Non-contrast Cerebral Blood Volume Images Synthesis from Arterial Spin Labeling and Non-contrast Standard MRI Using Deep Learning Network

Motivation: High quality brain perfusion maps without contrast agent has potential values in in clinical scenarios. Goal(s): To explore the possibility of deep learning model for synthesizing non-contrast cerebral blood volume (CBV) maps from arterial spin labeling (ASL) and non-contrast standard MRI sequences. Approach: Quantitative CBV maps of all participants were acquired by using the DSC-PWI sequence and synthesized by training a 3D incrementable encoder-decoder network on small sample size. Results: Deep learning model produced high-quality non-contrast CBV maps and the synthetic non-contrast CBV maps have better performance in glioma grading, prognosis prediction and differential diagnosis between tumor recurrence and treatment response than ASL. Impact: Patients undergoing radiochemotherapy with fragile vessels or adverse reactions to gadolinium contrast could benefit from synthetic non-contrast CBV methods.