DIAGNOSTIC PERFORMANCE OF MACHINE LEARNING TECHNOLOGY USING OPTICAL COHERENCE TOMOGRAPHIC IMAGE IN RETINAL DISEASES PRESENTED WITH SUBRETINAL FLUID.

Abstract

To study the diagnostic performance of machine learning in the diagnosis of three retinal diseases presented with subretinal fluid: central serous chorioretinopathy, polypoidal choroidal vasculopathy, and Vogt-Koyanagi-Harada disease by using optical coherence tomography (OCT) images. Optical coherence tomography scans from 259 patients presented with subretinal fluid from central serous chorioretinopathy, polypoidal choroidal vasculopathy, and Vogt-Koyanagi-Harada disease, and 108 patients with no subretinal fluid were used to train the machine learning. Three methods include using all scans from each eye, using only one-line across fovea images and using both infrared and OCT images. The result was reported in precision, recall, F-1 score, accuracy, and area under the curve. Using one-line across the fovea OCT images combined with an infrared photo achieved the best performance with an overall area under the curve score of 0.965 and an accuracy of 87.10%. The precision, recall, and F1-score were 88.43%, 87.19%, and 87.18%, respectively. The authors introduced a new aspect of developing a machine learning model to differentiate among three major retinal diseases presented with subretinal fluid from OCT images. The use of one-line across fovea OCT images combined with infrared photograph showed the highest diagnostic performance because of the possible imaging biomarkers.