Improving Diabetic Retinopathy Diagnosis with Transfer Learning and Filtered Fundus Images: An Automated Approach

Diabetic retinopathy (DR) is a critical eye disease affecting individuals with diabetes, leading to potential vision loss if not diagnosed and treated promptly. In this study, we proposed an automated system for the detection of diabetic retinopathy. This system incorporates a preprocessing stage, wherein a 3D image is constructed from the optimal output of a selection of three filters. Subsequently, a model is trained through transfer learning, enabling it to extract pertinent features. Classification is performed through the utilization of a fully connected layer. Experimental results on the Messidor dataset demonstrate the effectiveness of the proposed methodology. EfficientNet achieves the highest accuracy of 95% among the explored deep models, highlighting its potential for accurate diabetic retinopathy detection.