Discriminative Coronary Artery Tracking via 3D CNN in Cardiac CT Angiography

Extraction of the coronary artery centerline from cardiac CT angiography (CCTA) is a challenging yet prerequisite task for subsequent diagnosis in clinical practice. In this paper, a discriminative coronary artery tracking method (DCAT) is proposed to (semi) automatically extract coronary artery centerlines. It consists of two parts: a tracker and a discriminator. The tracker outputs orientation and radius of the vessel at each location, which is used to extract vessel-like objects. The discriminator provides a learning-based stop criterion during tracking, which can distinguish coronary artery from other vessel-like objects. We train the tracker and the discriminator simultaneously, which are proved to be helpful to each other. We evaluate the DCAT on the public dataset in CAT08 challenge and our method outperforms state-of-the-art methods. Furthermore, training of the discriminator only needs coarsely labeled centerline annotations, that enables training the DCAT model on a large set of data all of which have centerline annotations, but only a small fraction of which have accurate centerline and radius annotations. This reduces annotations effort greatly. Experimental results on a private collected clinical dataset demonstrate the effectiveness of this training schema.