A Deep-Learning-Based Framework for Automatic Segmentation And Labelling of Intracranial Artery

Automatic segmentation and labelling of intracranial arteries is important for the clinical diagnosis and research of cerebrovascular disease, but inter-individual differences in intracranial artery structure pose a serious challenge to automatic processing pipeline. Existing approaches model the artery labelling task as a centre-line classification problem, neglecting the significance of the image-level vessel segmentation and labelling for clinical research. In this paper, we propose a deep learning-based automated processing pipeline for joint segmentation and labelling of intracranial arteries, and further again a centre-line vessel type prediction algorithm based on voting strategy that is capable of obtaining both image-level and centre-line-level vessel labelling results. We used a private dataset containing 167 individual MRA scans and the public dataset TubeTK for training and testing. The experimental results show that our approach achieves a labelling dice score of 88.3% for 21 intracranial arteries and an average centre-line prediction accuracy of 95.5%, showing stable and robust results.