3‑D TEE Mitral Valve Segmentation and Mesh Reconstruction with Real‑Time Quality Assurance

Phat Kim Huynh, Jacques Kpodonu, Minh Huu Nhat Le, Dang Nguyen, Phi Huynh, Thuan Quang Phan, Olabiyi Hezekiah Olaniran, Tam Tran, Heath Rutledge-Jukes, Yanwen Xu, Dinh Hoang Nguyen

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: mitral valve, transesophageal echocardiography, image segmentation, vision transformers, quality assurance, mesh reconstruction

TL;DR: A transformer-based pipeline segments 3-D TEE mitral valves, flags unreliable cases via single-pass entropy, and exports watertight meshes

Abstract: Accurate, real‑time segmentation of the mitral valve (MV) from three‑dimensional transesophageal echocardiography (3‑D TEE) remains a technical bottleneck for intra‑procedural guidance and patient‑specific modeling. We present a vision‑transformer based pipeline that performs voxel‑wise MV segmentation, entropy‑driven quality assurance (QA), and instantaneous mesh export on the public MVSeg2023 dataset. A Swin‑UNETR backbone was trained on 105 training and 30 validation volumes (voxel size 0.6 mm). The model employs temperature‑scaled logits and computes per‑voxel Shannon entropy to flag uncertain predictions. Leaflet meshes are generated on‑the‑fly via marching cubes and Taubin smoothing. On 40 held‑out test volumes the method achieved a Dice coefficient of $0.83\,\pm\,0.05$ and a 95$^{\text{th}}$‑percentile Hausdorff distance of 4.2 mm. 100% cases passed the QA gate ($\text{Entropy}_{95}<0.80$), with no significant correlation between entropy and Dice ($r=0.016,\,p=0.92$). End‑to‑end inference, including meshing, averaged 104 ms per volume on a single NVIDIA A100 GPU while peaking at 4.0 GiB of video memory. The proposed echo‑only transformer pipeline delivers state‑of‑the‑art accuracy, built‑in reliability estimates, and real‑time performance, satisfying key clinical usability requirements. To our knowledge this is the first 3‑D TEE MV segmentation framework that unites transformer representations, uncertainty‑aware QA, and mesh reconstruction within 0.1 s, paving the way for routine intra‑procedural deployment.

Track: 3. Imaging Informatics

Registration Id: YRNT5JKVRS7

Submission Number: 312