ABSTRACT Ultrasound imaging is widely used in clinical practice due to its noninvasive nature, real‐time capability, and cost‐effectiveness. However, accurate automatic segmentation still faces significant challenges due to inherent physical limitations such as noise, acoustic attenuation, and blurred tissue boundaries. In this study, we propose PAA‐Swin, a physics‐aware Transformer framework that integrates acoustic principles into deep learning. Specifically, the Physics‐Aware Attention (PAA) mechanism explicitly models acoustic wave propagation and depth‐dependent attenuation via positional encoding. Building upon this, the Attenuation‐Aware Feature Modulation (AAFM) achieves signal compensation by predicting local attenuation coefficients, effectively addressing degradation at greater depths. Additionally, Adaptive Multiscale Feature Aggregation (AMFA) employs parallel branches with varying dilation rates to capture diverse acoustic textures. Extensive experiments on the TN3K, BUSI, and CAMUS datasets demonstrate that PAA‐Swin achieves Dice coefficients of 85.29%, 86.33%, and 96.75%, respectively. These results represent significant improvements of 1.79%, 3.98%, and 3.3% over state‐of‐the‐art methods, validating the robustness and generalizability of physics‐constrained deep learning in ultrasound image segmentation.
Sun et al. (Sun,) studied this question.