Coronary artery segmentation with dual convolution-transformer U-Net for cardiovascular interventions

Computed Tomography Angiography (CTA) is widely used for diagnosis and treatment planning for Cardiovascular Disease (CVD). In image-guided Percutaneous Coronary Interventions (PCI), segmentation of the coronary artery from CTA is necessary for treatment planning and intraoperative image guidance, which could be time-consuming and error-prone if done manually. Many automatic segmentation algorithms have been proposed, but most of them used a relatively small amount of data and have not been validated for their generalizability. To help facilitate this, a dataset of 1000 CTA images (ImageCAS) has been created and made publicly available for robust automatic segmentation algorithm development and extensive validation. In this study, we employ a Dual Convolution-Transformer U-Net (DCT-U-Net) model that combines convolutional blocks with instance normalization, as is standard in U-Net models, and transformer blocks with multi-head self-attention, multilayer perceptron, and layer normalization. This additional transformer path allows the model to utilize information from all regions of the CTA scan, rather than local windows as is the case for convolution only models. DCT-U-Net was trained on the ImageCAS data, and its performance was compared to state-of-the-art models including nnU-Net. Compared to the ground truth manual segmentations, DCT-U-Net achieved a dice similarity score and Hausdorff distance of 0.859 and 16.019 mm, respectively, significantly outperforming existing models. DCT-U-Net is also computationally efficient with an average inference time of only 3.07 seconds. With excellent performance and fast computation time, DCT-U-Net has the potential to be used in clinical settings to enable accurate and efficient PCI treatment planning and interventions.