3D-CNN for Glaucoma Detection Using Optical Coherence Tomography

Abstract

The large size of raw 3D optical coherence tomography (OCT) volumes poses challenges for deep learning methods as it cannot be accommodated on a single GPU in its original resolution. The direct analysis of these volumes however, provides advantages such as circumventing the need for the segmentation of retinal structures. Previously, a deep learning (DL) approach was proposed for the detection of glaucoma directly from 3D OCT volumes, where the volumes were significantly downsampled first. In this paper, we propose an end-to-end DL model for the detection of glaucoma that doubles the number of input voxels of the previously proposed method, and also boasts an improved AUC = 0.973 over the results obtained using the previously proposed approach of AUC = 0.946. Furthermore, this paper also includes a quantitative analysis of the regions of the volume highlighted by grad-CAM visualization. Occlusion of these highlighted regions resulted in a drop in performance by 40%, indicating that the regions highlighted by gradient-weighted class activation maps (grad-CAM) are indeed crucial to the performance of the model.