High Precision Deep learning 4D MRA Vessel Segmentation: Technical Development and Initial Clinical Evaluation on Arteriovenous Malformation

Motivation: Vessel segmentation from 4DMRA may provide further information to aid in clinical diagnosis. However, currently most of the neural networks for MRA segmentation target static angiography. Goal(s): To design a generalized neural network for 4DMRA intracranial vessel segmentation with minimal preprocessing. Approach: A modified U-net architecture (4DST U-Net) was designed by leveraging both spatial [x,y,z] and temporal (t) dimensions for 4D MRA vessel segmentation. External validation on two AVMs and a healthy volunteer was tested for model generalizability. Results: The proposed deep learning vessel segmentation method outperformed the other three models. External validation with AVM data correctly detected the AVM lesions. Impact: This work developed a 4DST U-Net for 4D MRA vessel segmentation with minimal preprocessing. The generalizability of this neural network was demonstrated by the external validation on patients. Both features may facilitate a wider application of this technique across multi-sites.