Comparison between supervised and physics-informed unsupervised deep neural networks for estimating cerebral...

Area of Science:

• Neuroimaging
• Medical Physics
• Artificial Intelligence

Background:

• Cerebral blood flow (CBF) and arterial transit time (ATT) are crucial biomarkers in neuroimaging.
• Multi-delay arterial spin labeling (ASL) is a non-invasive technique for quantifying these parameters.
• Accurate estimation of CBF and ATT is essential for diagnosing and monitoring various neurological conditions.

Purpose of the Study:

• To implement and evaluate a physics-informed unsupervised deep neural network (DNN) for estimating CBF and ATT from multi-delay ASL data.
• To compare the performance of the unsupervised DNN with a supervised DNN and a conventional method.
• To assess the accuracy and noise immunity of these methods using both simulated and in vivo data.

Main Methods:

• Trained supervised and unsupervised DNNs using simulated multi-delay ASL data.
• Compared CBF and ATT estimation accuracy and noise robustness across methods using simulations.
• Evaluated performance on in vivo data and assessed noise-induced variations by adding Rician noise.

Main Results:

• Supervised DNN demonstrated unbiased CBF estimation, unlike other methods with positive bias.
• Supervised DNN showed less bias in ATT estimation, with all methods behaving similarly with increasing noise.
• In vivo studies confirmed supervised DNN yielded the most accurate CBF and ATT; unsupervised DNN offered superior noise immunity for ATT estimation.

Conclusions:

• Physics-informed unsupervised learning effectively estimates CBF and ATT from multi-delay ASL, outperforming conventional methods.
• Supervised DNN generally showed superior performance in accuracy and bias reduction.
• Unsupervised DNN demonstrated specific advantages in noise immunity for ATT estimation, highlighting its potential in challenging imaging scenarios.