Deep learning-derived left ventricular circumferential strain for cardiac MRI diagnosis of acute myocarditis with...

To evaluate the diagnostic performance of deep learning-derived left ventricular circumferential strain (DL-LVCS), a method that reduces operator dependency and analysis time compared with conventional feature tracking, in patients with acute myocarditis and preserved LVEF (≥ 55%), and to compare it with feature tracking-derived strain (FT-LVCS). This retrospective study, conducted between January 2021 and December 2024, included 142 patients with acute myocarditis, stratified into preserved EF (AMY-pEF, n = 106, mean age, 31 ± 9 years; 80 male) and reduced EF (n = 36, mean age, 33 ± 12 years; 27 male) groups, alongside 106 age- and sex-matched healthy controls (mean age, 32 ± 7 years; 78 male). A deep learning model integrating PPIM-Net (Pyramid Progressive Image Mapping Network) for segmentation and U-Net for motion tracking was developed to compute DL-LVCS. Conventional LVCS was also obtained. Diagnostic performance was evaluated using ROC analysis, and AUCs were compared with the DeLong test. In the AMY-pEF group, both FT-LVCS (-18.4%±1.2 vs. -19.8% ±1.3) and DL-LVCS (-8.0%±1.4 vs. -10.4%±1.8) were significantly impaired versus controls (both P < 0.001). For diagnosing AMY-pEF, FT-LVCS (cutoff=-19.3%) achieved an AUC of 0.806, with 67.2% sensitivity and 84.5% specificity. DL-LVCS (cutoff=-9.2%) demonstrated significantly higher AUC (0.887), sensitivity (85.9%), and specificity (79.2%) (DeLong P < 0.01). DL-LVCS enables quantification of myocardial deformation and demonstrates higher sensitivity than conventional LVCS for discriminating patients with CMR-defined acute myocarditis and preserved LVEF from healthy controls. It may serve as a supplementary CMR biomarker when standard functional metrics are within normal ranges.