Explainable WinDWS-SwinNet: Optimizing Ultrasound Deep Radiomics for Multipathology Breast Lesion Classification

Timely and accurate characterization of breast lesions into different pathological classes through noninvasive imaging techniques is crucial for early cancer diagnosis and effective patient management. This study aims to develop a novel, explainable attention-based deep network to analyze radiomic patterns for predicting heterogeneous breast cancer types using ultrasound imaging.

A multicentric retrospective dataset comprises brightness-mode ultrasound images of breast tissue collected across five different centers/databases. A total of 3256 images, having biopsy-proven 2068 benign and 1188 malignant lesions, were analyzed. A lightweight and computationally scalable network, WinDWS-SwinNet, is developed for binary and multipathology breast cancer classification. An optimized image preprocessing algorithm is proposed to enhance ultrasound radiomic features, thereby improving data generalizability and diagnostic accuracy.

The WinDWS-SwinNet architecture demonstrated superior performance in breast cancer classification, achieving an area under the curve of 0.96, an overall test accuracy of 95.79%, and robust class-wise results with 95.11% precision and 98.51% sensitivity for benign tumors and 97.13% precision and 90.86% sensitivity for malignant tumors. For multipathology classification, the model achieved a test accuracy of 93%, with class-wise accuracies ranging from 94% to 99.5%. External validation on the independent unseen dataset resulted in an accuracy of 95.36%, highlighting the robustness and generalizability of the proposed model beyond the training data.

The developed methodology shows promising results in characterizing breast lesions into pathological subtypes noninvasively, effectively mimicking real-world diagnostic workflows, thereby supporting radiologists in reducing diagnostic uncertainty and enhancing confidence in clinical decision-making.