<p>Despite the great advances in the clinical diagnosing system, breast cancer is still the most common and fatal disease to woman all over the world, which indicates that there is still a urgent demand of precise and explainable diagnosis system. In this work, we propose a segmentation-guided knowledge distillation framework, which combines the high accuracy of deep learning and clinical explanation ability for breast cancer detection on ultrasound images. The proposed method utilizes a teacher–student framework, and the ResNet-50 teacher is capable of transferring high-level diagnostic representations to the MobileNetV2 student model. Segmentation masks are used as an auxiliary input to guide the student to attend to the tumor areas during training and to improve the relevance of features. This combination provides a good balance for knowledge transfer with low computational complexity. Tested on the BUSI ultrasound dataset, the model outperformed traditional CNN and transfer learning baseline with an overall accuracy and F1-score of 94%. Furthermore, scalability studies demonstrated robust results for reduced fractions of the dataset with real-time inference at more than 200 FPS. The results suggest that distillation of segmentation to the classification task, in the proposed framework improves diagnostic ability and enhances the model explainability, hence it leads to robust and trustworthy AI solutions for breast cancer screening.</p>