Ori-Net: Orientation-guided Neural Network for Automated Coronary Arteries Segmentation

Abstract

Coronary artery disease (CAD) is one of the diseases with high mortality, and its diagnosis is often facilitated by coronary artery segmentation in coronary computed tomography angiography (CCTA) images. However, due to the low contrast, imaging artifacts, and contrast agents in CCTA images, coronary artery segmentation models often produce fragmented segmentations due to the lack of coronary artery topology knowledge. We found that coronary arteries have cylinder-like structures that can be represented as envelopes reconstructed by orientation-guided spheres of different radii. To fully use this prior knowledge, a novel orientation-guided neural networks model, called Ori-Net, is proposed to improve the connectivity of coronary artery segmentation in CCTA images. First, Ori-Net simultaneously produces the coarse segmentation, radius, and orientation of the coronary artery in a multi-task learning framework. Then, we propose an orientation-guide tracking method using the predicted orientation and radius. It reconstructs the coronary artery iteratively, and the reconstruction is fused with the coarse segmentation to improve the segmentation performance further. The proposed method provides a new way to leverage the coronary artery shape prior. In addition, we extend the connected component ratio metric for volume data to evaluate the connectivity of the segmentation results. In the experiments, we compared the proposed Ori-Net with state-of-the-art methods on two coronary artery segmentation datasets. It demonstrates that (1) Ori-Net improves over the state-of-art by 10% in connectivity metric. (2) Ori-Net can significantly boost coronary artery segmentation performance, for example, by over 5% improvement as measured by Dice.