases the image's contrast by dispersing the most common pixel intensity values by extending the image's intensity range [30]. Researchers also combined CLAHE with their proposed method to improve their performance. For instance, CLAHE was utilized in conjunction with unsharp masking filtering, with the effectiveness in demonstrating an enhancement for mass region segmentation [31]. In addition, breast cancer detection using a modified CLAHE method is used to sharpen the margins of the masses on three datasets [32]. Meanwhile, CLAHE, wavelet, and anisotropic diffusion combination were presented for mammography enhancement in [33] and obtained a sensitivity of 93% when tested on a limited number of abnormal and normal images from the mini-Mammographic Image Analysis Society (mini-MIAS) dataset. The introduction of multilevel Otsu's thresholding with wind-driven optimization for mass detection utilizing CLAHE enhancement on mini-MIAS and Digital Database for Screening Mammography (DDSM) mammogram datasets is conducted with 96.9% and 96.2% detection sensitivity [29].

Additionally, a different approach using top-hat transform-based mammography enhancement is established to increase the contrast between the suspicious area and normal breast tissues, increasing mass detection sensitivity using the proposed technique compared to unenhanced images [34]. Moreover, grayscale transformation applied by [35] helps reveal more information and increase contrast by selectively emphasizing or suppressing undesirable elements in the image, hence uniformizing the pixel distribution. Recently, a study to detect mass with its performance improved using contrast-based enhancement by employing a hyperbolic tangent function with an adjustable Tunicate swarm algorithm as optimization of the system via fitness function is demonstrated by [36] and shows improvement when compared to the CLAHE method. The use of another optimization through

hybridized fast and robust fuzzy c-means clustering (FRFCM) and particle swarm optimization before mass detection was proposed on the mini-MIAS with 96.6% sensitivity [37]. A classification system for mammogram cancer by [38] using improved multi-fractal dimension features also included a pre-processing subsystem for denoising the mammogram following the cancer region segmentation.

These methods produced good final performance. However, these studies applied a straightforward object detection algorithm to analyze their method's effectiveness for the images to be trained in a full-scale CAD system. Moreover, the enhancement methods did not take the effect of variation of breast density into consideration, with some methods causing the final mass to be indistinguishable from the dense tissue [27,31], where the final output is in the form of classification of mass and non-mass only. This could raise the issue of losing crucial mass features if continued to the cancerous mass classification stage later. The studies were also not tested against any image quality metrics as an essential aspect of any image enhancement method proposal, by using metric performance such as applied by [36], which is not considered the best in the analysis of enhancement for breast density as it relies on the contrast and intensity of the images.

Existing state-of-the-art object identification techniques such as Faster Region-based CNN (R-CNN) [39], You Only Look Once (YOLO) versions [40,41], and Single Shot Multi-Box Detector (SSD) [42] have been implemented in many vision studies for detection, following the image enhancement techniques. YOLO has been proven to be the most beneficial in terms of accurate and fast detection rate [43,44] compared to the other detection algorithms. For example, mass detection using the YOLO model was carried out as proved by Al-Antari et al. [45] and resulted in a detection accuracy of 98.96%. Similarly, [28] enhanced their approach by comparing feedforward CNN, ResNet-50, and InceptionResNet-V2 for classification before implementing the YOLO model for detection. Subsequently, this team [46] proposed a CAD system framework that classified breast masses into malignant and benign using Fully Connected Neural Networks (F-CNNs). This system framework first detected breast masses using the YOLO model with an overall accuracy of 99.7%. Meanwhile, [47] employed the YOLO fusion model for breast mass detection by fusing the best feature representation from single-class mass-based and calcificationbased training models to a multiclass model that combined the feature maps. Their best performance observed was 98.1% for mass lesion accuracy detection. In [48], fusion YOLO was used for detection by introducing new classes of normal and architectural distortion abnormality on final prediction with mass detection accuracy at 93% ± 0.118.

Based on the discussions, although different strategies were implemented to boost mass detection performance, the study has severe limitations that have been conducted to adapt the breast density variance effect through enhancement techniques before training the system. A fully automated mass detection based on density through CAD is crucial, especially with its link with 2.2-fold more cancer risk in clinical profiling for denser breasts reported [49]. Studies conducted by [10,18,24,25] all pointed to a decrease in the model's performance when trained using denser breast images. One of the earliest studies of mammograms that includes adaptation to breast density developed their model using density-based spatial clustering of applications with noise (DB-SCAN), highlighting the breasts' internal structure before training [25]. Likewise, the same method was applied by [24] on a different dataset to improve the method proposed by [25], where the author introduces a two-stage false positive reduction process through bilateral breast analysis. Even though it has good results in preparing the models based on breast density, limitations include if only unilateral breast is available, and asymmetrical factors for both breasts might affect the performance.

#### **3. Proposed Methodology**

This section discusses the overall methodology for completing the framework's three main phases, as shown in Figure 2. Each phase is discussed further in the following subsections.

**Figure 2.** Overall Proposed Methodology for Breast Mammogram Mass Classification.
