An Optimized Superpixel-Based Framework for Hard Exudates Segmentation Incorporating Multi-Stage Preprocessing

Abstract

Hard exudates (HEs) are among the earliest clinical manifestations of diabetic retinopathy (DR) and serve as important indicators for disease diagnosis and progression assessment. Accurate segmentation of HEs is therefore essential for early intervention and prevention of vision loss. However, automatic HEs detection remains challenging because lesion appearance, illumination conditions, and anatomically similar retinal structures such as the optic disc vary. This study proposes a knowledge-guided hard exudate detection framework based on a Superpixel-Based Framework (SBF) combined with adaptive feature analysis. The proposed method integrates retinal image enhancement, CIELAB color-space transformation, SBF superpixel generation, adaptive luminance–chromaticity thresholding, optic disc suppression, and morphological refinement to improve lesion localization and reduce false-positive detections. Unlike deep learning approaches that require large annotated datasets and substantial computational resources, the proposed framework leverages perceptually meaningful color information and retinal anatomical knowledge to achieve robust, interpretable segmentation. We evaluated the framework on the publicly available DIARETDB1 and STARE datasets and compared it with K-means Clustering, Fuzzy C-Means (FCM), Mean Shift Segmentation, Region Growing, U-Net, and Convolutional Neural Network (CNN) based segmentation methods. The proposed framework achieved the best performance on the Standard Diabetic Retinopathy Database (DIARETDB1), with 98.24% accuracy, 95.81% sensitivity, 99.12% specificity, a Dice coefficient of 95.58%, and an IoU of 91.42%. Similarly, on the STructured Analysis of the Retina (STARE) dataset, the method achieved 97.89% accuracy and 90.21% IoU. The results demonstrate that the proposed method provides accurate, computationally efficient, and interpretable HEs segmentation for automated DR screening systems.