Enhanced Three-Class Breast Ultrasound Classification Using Controlled Augmentation and ResNet50-CBAM with Explainability

Abstract

Breast cancer is one of the leading causes of cancers among patients around the world, where the early diagnosis of the disease by diagnostic ultrasound imaging technique becomes vital in achieving low mortality rates. Nevertheless, class imbalance continues to remain a problem in computer-aided automated diagnosis of breast ultrasounds, largely because of the scarcity of samples representing normal tissues compared to those representing pathological tissue. In order to resolve this crucial issue, a data augmentation scheme, targeting the normal category, is proposed, which utilizes the ResNet-50 neural network architecture with modifications, along with the Convolutional Block Attention Module (CBAM). The experimental dataset consists of both the publicly available Breast Ultrasound Images (BUSI) dataset as well as the proprietary clinical dataset. By applying controlled geometric and photometric transformations, the normal tissue class is augmented to 800 images from its initial value of 133, thus restoring balance between the classes and allowing the network to learn robust invariants of healthy anatomy. For high dimensional feature extraction, a customized ResNet-50 network is used, whereas the Convolutional Block Attention Module (CBAM) in the framework iteratively rescales the channels and spatial features to highlight the important acoustic features. The system improves the distinction between malignant, benign, and healthy tissues. An extensive experimental analysis on an independent test data set shows the outperformance of the proposed framework with an overall accuracy of 86.99%, precision of 88.07%, recall of 88.50%, and AUROC of 96.29%. Interpretability is ensured using the Grad-CAM visualization method, where automatic classification results align with pathological regions.