Background: Intraoperative decision making for mitral valve repair relies on rapid, reliable assessment of leaflet anatomy from three-dimensional transesophageal echocardiography (3-D TEE). Current workflows depend on manual or semi-automated segmentation that is time-consuming, operator-dependent, and inconsistently exportable to meshes for sizing and simulation. Hypothesis: We hypothesize that an AI-powered 3-D TEE segmentation pipeline with single-pass, entropy-based quality assurance (QA) will deliver clinically actionable, real-time leaflet segmentations and clean mesh exports, supporting intraoperative repair planning and device sizing. Methods: We trained a transformer-based 3-D segmentation model (Swin-UNETR) on the MVSeg-2023 TEE dataset that consists of 150 end-diastolic TEE volumes. Volumes underwent intensity normalization, isotropic resampling to 0.6-mm voxels, and foreground cropping before inference. Temperature-scaled logits were converted to voxel-wise Shannon-entropy maps, a case-level quality score was thresholded using the validation set to gate low-confidence outputs. The primary endpoint was class-averaged Dice overlap for anterior and posterior leaflets. Secondary endpoints included boundary error (95th-percentile Hausdorff distance (HD95), average symmetric surface distance (ASSD)), mesh quality (non-manifold-edge rate), and end-to-end latency. Results: On the held-out test set, class-averaged Dice was high (0.832 ± 0.051). Boundary errors were low (HD95 = 4.2 ± 2.1 mm and ASSD = 0.39 ± 0.013 mm). Meshes were topologically clean (non-manifold edges = 0.21 ± 0.07%) and exported in real time (end-to-end time = 122 ± 35 ms with segmentation time = 104 ± 34 ms and peak GPU memory ≈4.0 GB). All test cases passed the entropy-based QA gate; within accepted cases, entropy and Dice were uncorrelated (Pearson correlation coefficient r = 0.016, p-value = 0.92). Across all test studies, the pipeline produced standardized leaflet masks and mesh exports automatically, enabling immediate downstream intraoperative support. Conclusions: Our transformer pipeline for TEE mitral-leaflet segmentation achieved high overlap accuracy, clean instant meshes, and real-time performance while providing a transparent QA signal, supporting its potential for intraoperative guidance and downstream computational modeling. Next steps include multi-vendor, multi-pathology external validation and extension to temporally consistent 4-D TEE.
Nguyen et al. (Tue,) studied this question.