Denoising arterial spin labeling perfusion MRI with deep machine learning.

Area of Science:

• Medical Imaging
• Neuroimaging
• Artificial Intelligence

Background:

• Arterial Spin Labeling (ASL) perfusion MRI is a noninvasive quantitative method for cerebral blood flow (CBF) measurement.
• Low signal-to-noise ratio (SNR) presents a significant challenge in ASL MRI data processing.
• Deep learning (DL) offers advanced capabilities for complex data processing tasks like denoising.

Purpose of the Study:

• To propose and validate a DL-based algorithm, termed DL-ASL, for denoising ASL MRI data.
• To address the inherent low SNR limitations in ASL MRI.
• To leverage DL's flexibility for improved ASL data processing.

Main Methods:

• Development of the DL-ASL network utilizing convolutional neural networks (CNNs).
• Incorporation of dilated convolutions and wide activation residual blocks within the CNN architecture.
• Focus on preserving spatial resolution and accounting for inter-voxel correlations during model training.

Main Results:

• DL-ASL demonstrated substantial improvements in ASL CBF quality, particularly in SNR.
• Retrospective analyses indicated DL-ASL could reduce acquisition time by up to 75% without compromising CBF measurement quality.
• The algorithm achieved superior denoising performance compared to existing methods, evidenced by higher PSNR and SSIM scores.

Conclusions:

• DL-ASL provides enhanced denoising for ASL MRI, outperforming conventional techniques.
• The algorithm leads to significant reductions in total acquisition time by enabling fewer repetitions.
• DL-ASL holds potential for more efficient and high-quality ASL MRI examinations.