Recognizing Abnormalities in Fundus Images Using Vision Transformer for Ocular Diseases

In today’s scenario of increasing ocular diseases, ocular disease detection requires advanced diagnostic tools for early detection and treatment. Exploring the advanced deep learning technique, this study introduces a customized Vision Transformer (ViT) for recognizing abnormalities in fundus images. These images are crucial means for diagnosing conditions related to ocular diseases. On the other hand, ViT leverages an inherent self-attention mechanism. The proposed customized ViT enhances the extraction of relevant features from retinal images. ViTs are capable of outperforming convolutional neural networks (CNNs). We employed this methodology on a comprehensive dataset comprising annotated fundus images representing normal and abnormal classes. The proposed model demonstrates satisfactory performance having classification accuracy of 87.16% even being the dataset class-imbalanced. Furthermore, our approach facilitates gaining satisfactory performance over imbalanced datasets, as well as abnormal images that may have multilabels, allowing for the simultaneous detection of multiple ocular conditions within a single image. The findings indicate that ViT can effectively capture complex patterns in fundus images, thereby enhancing the potential for timely interventions in ocular health. This work contributes to the use case of transformer-based models in medical imaging. Through this study, we aim to pave the way for integrating advanced machine/deep learning techniques into clinical practice for better management of ocular diseases.