Self-supervised learning of 3D structure from 2D OCT slices for retinal disease diagnosis on UK biobank scans

This study presents a self-supervised learning framework for retinal disease classification using Optical Coherence Tomography (OCT) scans. To balance the contextual richness of 3D volumes with the computational efficiency of 2D architectures, we introduce a quasi-3D input generation strategy. Each input is constructed by stacking three OCT slices, sampled from channel-specific Gaussian distributions centered on the volume midplane, and arranged in a standard three-channel 2D format compatible with existing pre-trained models. These quasi-3D images are used to pre-train a Vision Transformer (ViT-Base) via a Masked Autoencoder (MAE) with a shared masking pattern, encouraging the model to reconstruct masked regions by encoding anatomical continuity across slices. Pre-training is conducted on 10,000 unlabeled OCT volumes from the UK Biobank. The encoder is then fine-tuned on the OCTA-500 dataset for three-class and four-class retinal disease classification tasks, including macular degeneration and diabetic retinopathy. The model achieves 92.57% accuracy on the three-class task, matching the performance of RETFound while using over 150 times less pre-training data and a smaller backbone.