Deep Learning Classification of Angle Closure based on Anterior Segment Optical Coherence Tomography.

Researchers

Journal

Modalities

Models

Abstract

To assess the performance and generalizability of a Convolutional Neural Network (CNN) model for automated, objective and high-throughput identification of primary angle closure disease (PACD) as well as PACD stage differentiation on anterior segment swept-source optical coherence tomography (AS-OCT). The spectrum of PACD include primary angle closure suspect (PACS), primary angle closure (PAC) and primary angle closure glaucoma (PACG) DESIGN: Cross-sectional PARTICIPANTS: Patients from three different eye centers across China and Singapore were recruited for this study. 841 eyes from the two Chinese centers were divided into 170 control eyes, 488 PACS, and 183 PAC + PACG eyes. An additional 300 eyes were recruited from Singapore National Eye Center as testing dataset, divided into 100 control eyes, 100 PACS, and 100 PAC + PACG eyes.Each participant underwent standardized ophthalmic examination including gonioscopy and AS-OCT imaging, and was classified by the presiding physician as either control, PACS, PAC or PACG. Deep Learning model Inception-v3 was used to train 3 different CNN classifiers: classifier 1 aimed to separate control vs. PACS vs. PAC+PACG; classifier 2 aimed to separate control vs. PACD; classifier 3 aimed to separate PACS vs. PAC+PACG. All classifiers underwent 5-fold cross validation and were evaluated on independent test sets from the same region, China. To assess the generalizability of these models across different cohorts of patients, trained classifiers were further tested using data from a different country, Singapore.Area under receiver operator characteristic curve (AUC), precision and recall.Classifier 1 achieved an AUC of 0.96 on validation set from the same region, but dropped to an AUC of 0.84 on test set from a different country. Classifier 2 achieved the most generalizable performance with an AUC of 0.96 on validation set from the same region and AUC of 0.95 on test set from a different country. Classifier 3 showed the poorest performance, with an AUC of 0.83 and 0.64 on datasets from the same region and different country respectively.CNN classifiers can effectively distinguish PACD from controls on AS-OCT with good generalizability across different patient cohorts. However, their performance is moderate when trying to distinguish PACS vs. PAC+PACG.Copyright © 2023. Published by Elsevier Inc.