Automated segmentation of blood-flow regions in large thoracic arteries using 3D-cine PC-MRI measurements

Purpose: Quantitative analysis of vascular blood-flow, acquired by phase-contrast MRI requires an accurate segmentation of the vessel lumen. In clinical practice, 2D-cine velocity-encoded slices are inspected, and mostly the blood-flow lumen is segmented manually. However, segmentation of time-resolved volumetric blood-flow measurements is a tedious and time-consuming task. Methods: We propose an automated segmentation of large thoracic arteries, solely based on the 3D-cine phase-contrast MRI (PC-MRI) blood-flow data. In contrast to previous work, we employ an active-surface model, which is fast and topologically stable. An active surface model requires an initial surface, approximating the desired segmentation. We introduce a novel method to generate this surface, based on a voxel-wise temporal maximum of the blood-flow speed intensities. The active surface model balances forces, based on the surface structure and image features derived from the blood-flow data. The segmentation results were validated using volunteer studies, including time-resolved 3D and 2D blood-flow data. The segmented surface was intersected with a velocity-encoded PC-MRI slice, resulting in a cross-sectional contour of the lumen. These cross-sections were compared to reference contours that were manually delineated on high-resolution 2D-cine slices. Results: We show that our automated approach closely approximates the manual blood-flow segmentations, with error distances on the order of the voxel size. The initial surface provides a close approximation of the desired luminal geometry. This improves the convergence time of the active surface, and facilitates parametrization. Conclusions: We present an active-surface approach to segment the vessel lumen, suitable for quantitative analysis of 3D-cine PC-MRI blood-flow data. As opposed to the variety of thresholding and level-set approaches, the active-surface model is topologically stable. We introduce a novel methodology to generate an initial approximate surface, and have inspected various features that underpin the segmentation model. We have validated the feasibility of our approach, and show that the resulting surfaces closely approximate manual segmentations.