Retinal Blood Vessel Segmentation Using Attention V-Net with Scale-Aware Evaluation

Abstract

Retinal blood vessel segmentation remains a significant challenge, especially for small blood vessels with diameters less than 3 pixels in the DRIVE dataset and less than 4 pixels in the STARE dataset, owing to their low contrast and narrow structures. The aim of this study is to improve small retinal blood vessel segmentation performance through an Attention V-Net architecture that integrates attention-gating mechanisms into the skip connections of a V-Net backbone to strengthen the feature representation of thin vascular structures. The research method involves training and evaluating the proposed model on the DRIVE and STARE datasets using a scale-aware evaluation framework based on pixel-pitch calibration, classifying blood vessels into small and large categories, and measuring performance using accuracy, sensitivity, specificity, precision, Dice coefficient, and IoU. The results show that for small vessel segmentation, the method achieves sensitivities of 0.7033 and 0.6984, Dice scores of 0.4720 and 0.4699, and IoUs of 0.3096 and 0.3079 on the DRIVE and STARE datasets, respectively. For large vessels, sensitivities of 0.9219 and 0.8851, Dice scores of 0.8031 and 0.8179, and IoUs of 0.6719 and 0.6933 are obtained. Global evaluation yields accuracies of 0.9475 and 0.9602, sensitivities of 0.8727 and 0.8719, and Dice scores of 0.8080 and 0.8268. In conclusion, Attention V-Net demonstrates consistent segmentation performance across vessel scales, and the scale-aware evaluation framework effectively reveals the performance gap between small and large vessel segmentation, providing a more clinically relevant assessment than conventional global evaluation for early diagnosis of retinal diseases.