Abstract Background Accurate quantification of left ventricular (LV) systolic function is fundamental in echocardiography. LV Global Longitudinal Strain (GLS) offers advantages over LV ejection fraction, being more sensitive, reproducible and offering better prognostic value. However, existing semi-automatic methods are time-consuming and heavily operator-dependent. Additionally, poor image quality and foreshortening decrease accuracy and reproducibility of GLS. As a result, these challenges lead to suboptimal efficiency and underutilization of GLS quantification in clinical practice. Purpose Our objective was to develop a deep learning (DL) application for real-time automated measurements of GLS, capable of providing tracking quality and foreshortening feedback during scanning. Subsequently, we aimed to validate the novel method in a prospective cohort study. Methods Deep learning networks were trained to identify apical views, identify timing of end-systole and end-diastole, conduct cardiac segmentation, estimate myocardial motion, assess LV length in each view for foreshortening detection, and present real-time visualization of tracking quality and GLS values for three cardiac cycles to the operator for review (Figure 1). The DL method was validated in a prospective cohort of 40 patients included regardless of image quality. Agreement with manual measurements obtained using semi-automatic software (EchoPAC version 204, GE HealthCare, Horten, Norway) was evaluated. Bland-Altman statistics and Pearson's correlation coefficient were used to evaluate agreement and correlation. Results Feasibility for real-time DL measurements of GLS was 98%, with one patient excluded due to poor image quality. There was good agreement (Bias 0.7%, LoA: -1.6 to 3.1%, figure 2) and excellent correlation (Pearson’s correlation coefficient of 0.93, p<0.001) between the real-time DL application and manual reference measurements for GLS. Using the DL application during scanning resulted in a 58% (95% CI 56%-61%) reduction in time required for scanning and post-processing to obtain GLS values, compared to using the semi-automatic method. Average time consumption was 2 minutes and 16 seconds for the DL method and 5 minutes and 27 seconds for the reference method. Conclusion Using deep learning, real-time fully automated measurement of GLS during scanning is feasible, time-saving, and provides good agreement with reference methods. Real-time measurements also offer the potential to reduce operator-related variability by integrating tools to provide image and tracking quality feedback to the user, ultimately benefiting patients.Application user interfaceAgreement between DL and reference
Read full abstract