Real-time deep learning echocardiographic global longitudinal strain improves agreement with cardiac magnetic resonance feature tracking compared to a conventional workflow

Background Left ventricular (LV) global longitudinal strain (GLS) improves detection of systolic dysfunction and prognostic assessment compared to ejection fraction, yet conventional semi-automatic workflows are time-consuming and operator dependent. Deep learning (DL)-based GLS analyses may address these challenges, but validation is challenged by the lack of a gold standard and vendor-specific echocardiography bias. Assessment against cardiac magnetic resonance feature tracking (CMR-FT) provides an independent comparator. Purpose To compare agreement and correlation between conventional echocardiographic GLS, real-time DL GLS, and CMR-FT as measures of myocardial deformation. Methods Consecutive patients referred for clinical CMR (n=35) underwent same-day CMR and two separate echocardiograms performed by different operators: (1) DL-assisted acquisition with real-time GLS measurements and (2) guideline-directed conventional acquisitions, Figure 1. DL-assisted acquisitions were analysed by two in-house–developed methods: Optical flow1, which estimates dense, pixel-wise myocardial motion across frames, and LocoTrack2, which follows discrete anatomical landmarks and derives strain from their relative displacement. Conventional acquired images were analysed by a blinded expert cardiologist using semi-automatic software (EchoPAC version 204, GE HealthCare). CMR-FT was analysed by a blinded expert radiologist using the cvi42 software (Circle Cardiovascular Imaging). Bland-Altman statistics, mean absolute difference (MAD), intraclass (ICC 2.1) and Pearson's correlation coefficients were used to evaluate agreement and correlation. Results Mean age was 55 ± 15 years, and 21 (60%) were male. Mean LV ejection fraction was 49 ± 7%. Comorbidities were common and included hypertension (40%), heart failure (17%), coronary artery disease (14%), and atrial fibrillation (14%). Mean GLS was 14.9 ± 3.9% with CMR-FT, 16.8 ± 3.4 with the conventional workflow, 14.6 ± 3.0 with Optical Flow, and 14.4 ± 3.3 with LocoTrack. Bias against CMR-FT was significantly lower for both DL methods compared with the conventional workflow (both p < 0.01; Figure 2). MAD relative to CMR-FT was also significantly lower with Optical Flow (MAD 2.0, p = 0.02) and LocoTrack (MAD 2.0, p = 0.03) than with the conventional workflow (MAD 2.8). ICCs were higher for the DL methods (ICC = 0.77 and 0.78) than for the conventional workflow (ICC = 0.65), consistent with stronger Pearson correlations (r = 0.80 and 0.79 vs 0.74). Conclusion A real-time DL-assisted GLS workflow demonstrated better agreement and stronger correlation with CMR-FT compared with the conventional workflow, resulting in a negligible and significantly lower bias. These observations underscore the value of complementary imaging modalities in validating DL-based methods in echocardiography.