Proceedings of the Annual Meeting of the Cognitive Science Society

Radiologists often implicitly rely on tissue backscattering characteristics to interpret noisy ultrasound (US) images under challenging imaging conditions. However, existing deep learning-based segmentation models rarely incorporate such intrinsic physical statistics into representation learning. To address this limitation, we propose Naka-SAM, a physics-informed ultrasound segmentation framework built upon the Segment Anything Model (SAM). Specifically, we design a Physical Prior Generator (PPG) to model ultrasound backscattering statistics through learnable Nakagami distribution estimation, where the shape parameter \(m\) characterizes local scattering concentration and the scale parameter \(\Omega\) reflects backscattered energy variations. These physics-informed priors provide domain-invariant tissue representations associated with underlying microstructural properties. Furthermore, a dual-stage gated fusion strategy is introduced to progressively integrate physical priors with both low-level structural features and high-level semantic representations, thereby improving robustness under noisy and low-contrast imaging conditions. Extensive experiments across seven ultrasound datasets demonstrate that Naka-SAM consistently outperforms both conventional segmentation methods and recent SAM-adapted frameworks, while exhibiting strong cross-domain generalization capability across heterogeneous ultrasound imaging scenarios.