Heuristic Tubular Growth with Adaptive Voxel Filtering for Coronary Artery Segmentation

In 3D coronary computed tomography angiography (CCTA), accurate coronary artery segmentation is crucial for diagnosis and further treatment of cardiovascular diseases. However, for 3D CCTA images, the extremely unbalanced voxel distribution between coronary artery and background voxels brings a huge challenge to data-driven model optimization. To address this problem, we propose a novel coronary artery segmentation framework which consists of two stages: growth and filtering. Specifically, in the growth stage, a local growth algorithm is proposed to segment and connect the targets patch wisely, ignoring the non-target patches. In the filter stage, voxel-level pseudo heart labels are generated from coronary artery ground truths to train a heart segmentation model to exclude irrelevant background voxels. Finally, a point cloud based segmentation model is introduced to remove the false positive voxels within the heart (e.g., veins) and get more accurate segmentation results. Compared with existing methods, our framework could avoid interference from redundant voxels and is more robust to unbalanced class distribution. Extensive experiments demonstrate that the proposed model outperforms the state-of-the-art both in quality and quantity (i.e., 6.1 % improvement). Additionally, we will release our full codes and a high-quality annotated coronary artery segmentation dataset (i.e., CORONARY-18) for future research.