Parametric cerebral blood flow and arterial transit time mapping using a 3D convolutional neural network.

Purpose: To reduce the total scan time of multiple postlabeling delay (multi-PLD) pseudo-continuous arterial spin labeling (pCASL) by developing a hierarchically structured 3D convolutional neural network (H-CNN) that estimates the arterial transit time (ATT) and cerebral blow flow (CBF) maps from the reduced number of PLDs as well as averages.

Methods: A total of 48 subjects (38 females and 10 males), aged 56-80 years, compromising a training group (n = 45) and a validation group (n = 3) underwent MRI including multi-PLD pCASL. We proposed an H-CNN to estimate the ATT and CBF maps using a reduced number of PLDs and a separately reduced number of averages. The proposed method was compared with a conventional nonlinear model fitting method using the mean absolute error (MAE).

Results: The H-CNN provided the MAEs of 32.69 ms for ATT and 3.32 mL/100 g/min for CBF estimations using a full data set that contains six PLDs and six averages in the 3 test subjects. The H-CNN also showed that the smaller number of PLDs can be used to estimate both ATT and CBF without significant discrepancy from the reference (MAEs of 231.45 ms for ATT and 9.80 mL/100 g/min for CBF using three of six PLDs).

Conclusion: The proposed machine learning-based ATT and CBF mapping offers substantially reduced scan time of multi-PLD pCASL.