Lightweight Shape-Aware Segment Anything for Cardiac Ultrasound Segmentation

Foundation models such as the Segment Anything Model provide strong promptable segmentation, but are computationally heavy and largely ignore medical priors. We propose a lightweight, prior-guided framework, tailored for multi-class cardiac segmentation. Built on EfficientSAM, our model replaces prompt interaction with learnable class tokens, enabling fully automatic multi-class prediction with only 8.99M parameters. Specifically, we introduce a Shape-Aware Decoder that groups organs by data-driven aspect ratios into compact, elongated, and highly strip-like categories, and assigns each group to a dedicated branch with morphology-specific convolutions. A Position Prior Attention module further fuses multi-branch logits with empirical spatial priors derived from training masks. Experiments on a curated fetal four-chamber dataset and the CAMUS benchmark show that our method achieves state-of-the-art accuracy with significantly fewer parameters.