Cerebral Artery Classification: Integrating Synthetic Models with MRA Data

The Circle of Willis (CoW) plays a critical role in cerebral vascular circulation, yet its high anatomical variability complicates accurate segmentation and artery classification, particularly in clinical settings. While Magnetic Resonance Angiography (MRA) images are commonly used for segmentation, the scarcity of comprehensive, annotated datasets limits the performance of machine learning models. In this study, we leverage synthetic models generated by VaMos, a vascular modeling tool, to augment the existing TopCoW dataset. By incorporating these anatomically plausible synthetic structures alongside actual MRA data, we improve artery classification using a deep learning model, specifically nnUNet, which is optimized for medical image segmentation. For smaller/thinner and underrepresented arteries, such as the posterior and anterior communicating arteries, the improvement was around 16%. This approach provides a novel solution to data scarcity, showing that synthetic models can effectively complement real data to improve segmentation performance in challenging anatomical regions.