3D Retinal Reconstruction and Deep Learning for Diagnosis of Dry and Wet Age-Related Macular Degeneration from Optical Coherence Tomography Images

Abstract

This study presents a 3D Optical Coherence Tomography (OCT) image reconstruction workflow, developed primarily for OCT systems that do not natively export volumetric data, together with a deep learning-based framework for classifying age-related macular degeneration (AMD) into its two primary subtypes, dry and wet AMD. The reconstruction module aligns and interpolates the gaps between sequential B-scans to produce high-quality volumetric visualizations of retinal structures and is used purely for visualization, independently of the classification pipeline. Classification is performed by a binary convolutional neural network trained directly on stacks of the original, noninterpolated 2D B-scans, treating each set of 25 aligned slices per eye as a single input tensor. Across five-fold cross-validation, the proposed model achieved an average F1-score of 0.84 ± 0.15, an accuracy of 0.88 ± 0.10, and an AUC of 0.92 ± 0.09, with a peak fold-level validation accuracy of 96.00%. These results indicate that preserving cross-slice spatial context substantially improves classification performance over single-slice or unaligned baselines. By supporting both visualization and automated subtype classification, the presented approach has the potential to assist ophthalmologists in the timely and accurate diagnosis and management of AMD.