SwinMedNet: Leveraging Swin Transformer for Robust Diabetic Retinopathy Classification from the RetinaMNIST2D Dataset

Diabetic retinopathy (DR) is a leading cause of blindness, making early detection crucial. This study investigates the performance of a Swin Transformer-based deep learning model to classify DR severity in fundus images. In this research, a Swin Transformer architecture was employed, known for its global context learning capabilities, for classifying fundus images into five levels of severity of DR: no apparent retinopathy, mild non-proliferative DR (NPDR), moderate NPDR, severe NPDR, neovascularization, and vitreous/preretinal hemorrhage (PDR). The model was trained and evaluated on a publicly available dataset of fundus images with corresponding DR labels. Performance metrics included accuracy and Area Under the ROC Curve (AUC) for each class. The Swin Transformer model achieved higher accuracy than state-of-the-art research works i.e. 56.8%. The AUC value, which was 83.4% also indicated strong discriminatory power for each class. Our model's performance surpassed previous studies utilizing different deep learning architectures, demonstrating the effectiveness of Swin Transformers for DR classification. This study demonstrates the potential of Swin Transformers for accurate and reliable DR classification in fundus images. This approach could contribute to automated DR screening systems, aiding in early diagnosis and timely intervention, potentially improving patient outcomes and preventing vision loss.