PMSA-Net: A Pectoral Muscle Suppressed Attention Network for Automated Breast Cancer Detection in Mammographic Images

Abstract

Breast cancer is one of the most common causes of cancer death in women, and early and accurate diagnosis is paramount to enhance survival of patients, and mammography remains the gold standard in diagnosis. In spite of the advancement of deep learning-based mammographic analysis, the current methods have limitations, such as the presence of pectoral muscle regions, background artifacts, and varying breast densities, and do not adequately consider clinically relevant lesion regions. These constraints may introduce ambiguity in features and represent poor classification performance, especially in mixed mammographic databases. In view of these challenges, a novel deep learning framework, called PMSA-Net (Pectoral Muscle Suppressed Attention Network), for automated breast cancer detection is proposed. The proposed pipeline is based on a single architecture that includes breast region extraction, pectoral muscle suppression, adaptive image enhancement, convolutional feature extraction, channel attention, spatial attention, and multi-scale feature fusion. In order to enhance the image quality, first some image enhancement techniques are used: Anisotropic Diffusion Filtering and Multi-Scale Contrast Limited Adaptive Histogram Equalization (MS-CLAHE). Then, irrelevant structures will be reduced by breast region extraction and suppressed by the pectoral muscle, and the visibility of the lesion will be improved by adaptive enhancement. The improved mammograms are then fed into a CNN backbone and channel attention and spatial attention modules to highlight diagnostically relevant areas. Low level feature and high-level feature are fused at multi-scale for robust classification. Experimental analysis shows that the accuracy, precision, recall, F1-score, sensitivity, specificity and AUC of PMSA-Net are 97.86%, 97.24%, 97.51%, 97.37%, 97.52%, 98.14% and 0.9912 respectively. Comparative analysis shows better results than other deep learning based on existing mammography. The key findings show that suppression of the pectoral muscle and attention guided feature learning lead to significant enhancement of discriminative power and fewer false predictions. The proposed PMSA-Net is a reliable and effective framework for mammographic breast cancer detection, and can be used to assist clinical decision-making in computer-aided diagnosis systems.