Effects of the Training Data Condition on Arterial Spin Labeling Parameter Estimation Using a Simulation-Based Supervised Deep Neural Network

Abstract

Objective A simulation-based supervised deep neural network (DNN) can accurately estimate cerebral blood flow (CBF) and arterial transit time (ATT) from multidelay arterial spin labeling signals. However, the performance of deep learning depends on the characteristics of the training data set. We aimed to investigate the effects of the ground truth (GT) ranges of CBF and ATT on the performance of the DNN when training data were prepared using arterial spin labeling signal simulation. Methods Deep neural networks were individually trained using 36 patterns of the training data sets. Simulation test data (1,000,000 points), 17 healthy volunteers, and 1 patient with moyamoya disease were included. The simulation test data were used to evaluate accuracy, precision, and noise immunity of the DNN. The best-performing DNN was determined by the normalized mean absolute error (NMAE), normalized root mean squared error (NRMSE), and normalized coefficient of variation over repeated training (CVNet). Cerebral blood flow and ATT values and their histograms were compared between the GT and predicted values. For the in vivo data, the dependency of the predicted values on the GT ranges was visually evaluated by comparing CBF and ATT maps between the best-performing DNN and the other DNNs. Moreover, using the synthesized noisy images, noise immunity was compared between the best-performing DNN based on the simulation study and a conventional method. Results The simulation study showed that a network trained by the GT of CBF and ATT in the ranges of 0 to 120 mL/100 g/min and 0 to 4500 milliseconds, respectively, had the highest performance (NMAECBF, 0.150; NRMSECBF, 0.231; CVNET CBF, 0.028; NMAEATT, 0.158; NRMSEATT, 0.257; and CVNET ATT, 0.028). Although the predicted CBF and ATT varied with the GT range of the training data sets, the appropriate settings preserved the accuracy, precision, and noise immunity of the DNN. In addition, the same results were observed in in vivo studies. Conclusions The GT ranges to prepare the training data affected the performance of the simulation-based supervised DNNs. The predicted CBF and ATT values depended on the GT range; inappropriate settings degraded the accuracy, whereas appropriate settings of the GT range provided accurate and precise estimates.