Abstract

Precise detection of the pectoral muscle in mediolateral oblique (MLO) mammograms is vital for the reliability of automated breast cancer detection systems. As a result, the bright, triangular morphology of the pectoral muscle is often indistinguishable from a malignant lesion, thereby contributing to low diagnostic accuracy and high false-positive rates for computer-aided diagnosis (CAD) tools. This publication describes a deep learning–inspired pectoral muscle segmentation and elimination approach grounded in U-Net and its advanced U²-Net counterpart to increase the accuracy, robustness, and consistency in mammographic preprocessing. We used a comprehensive full-field digital mammography dataset with pixel-level expert-annotated ground truth for training and evaluation. Standard preprocessing steps—including image resizing, intensity normalization, and binary mask preparation—allowed for uniformity and stability in heterogeneous samples. The dataset performed well in terms of segmentation based on its Dice Similarity Coefficient (DSC) and Intersection over Union (IoU) quantitation which scored an overall high (best validation Dice score = 0.7426; Dice = 0.8379, IoU = 0.7847) in independent assessments. The convergence simulation confirmed smooth optimization, good matching between training–validation losses, and minimal overfitting. The qualitative evaluation also further strengthened the ability of the model to maintain anatomical integrity of the pectoral muscle by means of an accurate boundary definition and the continuity of the shape, thus also overcoming blur, noise, and contrast variability. Compared with the linear geometry of the muscle boundary predicted by classical Hough line–based approaches, the proposed approach is capable of overcoming the restriction of the linear edge detection mechanism. The results of these experiments render the model a strong yet clinically relevant preprocessing module with the capability to facilitate breast cancer classification performance. Due to its powerful generalization performance, the framework offers potential for implementation in numerous clinical systems, including resource-limited settings. Subsequent expansions would involve coupling with attention mechanisms, transformer-based encoders, domain adaptation programs, and radiologist-in-the-loop validation to enhance clinical translation and the robustness of AI-supported mammographic analysis.

Statistics