基于深度学习算法的前房角超声生物显微镜图像自动评估系统

Automatic evaluation system for ultrasound biomicroscopy images of anterior chamber angle based on deep learning algorithm

摘要:

目的：探讨一种基于深度学习算法的前房角(ACA)超声生物显微镜(UBM)图像分析系统的临床应用价值。

方法：收集2021-01/2022-06于武汉大学人民医院眼科中心进行UBM检查的受试者675名1 130眼的UBM图像4 196张构建图像数据集。采用Unet++网络对ACA组织自动分割，并开发一种支持向量机(SVM)算法对房角开闭状态进行自动分类，同时开发一种自动定位巩膜突、测量ACA参数的算法。另选取黄石爱尔眼科医院的受试者127名221眼的UBM图像631张和武汉大学中南医院的受试者188名257眼的UBM图像594张评估该系统在不同环境下的性能。

结果：本研究构建的分析系统对房角开闭状态识别的准确度为95.71%； ACA角度参数测量值的组内相关系数(ICC)均大于0.960，ACA厚度参数测量值的ICC均大于等于0.884，且该系统对ACA参数的准确测量部分依赖于巩膜突的准确定位。

结论：本研究构建的智能分析系统能够准确有效地自动评估ACA图像，是一种有潜力的快速识别ACA结构的筛查工具。

Abstract:

AIM: To explore the clinical application value of analysis system for ultrasound biomicroscopy(UBM)images of anterior chamber angle(ACA)based on deep learning algorithm.

METHODS: A total of 4 196 UBM images were obtained from 675 patients(1 130 eyes)at the Eye Center of Renmin Hospital of Wuhan University from January 2021 to June 2022 were collected to build an image dataset. Using Unet++network to automatically segment ACA tissue, a support vector machine(SVM)algorithm was developed to automatically classify opening and closing of chamber angle, and an algorithm to automatically locate the sclera spur and measure ACA parameters was developed. Furthermore, a total of 631 UBM images of 127 subjects(221 eyes)at Huangshi Aier Eye Hospital and 594 UBM images of 188 subjects(257 eyes)at Zhongnan Hospital of Wuhan University were selected to evaluate the performance of the system under different environments.

RESULTS: The accuracy of the analysis system constructed in this study for chamber angle opening and closing was 95.71%. The intra-class correlation coefficient(ICC)values of all ACA angle parameters were greater than 0.960. ICC values of all ACA thickness parameters were greater than 0.884. The accurate measurement of ACA parameters depended in part on the accurate location of the scleral spur.

CONCLUSION: The intelligent analysis system constructed in this study can accurately and effectively evaluate ACA images automatically and is a potential screening tool for the rapid identification of ACA structures.