An Attention Guided Multi-scale Network with Channel-Enhanced Transformer for Coronary Arteries Segmentation

Coronary artery disease is a major cause of mortality and morbidity. Automatic segmentation of the coronary artery is a key step in the diagnosis of coronary artery disease. This study aims to propose an attention guided multi-scale fusion network with channel-enhanced Transformer for automatic segmentation of coronary arteries on coronary computed tomography angiography (CCTA). To improve the segmentation performance, an attention guided multi-scale fusion (AGMF) module and channel-enhanced transformer (C-Trans) module are proposed. The AGMF can effectively extract and fuse multi-layer features to locate and segment multiple small coronary arteries accurately. Specifically, the AGMF module not only can extract representative coronary artery features in the encoding process but also can accurately locate and segment coronary arteries in the decoding process. To select the effective channel features, the C-Trans is introduced to obtain the channel features with token. This module can effectively screen out the channel features in each layer of U-net to improve the accuracy of segmentation. Experimental results show that our proposed method obtains the IoU of 0.6996 and the Dice of 0.8224, which are more than 1% higher than the existing model. From the results, it is found that our proposed network outperforms the state-of-the-art 2D coronary artery segmentation method.