Diagnosing Retinal Eye Diseases: A Novel Transfer Learning Approach

This study rigorously evaluates the potential of transfer learning in diagnosing retinal eye diseases using advanced models such as YOLOv8, Xception, ConvNeXtTiny, and VGG16. All models were trained on the esteemed RFMiD dataset, which includes images classified into six critical categories: Diabetic Retinopathy (DR), Macular Hole (MH), Diabetic Neuropathy (DN), Optic Disc Changes (ODC), Tesselated Fundus (TSLN), and normal cases. The research emphasizes enhancing model performance by prioritizing recall metrics, a crucial strategy aimed at minimizing false negatives in medical diagnostics. To address the challenge of imbalanced data, we implemented effective preprocessing techniques, including cropping, resizing, and data augmentation. The proposed models underwent fine-tuning and were evaluated using established metrics such as accuracy, precision, and recall. The experimental results are compelling, with YOLOv8 achieving the highest recall rates for both normal cases (97.76%) and DR cases (87.10%), demonstrating its reliability in disease screening. In contrast, Xception showed a decline in recall after fine-tuning, particularly in identifying DR and MH cases, highlighting the need for a careful balance between sensitivity and specificity in model training. Notably, both ConvNeXtTiny and VGG16 exhibited significant improvements post-fine-tuning, with VGG16’s recall for normal conditions increasing dramatically from 40.30% to an impressive 89.55%. These findings clearly underscore the potential of utilizing pre-trained models through transfer learning for the effective detection of retinal eye diseases, ultimately contributing to improved patient outcomes in medical diagnostics.