A Transformer-Enabled Approach for Diabetic Retinopathy Classification from OCT Imaging

It is well established that diabetic retinopathy (DR) is a leading cause of blindness; therefore, there is an urgent need for automated and scalable mechanisms for screening DR patients. In this study, we propose a framework to predict DR from optical coherence tomography (OCT) images using a Vision Transformer (ViT) based architecture. This method is unique in that it does not require the traditional method of manually extracting retinal features from OCT images or manual segmentation of OCT images into different layers for the purposes of building machine learning models. Instead, the proposed framework builds a feature set directly from the OCT scans by means of a proprietary patch embedding and transformer-based self-attention model that can learn to identify discriminative patterns from the OCT scans. The proposed framework consists of a preprocessing step, as well as a training/testing partitioning of the datasets using a stratified partitioning approach and a transformer-based training process with an additional dense decision final layer utilizing transfer learning techniques (AdamW and cross-entropy loss) and utilizing backpropagation. In trials conducted on the OLIVES dataset, the proposed method converged very quickly and exhibited excellent generalization capabilities, achieving 98.35 % validation accuracy after eighty epochs. The results indicate that the ViT has the capability to interpret the OCT biomarkers and provides a solid basis for DR screening in a clinical setting. Furthermore, the proposed approach highlights the advantages of using transformer-based self-attention for the automated diagnosis of retinal diseases and the future potential for computer-aided healthcare for detecting retinal diseases.