Automatic 3D Segmentation of Closed Mitral Valve Leaflets on Transesophageal Echocardiogram

Abstract

Heart disease is a leading cause of death worldwide, with mitral valve (MV) disease being among the most prevalent pathologies. The MV, constitutes a complex three-dimensional apparatus which makes clinical assessment challenging. Therefore, it would be highly desirable to have a patient-adapted model of the mitral annulus shape and its leaflets, both for diagnosis and intervention planning, as well as
follow-up purposes.
The main objective of this work is two-fold: improve the valve segmentation’s quality using modern architectures and extend it to a sequence of 3D ultrasound recordings for the entire systolic phase. For training purposes, we used a dataset consisting of 108 volumes that were semiautomatically segmented using a commercially available package. We tested several network architectures and loss functions available in the MONAI package to investigate which ones are best suited for the task at hand. We aimed for fast processing times that were usable in practice. Our method was evaluated on 30 recordings and compared to annotations made by two expert echocardiographers. The comparison metrics include Average Surface Distance (ASD), Hausdorff Distance 95% (HSD 95%), as well as standard classification metrics. Our results were a Dice score of 77.06±13.18 % on the evaluation test and distance errors of 0.09±0.12 mm for ASD and 0.49±0.43 mm for HSD 95% and the segmentations were considered comparable to the ground truth by clinicians. The proposed annotation method was significantly faster than one of the previous works and yielded results comparable to the state-of-the-art using a noisier ground truth.