Ultrasound-based breast nodule malignancy diagnosis using CNN enhanced by knowledge distillation

Accurate discrimination of benign and malignant breast lesions is essential for early diagnosis, yet remains challenging due to lesion heterogeneity and interobserver variability in ultrasound assessment. We developed a deep learning (DL) framework integrating dual inverted bottleneck blocks (Dual-IB), adaptive hierarchical feature fusion, and knowledge distillation (KD) for breast ultrasound lesion classification. The model was trained and evaluated on public datasets and validated on an independent external cohort to assess its generalizability. Results demonstrated superior performance compared with baseline methods and resident radiologists, while maintaining robust performance across diverse datasets and imaging conditions. On the external validation dataset, the model achieved the best performance, with an accuracy (Acc) of 92.64% and an F1-score of 92.37%. Furthermore, class activation mapping (CAM) visualizations revealed focused and anatomically plausible attention patterns within lesion regions, aiding interpretation and suggesting its potential to assist in breast ultrasound lesion classification.