Arterial spin labeling MR image denoising and reconstruction using unsupervised deep learning

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Arterial spin labeling (ASL) is a non-invasive MRI technique for quantitative blood perfusion measurement.
• Clinical use of ASL is hindered by low signal-to-noise ratio (SNR), limited spatial resolution, and long acquisition times.

Purpose of the Study:

• To develop an unsupervised deep learning framework for denoising and reconstructing high-resolution ASL images.
• To improve SNR and accelerate imaging speed in ASL MRI.

Main Methods:

• An unsupervised deep learning framework utilizing subject-specific anatomical priors (e.g., T1-weighted images).
• Neural network trained from scratch using noisy ASL images or sparsely sampled k-space data.
• Validation with in vivo data from healthy subjects on a 3T MR scanner.

Main Results:

• The proposed framework significantly improved image quality and accelerated ASL imaging speed.
• Demonstrated superior performance compared to reference methods in both qualitative and quantitative analyses.
• Successfully enhanced SNR and spatial resolution of ASL images.

Conclusions:

• The unsupervised deep learning framework effectively addresses limitations of conventional ASL imaging.
• This method offers a promising approach for improving clinical applications of ASL by enhancing image quality and reducing scan times.
• Enables faster and more reliable assessment of tissue viability using ASL MRI.