Introduction: Accurate interpretation of two-dimensional transesophageal echocardiography (2D TEE) is essential for successful mitral valve repair, yet echocardiographic assessment is highly operator dependent and subject to variability. While automation has been explored in transthoracic echocardiography, standardized interpretation of 2D TEE at the scallop level remains largely unexplored. This study investigates the feasibility of automating key components of preoperative 2D TEE interpretation to support standardization of mitral valve surgery planning.
Methods: Preoperative 2D TEE recordings from 50 patients undergoing mitral valve repair were retrospectively collected. Separate algorithmic components were developed and evaluated for (1) automatic classification of standard mid-esophageal views and visible mitral scallops, (2) anatomical segmentation of mitral valve–related cardiac structures, (3) automated extraction of clinically relevant geometric measurements combined with end-systolic frame identification, and (4) automated detection of mitral valve prolapse. Deep learning–based models were trained using patient-level data splits, and performance was assessed using standard classification, segmentation, and agreement metrics. Measurement feasibility was evaluated in a subset of 10 mid-esophageal long-axis views.
Results: View classification achieved high accuracy across standard mid-esophageal views, while scallop-level classification showed moderate performance (accuracy 0.64), mainly limited by severe class imbalance. Segmentation of clinically relevant structures was moderate, with mean Dice scores of 0.76 for the mitral valve and 0.69 across measurement-critical structures. Automated geometric measurements showed good agreement with expert references in the mid-esophageal long-axis view, with limited systematic bias (Coaptation-septal distance −0.95 mm; aortic–mitral angle −3.28°; anterior-to-posterior leaflet (AL:PL) ratio −0.27; A2 length −3.85 mm). Larger deviations occurred in parameters sensitive to segmentation quality, particularly AL:PL ratio and A2 length, while wide limits of agreement reflected the small sample size. End-systolic frame identification was robust, with most predictions within two frames of expert annotation (mean difference 1.8). Prolapse detection achieved 0.90 accuracy, with one false positive.
Conclusion: This study demonstrates the feasibility of evaluating automated methods for multiple components of 2D TEE interpretation in mitral valve repair planning. View recognition, anatomical segmentation, automated measurement extraction, and prolapse detection showed encouraging performance within the limitations of the available dataset. Scallop-level interpretation remains constrained by limited data availability and class imbalance. These findings motivate further development using larger, multi-center datasets and suggest that standardized automated analysis may ultimately support improved consistency and enable large-scale, outcome-driven research in mitral valve surgery.

Digital subtraction angiography (DSA) is widely used to assess vascular changes during endovascular thrombectomy, but manual comparison of image pairs is challenging and subjective. This work presents a deep learning-based method for arterial change detection in cerebral DSA by combining inter-sequence registration with a siamese spatiotemporal U-Net for joint arterial segmentation and change detection. Experiments on data from the MR CLEAN registry demonstrate that the proposed registration method significantly improves accuracy compared with the best-performing baseline (p=0.0054) and that the change-detection network consistently identifies arterial changes in successfully co-registered DSA pairs (Dice = 0.70). A preliminary reader study indicated that marking these changes can improve inter-rater agreement of extended thrombolysis in cerebral infarction (eTICI) scoring and increase the detection rate of emboli in new territory (ENT).