Enhanced parameter estimation in multiparametric arterial spin labeling using artificial neural networks

Area of Science:

• Neuroimaging
• Medical Physics
• Artificial Intelligence in Medicine

Background:

• Multiparametric arterial spin labeling (MP-ASL) is crucial for quantifying cerebral blood flow (CBF) and arterial cerebral blood volume (CBV_a).
• Low signal-to-noise ratio (SNR) in MP-ASL limits accuracy and necessitates complex, time-consuming parameter estimation.
• Deep neural networks (DNNs) present a promising solution to overcome these MP-ASL limitations.

Purpose of the Study:

• To develop and evaluate simulation-based DNNs for MP-ASL parameter estimation.
• To compare the performance of a supervised DNN (DNN_Sup), a physics-informed unsupervised DNN (DNN_Uns), and the conventional lookup table (LUT) method.
• To assess DNN performance using both simulated and in vivo data.

Main Methods:

• MP-ASL data acquired during resting state and a breath-holding task.
• Evaluation of accuracy and noise immunity using resting-state data.
• Statistical comparison of CBF and CBV_a between resting and task states; assessment of reproducibility.

Main Results:

• DNN_Sup exhibited higher accuracy, noise immunity, and a six-fold faster computation time compared to LUT.
• All methods detected task-induced increases in CBF and CBV_a, with larger effect sizes observed for DNN_Sup and DNN_Uns.
• Reproducibility was comparable and satisfactory across all tested methods (DNN_Sup, DNN_Uns, and LUT).

Conclusions:

• Supervised DNNs (DNN_Sup) significantly outperform unsupervised DNNs (DNN_Uns) and lookup table methods (LUT) in MP-ASL parameter estimation.
• DNN_Sup offers superior accuracy, noise immunity, and computational efficiency for MP-ASL.
• DNNs represent a powerful tool for advancing quantitative neuroimaging with MP-ASL.