Multi-Disease Classification in Retinal Oct Using Deep Transfer Learning: A Comparative Study

Abstract

Timely and precise detection of retinal diseases is essential for preventing permanent vision loss. However, the analysis of Optical Coherence Tomography (OCT) images is often time-taking and subject to inter-clinician variability. This paper presents deep learning architectures for multi-disease classification of OCT images using transfer learning and finetuning techniques with four state-of-the-art convolutional neural network architectures: MobileNetV2, DenseNet121, ResNet50, and EfficientNetV2. The present work incorporates standardized preprocessing, comprehensive data augmentation, and a twostage training strategy consisting of classifier-head training followed by progressive fine-tuning. To ensure a fair comparison among models, a unified training and evaluation pipeline was implemented using the OCT-C8 dataset, consisting of eight image categories that include seven retinal disease classes and one healthy class. Model performance was evaluated using accuracy, precision, recall, and F1-score metrics. Experimental results demonstrate that ResNet50 achieved the highest classification accuracy of 97.25%, followed closely by DenseNet121 with 97.11%. Additionally, Grad-CAM visualisations were employed to interpret model predictions by identifying the most influential regions within OCT images. The study highlights the effectiveness of transfer learning for OCT image classification and provides a systematic comparison between lightweight and heavyweight CNN architectures. Although the proposed framework demonstrates strong classification performance, it is intended as a comparative research study and is not designed for clinical deployment at this stage.