A hybrid convolutional and transformer network for segmentation of coronary computed tomography angiographic slices

Abstract

AbstractCoronary artery three‐dimensional reconstruction is essential for preventing, diagnosing, and treating coronary heart disease. This study introduces SegUnet, a lightweight hybrid CNN‐Transformer network for pixel‐level segmentation of coronary artery computed tomography angiography (CTA) slices to enhance the precision of coronary artery reconstruction. The overall SegUnet adopts a U‐shaped encoder‐decoder structure, with a hierarchical Transformer structure serving as the encoder to extract global contextual information; the decoder is enhanced with spatial attention mechanism (SAM) and residual structure to improve its perception and adaptive adjustment abilities for coronary artery vessels. In contrast to other methods, we propose an interpolation‐based multi‐feature fusion bridge to integrate multi‐scale features between encoder levels, capturing their semantic dependencies. Experimental results show that SegUnet achieves a prediction accuracy of 97.23% on the test set of coronary arteries 2D slices. Moreover, the network exhibits excellent performance on Synapse Multi‐Organ Segmentation, highlighting the superiority, effectiveness, and robustness of our proposed method. The outstanding performance of SegUnet in coronary artery segmentation has elevated the quality of three‐dimensional reconstruction of the coronary arteries, exerting a positive impact on the enhancement of diagnosis and treatment of coronary artery diseases.