Probabilistic labeling of cerebral vasculature on MR angiography

Abstract

Automatic labeling of cerebrovascular territories would greatly advance our ability to systematically study large datasets and also provide rapid decision support during the assessment of stroke patients, for example. Previous attempts have been challenged by the wide inter-subject variation in vascular topography. We investigate the use of a probabilistic model that learns the configurational characteristics of vascular territories to better annotate the cerebrovasculature. In the George Mason Brain Vasculature database, we identified patients with MRA reconstructions segmented into seven major regions (left and right MCA, PCA, and ACA and Circle of Willis). We then augmented these labels by manually segmenting the MCA territory into an additional eight regions. Among 54 patients that met the inclusion criteria, 39 reconstructions were used as training input to the MCA, ACA, and PCA model among the 61 digital reconstructions of human brain arterial structures available. The model was then validated on an independent cohort of 15 patients. The MCA segmentation algorithm was trained and tested using leave-one-out crossvalidation. The algorithm was found to be 94 ± 5.2% accurate in annotating the seven major regions and 88 ± 9.3% accurate in annotating the MCA subterritories.