Optic Vision: Modified U-Net Techniques for Detecting Diabetic Retinopathy through Segmentation of Exudates

Diabetic retinopathy (DR) is a significant complication of diabetes that can lead to permanent vision loss if not identified early. This study aims to develop and evaluate a U-Net-based architecture specifically for the segmentation of exudates in fundus images, improving automated DR analysis. We designed the model with additional convolutional layers and dropout techniques to enhance feature extraction and reduce overfitting. The model was trained on the Indian Diabetic Retinopathy Image Dataset (IDRiD) for 100 epochs, utilizing binary cross-entropy loss and optimizing with the Adam optimizer. Performance evaluation included validation accuracy, validation loss, and Intersection over Union (IoU) metrics, with post-processing applying Otsu thresholding at a value of 0.2324 for exudates. Our method achieved a validation accuracy of 1.0 and a validation loss of 0.0178, indicating exceptional segmentation performance. These results suggest that our approach significantly enhances the detection of crucial retinal structures, facilitating earlier identification of diabetic pathologies and demonstrating superior performance compared to existing studies in retinal lesion segmentation. Future work may explore the model’s application to larger and more diverse datasets.