Advancing retinal image analysis: from preprocessing to lesion identification in diabetic retinopathy

Abstract

Diabetic retinopathy (DR) is a leading cause of vision loss, making early detection and classification vital for effective management. Despite advancements in machine learning and deep learning techniques, these methods often require substantial computational resources and specialized domain knowledge, which are not always readily available. To address these challenges, a robust and computationally efficient methodology is proposed for early-stage DR detection, leveraging classical image enhancement and multi-stage segmentation techniques. Utilizing publicly available datasets— DRIVE, CHASEDB1, and DIARETDB1, our study applied data augmentation techniques to bolster the training set and mitigate overfitting. The framework employs a range of image enhancement methods, including Median Filter, Weiner Filter, and contrast limited adaptive histogram equalization (CLAHE), followed by lesion segmentation techniques. Accurate isolation of critical features like blood vessels (BVs) and the optic disc is achieved through principal component analysis (PCA) and iterative self-organizing data analysis technique algorithm (ISODATA). The experimental analysis demonstrates that the proposed methodology achieves high performance, with sensitivity, specificity, and accuracy rates of 91.50%, 88.20%, and 90.40% for microaneurysms (MA); 92.20%, 89.10%, and 91.30% for hemorrhages (HE); and 93.10%, 90.20%, and 92.60% for exudates, respectively. The image enhancement techniques improved the peak signal-to-noise ratio (PSNR) to 21.80 and the normalized cross-correlation (NCC) to 0.812. These results indicate the effectiveness of the proposed methods in accurately detecting and classifying lesions in retinal images.