Clinical utility of a fully automated AI system for accurate left ventricular mass measurement in echocardiography

Abstract Background Accurate measurement of left ventricular (LV) myocardial mass (MM) using 2D transthoracic echocardiography (Echo) provides valuable prognostic information and assists in tailoring treatment plans. Purpose We sought to validate a fully automated artificial intelligence (AI) system for LV diameter and wall thickness measurements, including MM assessment, using Echo, compared to independent human operators. Methods We analysed consecutive patients who underwent diagnostic transthoracic echocardiography (Echo) at our institution. End-diastolic interventricular septal and posterior wall thicknesses (IVSD and PWD), along with end-diastolic diameter (EDD) from parasternal long-axis views, were independently measured by both a fully automated AI system and human operators for one-dimensional LV myocardial mass (MM) calculation (Figure 1). In a subgroup of patients who also underwent cardiac magnetic resonance (CMR) imaging within 7 days, Echo-based MM results were compared with those derived from CMR. The AI system automatically selected appropriate views from the complete Echo exam and measured the parameters without any operator involvement. Results were presented through a web-based interface. For this study, the automatically generated values were used for analysis. Agreement between measurement methods was assessed using Pearson’s correlation coefficient (R), mean absolute error (MAE), and Bland-Altman analysis (bias and limits of agreement LOA). Results The final cohort included 1,896 consecutive patients with complete LV MM measurements by both AI and human operators (feasibility: 97%). AI and human measurements showed excellent correlation for EDD (r = 0.87, MAE = 3.7 mm, bias = 2.3 mm, LOA = -10.5 to 6.0 mm), and good correlation for IVSD (r = 0.59, MAE = 2.2 mm, bias = 1.8 mm, LOA = -2.4 to 6.0 mm) and PWD (r = 0.53, MAE = 1.8 mm, bias = 1.1 mm, LOA = -2.8 to 5.0 mm) (Figure 1, left panel). LV MM values from AI and human showed strong correlation (r = 0.77, MAE = 27 g, bias = 18 g, LOA = -40 to 76 g) (Figure 1, right panel). However, in the CMR subgroup (n = 72) AI demonstrated significantly higher agreement with CMR-derived MM (r = 0.82, MAE = 23 g, bias = 3.6 g, LOA = -61 to 67 g ) compared to human assessment (r = 0.71, MAE = 28 g, bias = -7 g, LOA = -85 to 72 g) (p = 0.008 for difference). Conclusion These findings underscore the potential of an AI-based application for fully automated view selection and measurement of left ventricular wall thickness and diameter. Although correlations for wall thickness measurements were lower - likely due to intrinsic anatomical variability and echocardiographic limitations - the AI system demonstrated stronger agreement with CMR-derived LV MM compared to human operators. This supports its integration into clinical workflows for scalable, reproducible, and operator-independent quantification.Figure 1