Spatio-Temporal correspondence attention network for vessel segmentation in X-ray coronary angiography

Background: The segmentation of contrast-filled vessels from X-ray coronary angiography (XCA) image sequences is an important step in the diagnosis and treatment of coronary artery disease. However, accurate and complete extraction of blood vessels is particularly challenging due to the poor quality of XCA images and the complex structure of blood vessels. The existing coronary artery segmentation methods mainly focus on pixel segmentation while ignoring the temporal information of the angiographic sequence, which further aggravates the fragmentation and loss of segmentation results. Methods: We propose a spatio-temporal correspondence attention network, which is a novel encoder-decoder structure consisting of one spatio-temporal correspondence block and two attention blocks. The spatio-temporal correspondence block establishes the relationship between the segmented frame and the previous frame, extracting spatio-temporal features from the previous frame to enhance feature representation of the current segmented frame. Simultaneously, the spatial attention block and channel attention block establish spatial and channel dependence relationship of the current segmented frame, enhancing foreground vessel segmentation ability. In addition, in order to solve the category imbalance problem in XCA images, we combined CE loss and Dice loss functions to train the proposed deep learning network. Results: The dice obtained by our method is 89.18 ± 0.07 %, the precision is 90.65 ± 0.32 %, and the recall is 87.76 ± 0.37 %. Conclusion: Compared with other advanced methods, our method demonstrates excellent results in both qualitative and quantitative measures. This suggests that integrating spatio-temporal information into the deep learning framework can enhance the segmentation of blood vessels in coronary angiography images.