A method for automatic 3D vasculature segmentation in ex vivo MRI using synthetic data

Motivation: Cerebral vascular anatomy is involved in several diseases, but only the largest vessels can be routinely imaged in vivo. Ex vivo MRI offers an alternative at higher resolution (~100 μm) where more vessels are visible. However, it has poor signal-to-noise ratio and nonspecific contrast, and obtaining accurate vessels manual annotations is extremely difficult. Goal(s): Our goal was to develop a method for automatic 3D vessel segmentation in ex vivo MRI. Approach: To overcome the lack of vessels manual annotations, we trained a neural network on synthetic data. Results: The segmentation method tested on real ex vivo MRI images achieved human-level performance. Impact: Our method for 3D vessel segmentation in ex vivo MRI can be used to build a whole-brain vascular atlas, and study inter-subject variability. It can also be adapted to microscopy and neuropathology, and to other tubular structures (axons and fascicles).