The Effect of Contrast Enhancement on Retinal Blood Vessel Segmentation Using CAS-UNet with Coordinate Attention

Abstract

Low contrast variation, uneven intensity distribution, and the presence of noise in retinal fundus images pose major challenges for blood vessel segmentation, particularly regarding thin and complex structures. These conditions make it difficult for models to accurately distinguish between blood vessels and the background. This study aims to analyze the impact of contrast enhancement techniques on retinal blood vessel segmentation performance using a CAS-UNet architecture modified with Coordinate Attention (CA). The methodology involves three preprocessing scenarios: Grayscale, Grayscale + CLAHE, and Grayscale + CLAHE + Gamma Correction. The model was trained using the DRIVE and CHASE_DB1 datasets with an 80:20 data split, an SGD optimizer, a learning rate of 0.01, and a combined BCE and Dice loss function over 50 epochs. Evaluation was conducted using a confusion matrix based on accuracy, sensitivity, specificity, F1-score, and IoU metrics. The results indicate that the Grayscale + CLAHE combination yielded the best performance—achieving a sensitivity of 81.46%, an F1-score of 81.63%, and an IoU of 69.01%—while also improving the detection of small blood vessels more consistently. These findings demonstrate that the appropriate application of contrast enhancement plays a crucial role in improving the quality of medical image segmentation.