Automatic 3D mitral valve leaflet segmentation and validation of quantitative measurement. (January 2023)

• Record Type: Journal Article
• Authors: Chen, Jinhui; Li, Hanzhao; He, Gaowei; Yao, Fengjuan; Lai, Lixuan; Yao, Jianping; Xie, Longhan
• Is Part Of: Biomedical signal processing and control. Volume 79(2023)Part 2
• Publication Date: 2023-01
• Subjects: Mitral valve -- Image segmentation -- Deep learning -- Quantitative measurement -- Pre-training; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104166
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: 3D transesophageal echocardiography (TEE) is widely used in the diagnosis of mitral valve disease and is also well suited for guiding cardiac interventions. The aim of this work is to achieve patient-specific 3D TEE mitral valve leaflet segmentation without any user interaction and to assess the feasibility of 3D quantitative measurements on automatic segmentation model. We suggested a novel pre-training strategy to better implement automatic segmentation. The strategy refers to classify the diastolic and systolic states of the mitral valve through a 3D convolutional neural network architecture, and then use the pretrained weights obtained from the classification task to initialize the parameters of the 3D segmentation deep learning framework. To determine the accuracy of geometric parameters of segmentation model, the measurements of the segmentation model were compared with those obtained by the clinical software. Statistical analysis was performed by using Intraclass Correlation Coefficient and Bland–Altman method. Fourteen 3D volumes were used to evaluate the segmentation performance. The results show a Dice Similarity Coefficient (DSC) of 0.877±0.027 and an Average Surface Distance (ASD) of 0.925±0.392 mm. Twenty-eight 3D volumes were used for the quantitative measurement. The statistical results show that the mitral annular parameters have a good agreement between segmentation model and clinical software except for the annular height. We developed a fully automatic methodology to segment the mitral valve leaflet from 3D TEE and demonstrated the feasibility of improving segmentation performance with the proposed pre-training strategy. The automatic segmentation model was proved to be reliable for performing quantitative measurements of mitral valve annulus dimensions. The results indicate that the precision of the automatic segmentation methodology could pave the way for application in quantification, modeling and surgical planning tools.