Effects of the Training Data Condition on Arterial Spin Labeling Parameter Estimation Using a Simulation-Based...

Area of Science:

• Neuroimaging
• Medical Physics
• Artificial Intelligence

Background:

• Deep neural networks (DNNs) show promise for estimating cerebral blood flow (CBF) and arterial transit time (ATT) from arterial spin labeling (ASL) signals.
• The accuracy of these DNNs is highly dependent on the characteristics of the training dataset, particularly the ground truth (GT) ranges used.

Purpose of the Study:

• To investigate the impact of ground truth (GT) ranges for CBF and ATT on the performance of simulation-based supervised DNNs.
• To determine optimal GT ranges for training data to enhance the accuracy and reliability of ASL signal analysis.

Main Methods:

• Trained DNNs using 36 distinct training data patterns derived from ASL signal simulations.
• Evaluated DNN performance using simulation test data (1,000,000 points) and in vivo data from healthy volunteers and a moyamoya patient.
• Assessed accuracy, precision, and noise immunity using metrics like NMAE, NRMSE, and CV Net.

Main Results:

• The highest DNN performance was achieved with GT ranges of 0-120 mL/100 g/min for CBF and 0-4500 ms for ATT.
• While predicted CBF and ATT values varied with GT ranges, appropriate settings maintained DNN accuracy, precision, and noise immunity.
• These findings were consistent in both simulation and in vivo studies.

Conclusions:

• The selection of GT ranges for training data critically influences the performance of simulation-based supervised DNNs for ASL analysis.
• Appropriate GT range settings are essential for achieving accurate and precise estimations of CBF and ATT, whereas inappropriate settings can degrade performance.