Improved Diabetic Retinopathy Grading Using Preprocessing-Aided Convmixer-Swin Transformer Architecture

Diabetic retinopathy is a serious and irreversible eye disease that often results from prolonged high blood sugar levels in diabetic patients. The condition progressively damages the retina, leading to a gradual decline in vision and, if untreated, may cause permanent blindness. This fatal disorder can be remedied with accurate diagnosis and identification at an early stage. This work presented a novel approach for properly classifying distinct DR grades, ensuring that early DR situations could not be overlooked. A hybrid custom model combining ConvMixer and Swin Transformer is proposed for classifying DR grades ranging from No DR to Proliferative DR. In addition, a preprocessing strategy is employed to enhance the overall performance. To evaluate the effectiveness of the proposed method, two publicly available datasets, APTOS and MESSIDOR, were utilized. Another dataset was developed by integrating these two datasets to somewhat minimize class imbalance. The proposed model achieved 90.3% accuracy and an 88.7% weighted kappa score, demonstrating its robustness and reliability for DR classification, even with imbalanced and limited datasets.