Explainable Artificial Intelligence for Early Detection of Glaucomatous Optic Neuropathy

Abstract

Background: Glaucoma is a leading cause of irreversible visual loss, and early structural damage may precede clinically detectable functional impairment. Explainable artificial intelligence may support earlier identification of glaucomatous optic neuropathy while providing visual information about image regions contributing to model predictions.

Objective: To evaluate the diagnostic performance of an explainable artificial intelligence model for glaucomatous optic neuropathy, determine its sensitivity for early-stage disease, and assess anatomical concordance of Grad-CAM saliency maps with clinically relevant optic nerve structures.

Methods: This prospective hospital-based diagnostic accuracy study included 60 adults undergoing glaucoma evaluation at a tertiary care teaching hospital in Pakistan from January to June 2026. Clinical examination, fundus photography, optical coherence tomography, and standard automated perimetry were used to establish the reference diagnosis. A previously trained convolutional neural network classified fundus images as glaucomatous or nonglaucomatous. Diagnostic performance was assessed using sensitivity, specificity, predictive values, and accuracy, while Grad-CAM maps were reviewed for anatomical concordance.

Results: Thirty patients had glaucomatous optic neuropathy and 30 were nonglaucomatous. The model achieved 90.0% sensitivity, 86.7% specificity, 87.1% positive predictive value, 89.7% negative predictive value, and 88.3% overall accuracy. Early-stage sensitivity was 83.3%, whereas all 12 moderate-to-advanced cases were detected. Forty-nine of 60 Grad-CAM maps (81.7%) were considered anatomically concordant with clinically relevant optic nerve regions.

Conclusion: Explainable AI demonstrated good diagnostic performance for glaucomatous optic neuropathy and clinically relevant spatial attribution in most cases. Lower sensitivity for early-stage disease and the small single-centre sample indicate that larger independent validation studies are required before routine clinical implementation.