SFIT-Net: Spatial Reconstruction Feature Interaction Transformer Retinal Vessel Segmentation Algorithm

Retinal vessel segmentation is crucial in the early diagnosis and treatment of fundus disorders, as it enables the delineation of vessels with varying thicknesses and aids in detecting disease symptoms. However, existing segmentation methods often overlook the local morphological structures and global dependencies of blood vessels, resulting in difficulties such as distinguishing blood vessels from background noise and accurately identifying small-caliber vessels. To address these challenges, we developed the Spatial Reconstruction Feature Interaction Transformer Retinal Vessel Segmentation Algorithm (SFIT-Net), which integrates convolutional neural networks (CNNs) for extracting local vessel features and Transformer models for efficiently capturing global relationships. This allows for better capturing of long-range dependencies between vessels and alleviating vessel discontinuities. Specifically, SFIT-Net first introduces the Spatial Reconstruction Unit (SRU) in the encoding section. This model employs weights to guide the network in focusing more on vascular information and other relevant features. The Spatial Recursive Unit (SRU) performs cross-reconstruction to mitigate semantic loss caused by downsampling and convolution. Then, the Feature Interaction Transformer (FIT) module in the bottom layer extracts distinctive vascular features and minimizes background interference. Finally, the TFA module fuses feature maps from different paths at the same scale in decoding, enhancing both high- and low-level representations and improving vessel information utilization. The SFIT-Net algorithm underwent extensive experiments on the DRIVE, CHASE-DB1, and STARE datasets. The results demonstrate that our method outperforms most methods in terms of accuracy, sensitivity, F1-score, and generalization ability. © 2025 Elsevier Ltd