Fully automated deep learning echocardiographic strain model for vessel-specific detection of coronary occlusion

Background Strain echocardiography is uniquely positioned for early risk-stratification in suspected acute coronary syndromes (ACS), yet its diagnostic utility is limited by time-consuming analyses, operator dependency, and rigid anatomical assumptions on coronary perfusion territories. Deep learning (DL) enables fully automated, fast, and reproducible quantification of segmental strain during image acquisition1–3, potentially improving detection of high-risk ACS. Purpose To validate a fully automated DL-based method that measures and integrates regional strain from all 18 myocardial segments into a transparent and interpretable vessel-specific occlusion prediction model. Methods We prospectively included participants with acute myocardial infarction undergoing coronary angiography. Longitudinal strain was measured in the three standard apical views by a fully automated DL method developed by our group. Standardized segmental strain values were automatically integrated into the occlusion prediction model, using learning-based segment-specific weights to evaluate the strain patterns and generate probabilities of LAD, CX, and RCA occlusion. Diagnostic performance for angiographically confirmed occlusion was evaluated using ROC statistics and compared with conventional regional strain based on hypothetical perfusion territories. Results 1232 participants were included (mean age 74 ± 12, 77% male), of whom 58% presented with NSTEMI and 42% with STEMI. 23% had a prior history of coronary artery disease. Coronary occlusion was confirmed angiographically in 654 participants. The DL-based model had a feasibility of 94% and completed measurements within 1 second. Our model demonstrated high predictive performance for LAD (AUC 0.84), for CX (AUC 0.83), and RCA (AUC 0.80, Figure 1), consistently outperforming conventional territorial strain (AUCs 0.79, 0.67, and 0.59, respectively). All improvements were significant (p <0.0001), with the largest improvements for CX and RCA. Model outputs included interpretable vessel-specific probability maps that visualized the predicted occlusion patterns (Figure 2). Conclusion A fully automated DL method integrating weighted longitudinal strain from all 18 segments accurately identified vessel-specific coronary occlusion, outperforming conventional territorial strain. This approach may facilitate early risk stratification and improve detection of occlusive ACS.