Multi-task learning with orthogonal feature disentanglement for breast tumor segmentation and classification in ultrasound images
Junwei Yang, Shenhai Zheng, Junyu Zhu, Chengman Jiang · Biomedical Signal Processing and Control · 2026
In breast tumor image analysis, multi-task computer-aided diagnosis systems are often insensitive to local texture features in ultrasound images, and existing frameworks generally face the challenge of insufficiently mining and effectively decoupling features between classification and segmentation tasks. To address this, this paper proposed a multi-task learning method for breast ultrasound image tumor segmentation and classification using orthogonal feature disentanglement, aiming to deeply integrate and utilize the intrinsic relationships inter-task. First of all, to leverage the feature differences inter tasks, we designed an orthogonal dual-attention decoupling module to collaboratively support both classification and segmentation. Then, a contrast-aware dynamic fusion convolution module is introduced to enhance the perception of local structural variations and improve the accuracy of boundary localization. Furthermore, a cross-task orthogonal feature decoupling layer is constructed to extract task-specific features from mixed features, further improving the performance of individual tasks. Experimental results show that the proposed method achieves a segmentation performance of DSC = 84.25%, HD = 23.80, IOU = 76.52%, and a classification performance of F1 = 83.60%, ACC = 88.69%, AUC = 94.54% on the BUSI dataset. On the BUS-B dataset, it achieves a segmentation performance of DSC = 87.29%, HD = 12.98, IOU = 80.42%, and a classification performance of F1 = 86.81%, ACC = 90.13%, AUC = 91.91%. The results demonstrate that the proposed method not only improves the performance of breast tumor ultrasound image segmentation and classification but also provides new insights for the research on multi-task feature decoupling mechanisms.