Abstract WP16: Automated Vessel Identification In The Circle Of Willis Using Deep Learning Models To Assist AI Based Lesion Detection In Cerebral Vasculature

Introduction: Vascular diseases are among the top causes of death in the world. Automated approaches which allow comparison of vasculature enable both large population studies of vasculature and assist objective diagnosis of vascular disease. Accurate and robust methods for vessel labeling of angiography are an important step to automatically localize lesions or other features of interest. Hypothesis: using a deep learning approach, a model can be developed to identify cerebral arteries. Methods: We used 152 cerebral TOF-MRA angiograms from three publicly available datasets and automatically segmented vessels using a pre-trained nnU-net model. With the segmented vessels, we manually created reference labels for different cerebral arteries using 3DSlicer. After extracting vessel centerlines from the segmentations using VesselVio, we then associated each centerline point with the reference labels to assign individual arterial segments. Finally, in an ablation-type study, several PointNet++ point cloud models were trained and evaluated, using the centerline coordinates, local vessel radius and/or centerline point connectivity information as inputs. Results: The model trained only on centerline coordinates resulted in a high true positive rate (TPR) of 0.92, while models trained with additional features resulted in significantly higher (p<0.05) TPR of 0.95. Different PointNet++ models favored improved labeling of certain arteries, while the internal carotid artery, and posterior communicating artery were consistently the most and least accurately labeled arteries, respectively. Conclusion: We developed a complete pipeline utilizing a deep learning PointNet++ for automatic cerebral vessel identification that labels vessels based on automatically extracted centerlines. Results show that utilizing vessel radius as an additional feature significantly improved cerebral vasculature labeling. The overall TPR of 0.95 shows that our model can be used for fast and reliable artery labeling of cerebral MRA images.