Enhancing Exudate Segmentation in Fundus Images using CLAHE and ResNet U-Net Architecture

Diabetic Retinopathy remains a leading cause of irreversible blindness worldwide, with hard exudates serving as a critical early biomarker for diagnosis. However, automated exudate segmentation is often hampered by low image contrast and the visual similarity between lesions and the optic disc. The proposed method is an optimized segmentation framework utilizing a U-Net architecture with a ResNet-18 backbone. To enhance lesion discriminability, a preprocessing pipeline consisting of background removal and contrast-limited adaptive histogram equalization (CLAHE) was implemented. The model was trained and evaluated on 81 fundus images from the IDRiD dataset with a 512×512 input resolution. The experimental results demonstrate that the integration of CLAHE and ResNet U-Net outperformed standard U-Net models, achieving a Dice score of 76.84%, sensitivity of 78.62%, specificity of 99.52%, and accuracy of 99.08%. These findings suggest that the proposed method improves the detection of hard exudates compared to baseline models, providing a viable and computationally efficient approach for assisting in early diabetic retinopathy diagnosis.