Fully Automatic AI-Based Quantification of LV Mass in Echocardiography: A Multimodality Validation.
Authors
Affiliations (4)
Affiliations (4)
- Department of Internal Medicine and Cardiology, Herzzentrum Dresden, Technische Universität Dresden, Dresden, Germany. Electronic address: [email protected].
- Institute and Polyclinic for Diagnostic and Interventional Radiology, Faculty of Medicine and University Hospital Carl Gustav Carus, Technische Universiät Dresden, Dresden, Germany.
- Department of Internal Medicine and Cardiology, Herzzentrum Dresden, Technische Universität Dresden, Dresden, Germany.
- Institute for Structural Analysis, Technische Universität Dresden, Dresden, Germany.
Abstract
Accurate assessment of left ventricular (LV) myocardial mass is critical for guiding treatment decisions. Measurements based on echocardiography are limited by operator variability. Artificial intelligence (AI)-based methods promise improved precision and reproducibility, but they require validation against high-resolution reference standards such as computed tomography (CT). The aim of this study is to compare the accuracy of AI-based vs expert manual echocardiography LV mass measurements using CT as the reference and to evaluate real-world test-retest reliability. In 218 patients undergoing echocardiography and CT, the authors analyzed LV mass, interventricular septal diameter (IVSD), end-diastolic diameter, posterior wall diameter (PWD), and end-diastolic volume. LV mass was assessed by the conventional linear method as well as a novel hybrid method combining end-diastolic tracings with mean wall thickness. AI- and expert-based (Expert) echocardiographic measurements were compared with CT using intraclass correlation coefficient (ICC) and mean absolute percentage error (MAPE). Increased relative wall thickness (RWT >0.42;+) and abnormal LV mass was evaluated using receiver-operating characteristic analysis (area under the curve [AUC]). Test-retest reliability was assessed in a bedside cohort (n = 40) using coefficient of variation. AI echocardiography showed strongest agreement with CT for LV mass assessment using the hybrid method (ICC: 0.76; MAPE: 0.16) and lower agreement for the linear method (ICC: 0.49; MAPE: 0.44). Nevertheless, compared with Expert echocardiography, AI reduced measurement error by ∼20%, mainly due to more consistent IVSD and PWD measurements, and reclassified 34% of RWT+ cases. Diagnostic accuracy for detecting increased LV mass was higher with AI echocardiography than with Expert echocardiography (AUC: 0.78 vs 0.71 for linear; 0.84 vs 0.77 for hybrid). Test-retest reliability was highest with AI echocardiography for both mass methods (coefficient of variation: 7.8% and 8.1%), and reproducibility of manual assessment varied with operator skills. AI echocardiography provides superior accuracy, diagnostic performance, and reproducibility for LV mass assessment compared with expert evaluation, supporting its clinical integration to improve standardization in cardiac imaging.