Swin-transformer based recognition of diabetic retinopathy grade

Diabetic Retinopathy (DR), a common diabetes-related disorder, is a leading driver of blindness worldwide. Quick detection and precise staging are essential for effective management and vision preservation. This study explores the Swin Transformer, an advanced deep learning framework with a multi-layered setup and a unique sliding window method, to create an automated tool for DR stage assessment. Utilizing the APTOS 2019 Blindness Detection dataset, the system accurately identifies small retinal signs like microaneurysms and more pronounced features such as hemorrhages, achieving high precision. Improved preprocessing, including image enrichment and calibration, enhances its versatility. Results indicate that this approach outperforms traditional Convolutional Neural Networks (CNNs) in precision, computational thrift, and growth potential, with a test accuracy of 99.57% and a test loss of 0.0220.