PhyJointNet: Attention-Guided Joint Segmentation & Classification of Breast Ultrasound with Latent Feature Augmentation

Ultrasound imaging is widely used for early breast cancer detection, but building automated systems that can both localize segmentation and assess classification remains challenging. These challenges stem from task interference, class imbalance in clinical datasets, and variability in image appearance due to differences in ultrasound hardware. To address this, we propose PhysJointNet, a physics-based augmentation framework for joint segmentation and classification in breast ultrasound. The framework enhances robustness and generalization across imaging devices while supporting effective multi-task learning. It includes a physics-informed augmentation module that uses a frozen descriptor to capture and align scanner-specific characteristics, reducing hardware-induced variations in texture and signal gain. Experimental results on benchmark datasets show improved performance, achieving 0.853 in segmentation and 0.918 in classification, demonstrating the effectiveness of the proposed approach for reliable clinical screening.