Automated retinal disease classification from optical coherence tomography images using deep learning

Purpose To develop and comparatively evaluate a reproducible Deep Learning (DL) framework for automated multiclass retinal disease classification from Optical Coherence Tomography (OCT) images in a retrospective comparative diagnostic accuracy study. Method A publicly available OCT dataset containing 2064 images from 820 patients was used to classify seven diagnostic categories: Age-Related Macular Degeneration (AMD), Diabetic Macular Edema (DME), Epiretinal Membrane (ERM), normal retina, Retinal Artery Occlusion (RAO), Retinal Vein Occlusion (RVO) and Vitreomacular Interface Disease (VID). To avoid data leakage, the dataset was partitioned at the patient level into training (1481 images; 592 patients), validation (281 images; 105 patients) and test (302 images; 123 patients) sets. Eight Convolutional Neural Network (CNN) architectures (InceptionV3, VGG16, MobileNetV2, AlexNet, EfficientNetV2B2, ResNet50, DenseNet121 and Xception) were trained within the same experimental pipeline using pretrained weights, weighted cross-entropy loss, Adaptive Moment Estimation with decoupled weight decay (AdamW) optimisation, early stopping and checkpointing. Performance was assessed using overall accuracy, balanced accuracy, macro Area Under the Receiver Operating Characteristic Curve (AUROC), per-class precision, recall and F1-score, confusion matrices and Expected Calibration Error (ECE), with temperature scaling applied for probability calibration. Results All evaluated models demonstrated strong diagnostic performance on the patient-disjoint test set. Xception achieved the best overall classification results, with 93.05% accuracy and 87.39% balanced accuracy. ResNet50 and DenseNet121 also performed strongly, achieving accuracies of 91.72% and 91.39%, respectively, with macro AUROC values above 0.98. ResNet50 produced the highest macro AUROC (0.99021) and the lowest calibrated ECE (0.02171), indicating strong discrimination and reliable probability estimates. Per-class analysis showed better performance for dominant categories such as AMD and normal retina, whereas minority classes remained more challenging. Conclusions A reproducible DL framework for OCT-based retinal disease classification was developed and shown to perform effectively across seven retinal diagnostic categories. Among the evaluated architectures, Xception provided the best overall classification performance, while ResNet50 offered the strongest discrimination and calibration. These findings support the potential of DL as a reliable tool for automated retinal disease detection from OCT images. Clinical trial number Not applicable.