Explainable deep learning for the automated classification of macular diseases in OCT images

Abstract

Aim: Optical coherence tomography (OCT) is a widely used, noninvasive, rapid, and high-resolution imaging technique for diagnosing and monitoring macular diseases. Despite its clinical value, OCT image interpretation is time-consuming and requires expert knowledge, which may lead to inconsistencies in diagnosis. The objective of this study is to create an AI-based model that reliably and effectively categorizes macular diseases from OCT images, offering a workable solution in environments with restricted access to ophthalmology specialists.

Materials and Methods: A convolutional neural network model based on ResNet50 architecture was developed to classify OCT images into seven categories: age-related macular degeneration (AMD), diabetic macular edema (DME), epiretinal membrane (ERM), retinal artery occlusion (RAO), retinal vein occlusion (RVO), vitreomacular interface disease (VID), and normal (NO) controls. Grad-CAM was employed to enhance the interpretability of the model and support clinical usability.

Results: The model’s macro-averaged precision, recall, and F1-score were 0.943 (95% confidence interval [CI]: 0.941–0.960), 0.940 (95% CI: 0.941–0.960), and 0.940 (95% CI: 0.941–0.960), respectively, with an overall accuracy of 0.950 (95% CI: 0.941–0.960). Grad-CAM visualizations confirmed the model’s focus on relevant retinal regions, thus supporting diagnostic reliability and interpretability.

Conclusion: The explainable model demonstrated strong diagnostic performance and potential as a clinical decision-support tool, especially in environments with limited resources. The integration of explainable AI techniques, such as Grad-CAM, enhances trust in automated decision-making and offers significant potential in supporting non-expert users and early detection strategies.