Accuracy of a Deep Learning Model in Intracardiac Echocardiography

Résumé fourni par la source

BACKGROUND: Intracardiac echocardiography (ICE) is widely used during electrophysiology and structural heart procedures; however, image interpretation remains operator-dependent and procedural views are not standardized. Although artificial intelligence has been increasingly applied to transthoracic and transesophageal echocardiography, applications to ICE remain limited. OBJECTIVES: The objective of the study was to develop and evaluate Auto-Contour, a deep-learning pipeline for multistructure semantic segmentation of ICE anatomy and assess its feasibility for real-time procedural guidance. METHODS: In this retrospective multicenter study, 5,496 deidentified ICE cine loops from 249 procedures of unique patients, including routine clinical cases and the ViewFlex™ X first-in-human study, were analyzed. ICE experts classified each cine into 1 of 20 procedural views and annotated key anatomic structures, including the left atrium, left atrial appendage, pulmonary vein ostia, valves, cusps, papillary muscles, and left ventricle, at end-systole, and end-diastole, yielding 65,117 segmentations. A deep-learning segmentation model was trained using patient-level splits, standard augmentations, and early stopping. RESULTS: Segmentation performance was highest for larger cardiac chambers, with Dice scores of 0.94 for the left atrium and 0.82 for the left ventricle, and corresponding 95th-percentile Hausdorff distance values of 1.18 mm and 3.27 mm. Smaller structures also demonstrated acceptable performance, including the left atrial appendage, pulmonary veins, papillary muscles, and aortic cusps. The mean per-frame inference time was <0.03 seconds. CONCLUSIONS: Auto-Contour demonstrated robust multistructure segmentation of ICE anatomy with real-time inference, supporting prospective evaluation of artificial intelligence-assisted ICE for procedural standardization, efficiency, and safety.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.