DiffCAS: diffusion based multi-attention network for segmentation of 3D coronary artery from CT angiography

Automatic segmentation of 3D coronary arteries from computed tomography angiography (CTA) is an indispensable part of accurate and efficient coronary artery disease (CAD) diagnosis. However, it remains challenging due to the complex anatomy of coronary arteries. Inspired by the denoising diffusion probabilistic model (DDPM), we propose a diffusion-based multi-attention network for 3D coronary artery segmentation from CTA. The proposed method is called DiffCAS in short. DiffCAS utilizes the denoising diffusion of the diffusion model to yield segmentation results. During the denoising diffusion, the Swin Transformer is adopted to extract semantic information from CTA images, and an adaptive residual feature enhancement (ARFE) module is proposed as denoising encoder in the diffusion model, a feature fusion attention (FFA) module is coined to fuse the features from Swin Transformer and ARFE encoders, so as to improve the segmentation performance. Experimental results and comparisons on the ASOCA and ImageCAS datasets show that the proposed DiffCAS outperforms some SOTA networks in terms of Dice coefficient that are 84.41% and 84.59%, on ASOCA dataset and ImageCAS dataset, respectively.