BFR-Unet: A Full-Resolution Model for Efficient Segmentation of Tiny Blood Vessels

Abstract

Retinal blood vessel segmentation plays a crucial role in diagnosing retinal diseases, where accurate and complete vessel segmentation is essential for reliable diagnosis. Currently, U-Net remains one of the most widely used architectures for retinal blood vessel segmentation. However, due to the complexity and variability of retinal structures, the blood vessel edges are often very thin, and the low contrast of retinal images further complicates accurate segmentation. These challenges frequently result in U-Net models failing to precisely capture vessel boundaries. To address this issue, a novel full-resolution retinal blood vessel segmentation network, termed BFR-Net, is proposed. The BFR-Net is composed of three primary modules: the multi-residual convolution module, the boundary attention module, and the feature fusion module. The multi-residual convolution module, forming the backbone of the network, enables effective extraction of contextual information across the full resolution. The boundary attention module processes outputs from both the backbone and different network levels to capture detailed edge features, thus enhancing the segmentation performance. Finally, the feature fusion module integrates features from the backbone and boundary attention modules, further improving overall network performance. The performance of the proposed model is evaluated on three commonly used retinal vessel segmentation datasets. Experimental results demonstrate that BFR-Net achieves advanced performance, particularly in segmenting vessel edges and small blood vessels. Specifically, on the DRIVE and CHSAE_DB1 datasets, the Se and F1 scores are 0.8646, 0.8244, 0.8838, and 0.8108, respectively. These results demonstrate that the proposed network exhibits excellent performance in segmenting vessel boundaries and fine vessels.