Int J Comput Assist Radiol Surg. 2026 Oct 24. doi: 10.1007/s11548-026-03797-8. Online ahead of print.
ABSTRACT
PURPOSE: Cardiovascular diseases are the leading cause of mortality worldwide. Among them, myocardial infarction (MI), which is characterized by abnormal myocardial motion, remains a critically important condition to detect. Accurate diagnosis of MI requires detailed motion analysis; however, motion-based assessment is time-consuming and imposes a considerable burden on physicians. Moreover, most previous studies have relied on local motion descriptors or limited shape features, resulting in suboptimal detection performance.
METHODS: To address these limitations, we propose a deep learning framework, termed MambaMorph-Motion Infarction Detector (MMMID), for myocardial infarction detection from cardiac CT images. The proposed pipeline integrates semantic segmentation, sequential deformable registration using MambaMorph, and a DenseNet-based classification module to effectively leverage rich spatiotemporal myocardial motion information across the cardiac cycle.
RESULTS: The proposed method enables accurate tracking of myocardial motion aligned with anatomical structures by associating time-resolved motion fields with corresponding CT volumes. MMMID achieved an accuracy of 91.92% and an AUC of 0.9335 on an in-house cardiac CT dataset. Compared with the CT-only configuration, the integration of sequential motion fields improved recall, suggesting that motion information provides complementary diagnostic cues for myocardial infarction detection rather than replacing anatomical CT information.
CONCLUSION: The proposed framework enables myocardial infarction detection by jointly leveraging cardiac anatomy and estimated myocardial motion. Although CT appearance remained a strong source of discriminative information, sequential motion fields contributed complementary functional information, particularly for improving sensitivity to infarction cases. MMMID may serve as a supportive tool for automated cardiac CT analysis and could help reduce the burden of motion-based assessment in clinical workflows.
PMID:42926307 | DOI:10.1007/s11548-026-03797-8