Deep Learning for Differentiating Glaucomatous and Non-glaucomatous Optic Neuropathy with Retinal Nerve Fiber Layer and Optic Disc Photographs
13 Dec 2025
10:15
10:30
Yun Jeong LeeSouth Korea
Speaker
Deep Learning for Differentiating Glaucomatous and Non-glaucomatous Optic Neuropathy with Retinal Nerve Fiber Layer and Optic Disc PhotographsPurpose: To develop a deep learning (DL)-based algorithm to differentiate glaucomatous optic neuropathy (GON) and non-glaucomatous optic neuropathy (NGON) with retinal nerve fiber layer (RNFL) and optic disc photographs.
Methods: A total of 765 image pairs (618 GON, 147 NGON) were retrospectively collected and preprocessed using histogram matching and region-of-interest cropping based on Hough circle detection. An external validation dataset consisting of 177 pairs (103 GON, 74 NGON) was also used. DL models were developed using ResNet34 for optic disc images and DenseNet121 for RNFL images, with feature-level fusion implemented via cross-attention mechanisms. Model training involved data augmentation, class imbalance correction, and five-fold cross-validation. Model interpretability was assessed using Grad-CAM visualization.
Results: The proposed model demonstrated robust performance in both internal and external datasets. In the internal validation set, the DL model achieved an AUC of 0.98 with the RNFL images, which was comparable to that with the optic disc (AUC of 0.99, P = 0.23) or combined RNFL and optic disc images (AUC of 0.98, P = 0.70), and was significantly superior to that with the masked RNFL (AUC of 0.94, P < 0.05) or combined masked RNFL and optic disc images (AUC of 0.96, P < 0.05). In the external validation set, the fusion model integrating both RNFL and optic disc images achieved superior diagnostic performance compared to single-modality models, with the highest AUC (0.88).
Conclusions: The proposed multi-input DL model effectively differentiates between GON and NGON using RNFL and optic disc photographs. By integrating structural features via cross-attention, the model achieves consistent diagnostic performance, even in external datasets. This suggests the potential value of our DL model in clinical practice by helping clinicians make accurate diagnoses and treatment decisions.