ARTIFICIAL INTELLIGENCE'S ROLE IN DIFFERENTIATING THE ORIGIN FOR SUBRETINAL BLEEDING IN PATHOLOGIC MYOPIA.

Abstract

To identify salient imaging features to support human-based differential diagnosis between subretinal hemorrhage (SH) due to choroidal neovascularization (CNV) onset and SH without CNV (simple bleeding [SB]) in pathologic myopia eyes using a machine learning (ML)-based step-wise approach. Four different methods for feature extraction were applied: GradCAM visualization, reverse engineering, image processing, and human graders' measurements. GradCAM was performed on a deep learning model derived from Inception-ResNet-v2 trained with OCT B-scan images. Reverse engineering consisted of merging U-Net architecture with a deconvolutional network. Image processing consisted of the application of a local adaptive threshold. Available OCT B-scan images were divided in two groups: the first group was classified by graders before knowing the results of feature extraction and the second (different images) was classified after familiarization with the results of feature extraction. Forty-seven and 37 eyes were included in the CNV group and the simple bleeding group, respectively. Choroidal neovascularization eyes showed higher baseline central macular thickness ( P = 0.036). Image processing evidenced in CNV eyes an inhomogeneity of the subretinal material and an interruption of the Bruch membrane at the margins of the SH area. Graders' classification performance improved from an accuracy of 76.9% without guidance to 83.3% with the guidance of the three methods ( P =0.02). Deep learning accuracy in the task was 86.0%. Artificial intelligence helps identifying imaging biomarkers suggestive of CNV in the context of SH in myopia, improving human ability to perform differential diagnosis on unprocessed baseline OCT B-scan images. Deep learning can accurately distinguish between the two causes of SH.