EchoSAM2: enhancing anatomical consistency in few-shot cardiac ultrasound segmentation via sequence-wise memory and registration

Introduction Few-shot segmentation in echocardiography is highly valuable for rapid clinical assessment, yet remains challenging because expert annotations are scarce, cardiac morphology varies across patients, and ultrasound videos exhibit complex temporal motion. Although the Segment Anything family provides strong generic priors, its direct deployment on echocardiographic videos often suffers from unstable frame-wise predictions and limited adaptation under low-label conditions. Methods We propose EchoSAM2, a few-shot framework tailored to cardiac ultrasound video analysis. The method introduces support-conditioned, sequence-level memory propagation to preserve anatomical coherence throughout the cardiac cycle. To stabilize inference, we design an Adaptive Start-Frame Selection (ASFS) module that automatically identifies a favorable initial frame and incorporate a lightweight registration network to efficiently align support and query frames. Results Experiments on the CAMUS and EchoNet-Dynamic datasets demonstrate that EchoSAM2 achieves new state-of-the-art performance for few-shot segmentation and clearly outperforms existing video segmentation and SAM-adaptation baselines under low-annotation settings. EchoSAM2 also narrows the performance gap between few-shot and fully supervised models. Discussion These results demonstrate the effectiveness of support-conditioned temporal propagation, adaptive start-frame selection, and efficient frame alignment for few-shot echocardiographic video segmentation. EchoSAM2 shows practical potential for automated cardiac quantification and computer-aided diagnosis in data-constrained clinical settings.