HawkNet: an intelligent bio-inspired optimization based patch wise adaptive U-Net framework for retinal blood vessel segmentation in fundus images

Abstract

The accurate segmentation of retinal blood vessels plays a pivotal role in the early diagnosis and monitoring of health issues like diabetes, high blood pressure, and glaucoma. Traditional machine learning techniques such as matched filtering, morphological processing, and edge detection have a tough time to see tiny blood vessels clearly because eye images often have uneven lighting, poor contrast, and blood vessels that twist and turn in complex ways. To overcome these challenges, this work introduces an improved Patch-wise U-Net method with Harris Hawk optimization (HHO). The patch system split fundus images into overlapping patches, enabling the network to focus on localized vessel features such as fine capillaries, bifurcations, and vessel boundaries that are often missed in global segmentation. Meanwhile, HHO is employed to fine-tune the U-Net’s hyperparameters and learning weights, achieving faster convergence and enhanced segmentation accuracy without manual tuning. To establish generalizability and reliability, the proposed framework is rigorously tested on two standard retinal image repositories DRIVE, and STARE. Experimental evaluation demonstrates significant improvement in key performance metrics, including accuracy of 97.8%, precision 95.8%, recall 96.5%, f1-score 97.3%, IoU of 96.2%, and dice coefficient (DC) 94.3%, highlighting the model’s capability to capture thin and thick vessels structures while minimizing false detections. Additionally, qualitative segmentation further confirm that the proposed framework effectively preserves the continuity and morphology of thin and thick vessels, even in regions with low contrast or uneven illumination. Overall, the proposed method visual inspection has revealed that the suggested method can segment thin and thick vessels with greater accuracy than previous methods. It also demonstrates its potential for real-life clinical application.