Spatial-AttVesNet: Compact U-Net with Attention for Retinal Vascular Image Segmentation

Ophthalmologists have long focused on studying the retinal vasculature to help diagnose disorders such as Diabetic Retinopathy (DR), Macular Degeneration (AMD), Retinal Vein or Artery Occlusion, among others.Segmenting the structure of retinal vessels from fundus images remains a challenging task due to the large variability of structures and the non-homogeneity of their intensities.This study introduces a two-step framework to improve the segmentation of retinal blood vessels.First, we establish a preprocessing phase to simulate the most common distortions encountered in fundus images, including orientation variations, photometric distortions, and color inconsistencies.A second phase focuses on segmentation using an enriched U-Net architecture to simplify the fusion of encoded and decoded features.Experimental results demonstrate that our proposed method Spatial-AttVesNet achieves superior performance compared to traditional U-Net, particularly in terms of precision 97.23%, Specificity 99.09%, sensitivity 78.08%, Dice coefficient 82.06% and IOU 69.79%.The proposed method is evaluated on publicly available tree datasets as baselines (DRIVE, STARE, and Chase-DB1), respectively, achieving an accuracy of 97.23% ,96.56%, and 97.74% which is 0.12%, 0.55% and 1.56% higher than that of the U-Net baseline.Finally, we replicate our method on our proprietary dataset MDRID 2025 (Moroccan Diabetic Retinopathy Image Dataset 2025) composed of 500 retinal images manually annotated and classified into 5 classes according to the recommendations of the ALFEDIAM (French Language Association for the Study of Diabetes and Diseases) that address the scarcity of publicly available segmented images to train deep learning models.Our proposed method outperforms existing methods and opens new avenues of research.Our code and evaluation methods are available on Github at: https://github.com/SanaeLamti/Spatial-AttVesNet-A-Low