Automated Generation of Cerebral Blood Flow Maps Using Deep Learning and Multiple Delay Arterial Spin-Labelled MRI
The purpose of this work was to develop and evaluate a deep learning approach for estimation of cerebral blood flow (CBF) and arterial transit time (ATT) from multiple post-label delay (PLD) arterial spin-labelled (ASL) MRI. Six-PLD ASL MRI was acquired on a 1.5T or 3T system among 99 older males and females with and without cognitive impairment. We trained and compared two network architectures: standard feed-forward convolutional neural network (CNN) and U-Net. Mean absolute error (MAE) was evaluated between model estimates and ground truth obtained through conventional processing. The best-performing model was re-trained on inputs with missing PLDs to investigate generalizability to different PLD schedules. Relative to the CNN, the U-Net yielded lower MAE on training data. On test data, the U-Net MAE was 8.4{+/-}1.4 ml/100g/min for CBF and 0.22{+/-}0.09 s for ATT. Model uncertainty, estimated with Monte Carlo dropout, was associated with model error. Network estimates remained stable when tested on inputs with up to three missing PLD images. Mean processing times were: U-Net pipeline = 10.77s; ground truth pipeline = 10min 41s. These results suggest hemodynamic parameter estimation from 1.5T and 3T multi-PLD ASL MRI is feasible and fast with a deep learning image-generation approach.