UKnow-Net: Knowledge-Enhanced U-Net for Improved Retinal Vessel Segmentation

Abstract

Retinal vessel segmentation plays a critical role in diagnosing and managing ophthalmic and systemic diseases, as abnormalities in retinal vasculature can indicate disease progression. Traditional manual segmentation by expert ophthalmologists is time-consuming, labor-intensive, and prone to variability, underscoring the need for automated methods. While deep learning approaches like U-Net have advanced retinal vessel segmentation, they often struggle to generalize across diverse datasets due to differences in image acquisition techniques, resolutions, and patient demographics. To address these challenges, I propose UKnow-Net, a knowledge-enhanced U-Net architecture designed to improve retinal vessel segmentation across multiple datasets. UKnow-Net employs a multi-step process involving knowledge distillation and enhancement techniques. First, I train four specialized teacher networks separately on four publicly available retinal vessel segmentation datasets—DRIVE, CHASE_DB1, DCA1, and CHUAC—allowing each to specialize in the unique features of its respective dataset. These teacher networks generate pseudo-labels representing their domain-specific knowledge. We then train a student network using the ensemble of pseudo-labels from all teacher networks, effectively distilling the collective expertise into a unified model capable of generalizing across different datasets. Experiments demonstrate that UKnow-Net outperforms traditional handcrafted networks (such as U-Net, UNet++, and Attention U-Net) and several state-of-the-art models in key performance metrics, including sensitivity, specificity, F1 score, and Intersection over Union (IoU). Specifically, our two variants, UKnowNet-A and UKnowNet-B, show well performance; UKnowNet-A, trained solely on pseudo-labels, achieved higher sensitivity across all datasets, indicating a superior ability to detect true positives, while UKnowNet-B, which combines pseudo-labels with ground truth annotations, achieved balanced precision and recall, leading to higher F1 scores and IoU metrics. The integration of pseudo-labels effectively transfers the collective expertise of the teacher networks to the student network, enhancing generalization and robustness. I aim to ensure fair comparison and reproducibility in future research by publicly sharing our source code and model weights.