SegMotion-Net: Segmentation-guided motion analysis for early myocardial infarction detection from echocardiography video.

Myocardial infarction (MI) remains a major clinical challenge, and early detection is essential to prevent irreversible myocardial damage. However, echocardiography-based MI detection relies heavily on physicians' subjective interpretation, leading to inter-observer variability and suboptimal performance. To address these challenges, we propose SegMotion-Net, an interpretable framework that integrates segmentation-derived anatomical priors with motion representation learning for MI detection. Specifically, a task-specific memory mechanism is employed to aggregate spatiotemporal features across cardiac cycles, providing contextual cues for accurate left ventricular (LV) wall segmentation. To mitigate error accumulation during memory updating, a memory enhancement module refines stored representations using predictive masks. Furthermore, we introduce an LV wall motion dynamics analysis module to capture temporally coherent and region-specific motion patterns associated with MI. On the public HMC-QU dataset, SegMotion-Net achieves a Dice coefficient of 93.5% for LV wall segmentation, and an AUC of 86.7% together with an F1 score of 87.5% for MI classification. External validation across multiple private datasets provides additional evidence of cross-center robustness. Notably, SegMotion-Net achieves performance approaching that of experienced cardiologists under echocardiography-only evaluation settings. By explicitly modeling LV wall segmentation and motion dynamics, the proposed framework provides interpretable and clinically meaningful decision support for MI detection.