Coronary Centerline Extraction via Optimal Flow Paths and CNN Path Pruning

We present a novel method for the automated extraction of blood vessel centerlines. There are two major contributions. First, in order to avoid the shortcuts that minimal path methods are prone to we find optimal paths in a computed flow field. We solve for a steady state porous media flow inside a region of interest and trace centerlines as maximum flow paths. We explain how to estimate anisotropic orientation tensors which are used as permeability tensors in our flow field computation. Second, we introduce a convolutional neural network (CNN) classifier for removing extraneous paths in the detected centerlines. We apply our method to the extraction of coronary artery centerlines found in Computed Tomography Angiography (CTA). The robustness and stability of our method are enhanced by using a model-based detection of coronary specific territories and main branches to constrain the search space[15]. Validation against 20 comprehensively annotated datasets had a sensitivity and specificity at or above 90%. Validation against 106 clinically annotated coronary arteries showed a sensitivity above 97%. 1 Motivation and Overview The automatic segmentation of coronary arteries in Computed Tomography Angiography (CTA) facilitates the diagnosis, treatment and monitoring of coronary artery diseases. An important step in coronary segmentation is to extract a centerline representation that supports visualizations such as curved planar reformatting or that supports lumen segmentation methods for quantitative assessments such as stenosis grading. In this work, our focus is to detect the full coronary tree of centerlines including the distal part of side branches for better visualization and quantification of the coronary anatomy. Coronary arteries constitute only a small portion of a large CTA volume because of their thin tubular geometry. Their centerline extraction is not an easy task due to nearby heart structures or coronary veins. The majority of existing centerline extraction techniques compute centerline paths by minimizing a vesselness or medialness cost metric such as Hessian based vesselness [6], flux based medialness [2, 8] or other tubularity measures [9] along the paths. However, the cumulative cost nature of these methods makes them very sensitive to the underlying cost metric causing them to easily make shortcuts through nearby non-coronary structures, especially when there are pathologies, large contrast variations, bifurcations or imaging artifacts along the true path. In addition, these methods are sensitive to length and curvature of the vessel. Vessel centerlines have frequently been found through minimum paths, minimum spanning trees or tracking algorithms such as Kalman filtering. These are generally run on hand-crafted vesselness measures computed from the image data. A review of 3D vessel segmentation can be found in [9]. More recent work evaluates consistency among multiple minimum paths [11] or considers multiple hypothesis tracking [7]. Constraints on possible paths have been applied through constrained optimization including a term for the conservation of network flow in [13] or through prior shape models [15]. The main contributions of this paper are 1) a formulation of centerline extraction as finding the maximum flow paths in a steady state porous media flow, 2) a learning based estimation of anisotropic vessel orientation tensors and their use as permeability for a flow computation, 3) a CNN based branch classifier for distinguishing true centerlines from leakages, and 4) the use of model-based coronary specific territories and main branches for robustness and efficiency.