Two-Stage Fine-Tuned Transfer Learning for Retinal Disease Classification from OCT Images: Comprehensive Model Comparison and Explainable AI Analysis

Abstract

This study presents a two-stage fine-tuned transfer learning approach for the automatic classification of retinal diseases from Optical Coherence Tomography (OCT) images. Four different ImageNet pre-trained deep learning architectures, namely ResNet50V2, DenseNet121, EfficientNetV2B0, and ConvNeXtTiny, were comprehensively evaluated on the OCTMNIST dataset at a resolution of 224×224. In the proposed two-stage training strategy, only the classification head is trained in the first stage, while all network parameters are fine-tuned in the second stage. The highest classification performance was achieved with the ConvNeXtTiny model, achieving 93.60% accuracy, 0.9809 AUC, and 0.9147 Cohen's Kappa. These results indicate a 16.0% increase in accuracy compared to the MedMNIST benchmark. To ensure the clinical interpretability of the models, the study performed a comprehensive explainable artificial intelligence (XAI) analysis using Grad-CAM, LIME, and SHAP methods. The experimental results demonstrate that fine-tuned transfer learning achieves high performance in retinal disease diagnosis and that XAI methods play a critical role in enhancing the clinical reliability of the models.