Preliminary study on the quantitative assessment model of mitral regurgitation in echocardiography based on fully convolutional networks: automatic identification and measurement of regurgitant radius
10.3760/cma.j.cn131148-20240726-00409
- VernacularTitle:基于全卷积神经网络的超声心动图二尖瓣反流定量评估模型初步研究：反流半径的自动识别与测量
- Author:
Lu ZHONG
1
;
Hongning SONG
;
Bo HU
;
Qing DENG
;
Jinling CHEN
;
Qing ZHOU
;
Fengxia JIANG
;
Sheng CAO
Author Information
1. 武汉大学人民医院超声影像科，武汉 430060
- Publication Type:Journal Article
- Keywords:
Echocardiography;
Mitral regurgitation;
Fully convolutional networks;
Artificial intelligence
- From:
Chinese Journal of Ultrasonography
2025;34(2):98-106
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop an artificial intelligence system using fully convolutional neural networks（FCN）to assist echocardiographers in the quantitative assessment of mitral regurgitation（MR）severity.Methods:From August 2021 to June 2024，echocardiographic images of 441 patients with MR were prospectively collected from Renmin Hospital of Wuhan University and the Central Hospital of Wuhan. After screening，a total of 269 patients（4 917 frames）were included in the study. Of these，3 644 frames（128 patients）of apical four-chamber color Doppler MR flow convergence images from Renmin Hospital of Wuhan University were selected as the training/validation set，while images from 121 patients（813 frames）were used as the internal test set. Additionally，images from 20 patients（460 frames）from the Central Hospital of Wuhan were selected as the external test set. The FCN algorithm was employed to capture features and segment the MR color region on the left atrial side，simultaneously outputting the regurgitant radius（r）for the calculation of the effective regurgitant orifice area and regurgitant volume. The severity of MR was then classified according to the 2017 guidelines of the American Society of Echocardiography. The segmentation and classification performance of the model was evaluated，and the measurement results of the AI system was compared with that of both senior and junior physicians.Results:In the internal test set，the accuracy of r identification for cases classified as Grade Ⅰ to Ⅳ was 0.48，0.81，0.86，and 0.87，respectively. In the external test set，the accuracy of r identification for cases classified as Grade Ⅰ to Ⅳ was 0.60，0.77，0.64，and 0.77，respectively. The average accuracy of MR classification in the internal and external test sets was 0.91 and 0.88，respectively.Conclusions:The FCN model is capable of segmenting the left atrial side regurgitant areas in apical four-chamber heart color Doppler images，aiding physicians in obtaining quantitative assessment parameters for MR，and assisting junior physicians in accurately assessing the severity of MR.
