Clinical knowledge constrained multi-task learning framework for breast cancer diagnosis using ultrasound videos

Ultrasound is widely used in breast cancer diagnosis due to its cost-effectiveness, non-invasiveness, and radiation-free properties. Computer-aided diagnostic systems based on dynamic breast ultrasound videos have advanced rapidly. However, existing methods are often affected by redundant information, make limited use of clinical prior knowledge, and provide insufficiently interpretable diagnostic results. To address these limitations, a clinical knowledge constrained multi-task learning framework (CKC-Framework), inspired by radiologists' workflow, is proposed for breast cancer diagnosis in ultrasound videos. The framework employs a DINO-based object detection algorithm to select keyframes and tumor-containing segments from variable-length videos, filtering out disruptive background frames. A diagnostic attribute constraint module is designed to enhance intra-frame feature interpretability and provide radiologists with keyframe diagnostic attributes. In addition, a clinical prior constraint module is introduced to guide temporal weighting and improve cross-frame semantic consistency. Finally, a keyframe-guided spatiotemporal Mamba is proposed to integrate spatial features across frames in long sequences via spatiotemporal scanning. For evaluation, we constructed Breast-USV, a large-scale breast ultrasound video dataset, and used the public BUSV dataset for benchmarking. Experimental results demonstrate that the framework achieves an AUC of 94.59% and an accuracy of 91.43% on Breast-USV, and an AUC of 90.38% and an accuracy of 87.39% on BUSV, achieving superior performance compared with existing methods, with additional cross-dataset evaluation from Breast-USV to BUSV to assess out-of-domain generalization. Moreover, by providing diagnostic attributes and lesion localization, the framework may support radiologists' decision-making and facilitate more efficient post-acquisition review.