Automated generation of cerebral blood flow and arterial transit time maps from multiple delay arterial spin-labeled MRI

Purpose Develop and evaluate a deep learning approach to estimate cerebral blood flow (CBF) and arterial transit time (ATT) from multiple post-labeling delay (PLD) ASL MRI. Methods ASL MRI were acquired with 6 PLDs on a 1.5T or 3T GE system in adults with and without cognitive impairment (N = 99). Voxel-level CBF and ATT maps were quantified by training models with distinct convolutional neural network architectures: (1) convolutional neural network (CNN) and (2) U-Net. Models were trained and compared via 5-fold cross validation. Performance was evaluated using mean absolute error (MAE). Model outputs were trained on and compared against a reference ASL model fitting after data cleaning. Minimally processed ASL data served as another benchmark. Model output uncertainty was estimated using Monte Carlo dropout. The better-performing neural network was subsequently re-trained on inputs with missing PLDs to investigate generalizability to different PLD schedules. Results 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/100 g/min for CBF and 0.22 +/- 0.09 s for ATT. A significant association was observed between MAE and Monte Carlo dropout-based uncertainty estimates. Neural network performance remained stable despite systematically reducing the number of input images (i.e., up to 3 missing PLD images). Mean processing time was 10.77 s for the U-Net neural network compared to 10 min 41 s for the reference pipeline. Conclusion It is feasible to generate CBF and ATT maps from 1.5T and 3T multi-PLD ASL MRI with a fast deep learning image-generation approach.

Automated summary

This study developed and evaluated a deep learning approach to estimate CBF and ATT from multiple PLD ASL MRI. Different convolutional neural network architectures were trained and compared, and the U-Net model showed better performance. The U-Net model had lower MAE on both training and test data, and demonstrated good generalizability.