Automated Retinal Disease Classification Using Fundus Images: A Multi-Class Analysis

Early diagnosis of retinal disease is necessary to prevent vision loss and to facilitate early medical treatment. The proposed deep learning system evaluates retinal diseases in fundus images by employing ConvNeXt-Tiny, InceptionV3, Vision Transformer (ViT), and Swin Transformer models. The study uses the Retinal Disease Classification dataset containing 3,200 fundus images across 46 diverse classes of retinal diseases. The data was resized, normalized, and special data augmentation methods were used for improving model performance and overfitting avoidance. Each model was subjected to certain training methods like dynamic adaptation of learning rate, selected optimizers, and batch sizes aligned with the model architecture complexity. ConvNeXt-Tiny obtained the highest test accuracy at 98.4%, InceptionV3 at 97.57%, Swin Transformer at 93.44%, and ViT at 90.94%. It can be observed that convolution-based approaches outperform in retrieving fine local details, whereas transformer-based approaches prove efficient in identifying global dependencies. The main challenges from the research were the class imbalance problem alongside small dataset volume and high computational requirements. Dataset expansion, real-time clinical assessment, and use of ensemble learning and self-supervised methods will be resolved in the future to improve the accuracy of classification and enable wider clinical usage.