Vision Transformer-Based Swin Transformer for Automated Diabetic Retinopathy Severity Detection and Classification

Diabetes-related retinopathy (DR) can cause permanent vision loss if not detected early. CNNs are widely used to classify DR however, they can't detect global spatial linkages and small progression patterns needed to assess disease severity accurately. To address the major CNN issues, the proposed research creates STBDR-Net (Swin Transformer-Based Diabetic Retinopathy Network) to solve these problems. Swin Transformer hierarchical attention methods are used in this design. A 224x224 Gaussian-filtered EyePACS dataset is used to train the model; this strategy clarifies the lesion and ensures the transformer framework receives input with the exact resolution. There are five DR severity levels: nil, mild, moderate, severe, and proliferative. Complex preprocessing methods, including contrast normalization and attention-based filtering, ensure accurate classification for all five DR severity levels in STBDR-Net. STBDR-Net outperforms CNNs and baseline transformer models in terms of accuracy, F1-score, and weighted metrics. This is true when comparing the two methods. Mild and moderate DR might confuse automated systems. The model distinguishes mild and moderate DR well. STBDR-Net is a trustworthy and useful real-world DR screening tool. Visual transformer-based learning samplings aid scalable, accurate clinical decision-making.