Towards standardizing Mitral Transcatheter Edge-to-Edge Repair with deep-learning algorithm: a comprehensive multi-model strategy

Silvia Corona^1,2*Théo Godefroy³Olivier Tastet^4,5Denis Corbin^4,5Thomas Modine⁶Ralph Stephan Von Bardeleben²Frederic Lesage⁴Walid Ben Ali^2*

• ¹Universite de Montreal Montreal Heart Institute, Montreal, Canada
• ²Interventional and Structural Cardiology, Institut De Cardiologie de Montreal, Montreal, Canada
• ³Biomedical Engineering, Polytechnique Montreal, Montreal, Canada
• ⁴Polytechnique Montreal, Montreal, Canada
• ⁵Institut De Cardiologie de Montreal, Montreal, Canada
• ⁶UMCV, Centre Hospitalier Universitaire de Bordeaux Hopital Cardiologique, Pessac, France

Background: Severe mitral valve regurgitation requires comprehensive evaluation for optimal treatment. Initial screening uses transthoracic echocardiography (TTE), followed by transesophageal echocardiography (TEE) to determine eligibility for adequate intervention. Mitral Transcatheter Edge-to-Edge Repair (M-TEER) indications are based on detailed and quality valve and sub-valvular apparatus assessment, including anatomy and regurgitation pathophysiology. Aim: To develop AI algorithms for standardizing M-TEER eligibility assessment using TTE and TEE echocardiograms, supporting all stages of mitral valve regurgitation evaluation to assist non-expert centers throughout the entire process, from severe mitral valve regurgitation diagnostic to M-TEER procedure. Methods: Three deep learning algorithms were developed using echocardiographic data from M-TEER patients performed at Montreal Heart Institute (2018-2025). 1) ECHO-PREP was trained to identify key diagnostic views in TTE (n=530) and diagnostic and procedural views in TEE (n=2222) examinations to determine the level of quality images needed to do a M-TEER. 2) 4D TEE segmentation with automated mitral valve area (MVA) quantification (n=221), and (3) 2D TEE scallop-level segmentation of leaflets and sub-valvular structures (n=992). Results: Preliminary results on test sets showed 95.7% accuracy in TTE view classification and 91% accuracy for TEE view classification. The 4D segmentation module demonstrated excellent agreement with manual MVA measurements (R=0.84, p<0.001), successfully discriminating patients undergoing M-TEER from those referred for surgical replacement (p=0.046 for AI predictions). The 2D scallop-level analysis achieved a mean Dice score of 0.534 across 11 anatomical structures, with better performance in commonly represented configurations (e.g., A2-P2, P1-A2-P3). Conclusion: ECHO-PREP demonstrates the feasibility of an integrated AI-assisted workflow for MR assessment, combining quality control, dynamic 4D valve quantification, and scallop-level anatomy interpretation. These results support the potential of AI to standardize M-TEER eligibility, reduce inter-observer variability, and provide decision support across centers with different levels of expertise.