Towards Segmenting Cerebral Arteries from Structural MRI

The assessment of cerebral arteries and the Circle of Willis (CoW) structure in neuroimaging scans is crucial for diagnosing various cerebrovascular pathologies. While angiographic sequences such as time-of-flight magnetic resonance angiography (TOF-MRA) are indispensable tools for the assessment of cerebral arteries, extending segmentation methods to include structural MRI holds significant clinical promise. In this study, we introduce a novel methodology to address the task of segmenting cerebral arteries in structural sequences. Our main goal is to construct a large database of paired structural sequences with corresponding pseudo-labels. First, we train a segmentation model on a small subset of angiographic data with gold-standard ground-truth labels and then utilize this model to predict pseudo-labels for the remaining unlabeled portion of the database. As subjects in the database comprise both angiographic and structural sequences, we register structural sequences to the space of predicted pseudo-labels, constructing our paired dataset. Finally, we train a sequence-agnostic segmentation model on this dataset with pseudo-labels and conduct extensive evaluations, reporting results for both the full and CoW regions. Our results show that we can learn to segment cerebral arteries from structural MRI, where PD-weighted sequences had the highest scores for both regions. We share our trained models on GitHub to facilitate further experimentation and research (github.com/risc-mi/cerebral-artery-segmentation).