Development of a risk assessment tool for early detection of glaucoma based on machine learning methods

Abstract

Glaucoma is a progressive optic neuropathy and a major cause of irreversible blindness, making reliable early screening particularly important. This study proposes an automated glaucoma risk assessment framework based on color fundus photographs, combining a ResNet-50 classifier with cycle-consistent generative adversarial network (CycleGAN)-based synthetic data augmentation. The model was evaluated on the retinal image database for optic nerve evaluation for deep learning (RIM-ONE DL) dataset under limited and imbalanced data conditions. Training on real images alone resulted in an accuracy of 76.8%, recall of 74.2%, and area under the receiver operating characteristic curve (AUC-ROC) of 0.825. After adding 500 synthetic normal and 500 synthetic glaucoma images to the training set, accuracy increased to 91.4%, recall to 89.7%, and AUC-ROC to 0.943. Synthetic image quality was additionally evaluated using Fréchet inception distance (FID), kernel inception distance (KID), and learned perceptual image patch similarity (LPIPS) metrics and expert visual assessment. The results demonstrate that CycleGAN-based augmentation can substantially improve glaucoma classification performance under limited-data conditions while preserving clinically relevant retinal structures. The proposed approach may therefore serve as a supplementary decision-support tool for automated glaucoma screening.