Automated Diagnosis of Eight Retinal Diseases from OCT Images Using Deep Learning with Attention Mechanisms

Optical Coherence Tomography (OCT) is a critical imaging modality for diagnosing retinal diseases, yet manual interpretation of OCT scans remains time-consuming and subjective. This study proposes a robust deep learning framework for multi-class OCT image classification using the RetinalOCT-C8 dataset, which comprises eight clinically relevant categories: Age-related Macular Degeneration (AMD), Choroidal Neovascularization (CNV), Central Serous Retinopathy (CSR), Diabetic Macular Edema (DME), Diabetic Retinopathy (DR), Drusen, Macular Hole (MH), and Normal retina. We introduce a ResNet-based architecture with Attention Residual Classifiers (ARC), featuring: (1) A pretrained ResNet backbone with adaptive feature extraction (128-D embedding), (2) A novel Squeeze-and-Excitation (SE) enhanced attention block for class-specific feature refinement, and (3) Residual connections to preserve discriminative spatial-semantic hierarchies. The model achieves 97.75% accuracy and 97.75% macro F1-score on the validation set, with perfect 1.0 precision-recall for AMD, CSR, DR, and MH classes. Notably, it maintains high performance on challenging cases ($\mathbf{9 4. 8 6 \% ~ F 1}$ for CNV, $\mathbf{9 4. 8 1 \%}$ for DRUSEN).