Hiearachical Construction of Pulmonary Vascular Trees from CT Images Based on Combined Deep Learning and Graph Topology Optimization
Bobo Qin, Hongkai Wang · 2023
Automatic construction of pulmonary vascular trees (PVTs) from computer tomography (CT) images is a critical step for computer-assisted diagnosis and treatment planning of pulmonary diseases. However, robust and accurate construction of the PVTs is a challenging problem due to the difficulties in artery/vein (A/V) classification and touching vessels separation. Current studies use deep learning approaches for A/V classification but lack of the utilization of vessel tree topology knowledge. In this paper, we combine deep learning approach with graph topology optimization for more accurate PVT construction. We first use a 3-D U-Net model to achieve the initial A/V segmentation and obtain the voxel-level probability maps. Based on the initial segmentation, the initial A/V tree-graph is constructed, and the centerlines and radius of each vessel segment are calculated. Using the initial voxel probabilities and the geometric features of the vessel segments (e.g., center line direction, length and radius), a graph optimization step is performed to correct the inaccurate A/V classifications and break the erroneous connections between the touching vessels. Finally, a more accurate hierarchical tree graph is constructed. Experimental results illustrate that the combination of deep learning and graph optimization outperformed the state-of-the-arts (SOTA) methods which only uses deep learning. Our method also simultaneously generates a parametric model of the vascular tree, which is important for vessel morphometry measure, blood flow simulation and etc.