Abstract
Diabetic retinopathy (DR) is a common eye condition that affects one-third of patients with diabetes, leading to vision loss in both working-age and elderly populations. Early detection and intervention can improve patient outcomes and reduce the burden on healthcare. By developing robust computational techniques, we can advance automated systems for screening and managing diabetic retinopathy. Our specific goal is to detect and segment exudates and hemorrhages in fundus images. In this study, we used the iterative NICK thresholding region growing (INRG) method as a basis. To further improve our results in different applications, we incorporated the watershed separation algorithm (WS) and the Chi(2) feature selection method (Chi(2)) on expanded feature sets. These algorithms were combined with the INRG method to segment hemorrhages and exudates. The segmentation results were used to detect the hemorrhages and exudates, which in turn were used to detect diabetic retinopathy. To evaluate our approach, we compared the results against two traditional methods and two state-of-the-art methods, including the original INRG-HSV model. In terms of hemorrhage segmentation, the INRG with WS (INRG-WS) achieved the highest F-measure of 64.76%, outperforming all other comparative methods. For exudate segmentation, the model INRG-WS- Chi(2), which used the combined INRG method with WS and Chi(2) ranking on expanded feature sets, performed the best. When it came to hemorrhage detection, the INRG method without WS and using only hue, saturation, and brightness (INRG-HSV) achieved the highest accuracy of 90.27% with the lowest false negative rate (FNR) of 9.39%. For exudate detection, the model INRG-WS-HSV, which used the combined INRG method with WS and only hue, saturation, and brightness, offered the highest accuracy rate of 88.14% and the lowest FNR rate of 8.75%. To detect diabetic retinopathy, we compared the performance of our best hemorrhage detection model (INRG-HSV) and exudate detection model (INRG-WS-HSV) against a state-of-the-art method. Our models significantly outperformed the state-of-the-art method (DT-HSVE), achieving an accuracy of 89.89% and an FNR of 3.66%.