Deep learning segmentation of small blood vessels and vessel density mapping based on high resolution black blood MRI

Motivation: Cerebral small vessel disease (cSVD) is linked to cognitive decline in older populations. Robust imaging of the small vessels can serve as a biomarker for cognitive decline. Goal(s): We aim to build a cerebral small vessel segmentation model robust to motion and image quality based on high resolution black blood MRI. Approach: We train a nnUNet model using in-house high quality data with simulated motion artifacts and evaluated trained model in a cohort of elderly persons at risk of cSVD. Results: The trained model showed higher performance than Jerman filter, and vessel density in the hippocampus was correlated with cognitive scores. Impact: A robust cerebral small vessel density estimation method can help for large-scale analysis of black blood MRI data to look for possible biomarkers of early cognitive decline. Our pipeline will be shared with the community.