Effective Deep Learning Framework for Retinal Abnormality Detection in OCT Imaging

Abstract

Retinal diseases such as age-related macular degeneration, diabetic macular edema, and central serous retinopathy can lead to irreversible vision loss if not detected early. Optical Coherence Tomography (OCT) imaging has emerged as a powerful, non-invasive tool for the visualization of retinal structural abnormalities. However, manual diagnosis from OCT scans requires specialized clinical expertise and can be time-consuming, especially for routine screening. In this study, we propose a comprehensive deep learning-based classification pipeline for the automatic detection of retinal diseases from OCT images. Five state-of-the-art convolutional neural network (CNN) architectures—DenseNet121, MobileNetV2, EfficientNetV2B2, Swin Transformer, and ResNet50—were trained and evaluated on three benchmark OCT datasets: OCTID, OCTDL, and OCT5k. Both quantitative analysis (accuracy, F1-score, precision, recall) and qualitative evaluation using Grad-CAM visualizations were conducted. Our results demonstrate that the proposed models, particularly EfficientNetV2B2 and Swin Transformer, exhibit high classification accuracy and strong generalization across datasets. The findings support the viability of deep learning in real-time OCT-based retinal disease screening.