Denoising of ASL Data Using Deep Learning Priors Generated From Distribution Remapping

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Arterial Spin Labeling (ASL) is crucial for non-invasive brain perfusion imaging.
• Conventional deep learning (DL) methods for ASL denoising struggle with limited data, leading to overfitting and poor generalization.
• Improving ASL data quality is essential for accurate cerebral blood flow (CBF) quantification.

Purpose of the Study:

• To develop an effective deep learning (DL)-based method for denoising arterial spin labeling (ASL) data.
• To address the limitations of existing DL methods in scenarios with insufficient training data.
• To enhance the signal-to-noise ratio (SNR) and generalizability of ASL imaging.

Main Methods:

• Employed data augmentation via Image-to-Image Schrödinger Bridge (I²SB)-based distribution remapping to create larger, diverse training datasets.
• Separately denoised in-distribution and out-of-distribution components of the ASL data using U-Net-based DL denoisers and Bayesian reconstruction with sparsity constraints.
• Trained DL denoisers on remapped public ASL datasets to capture in-distribution features and reconstructed out-of-distribution features for improved image quality.

Main Results:

• Simulation studies confirmed the effectiveness of distribution remapping for data augmentation in limited-data scenarios.
• The proposed method achieved an average SNR improvement of approximately 7 dB, outperforming state-of-the-art approaches in both simulation and in vivo experiments.
• Demonstrated robust and generalizable performance across various ASL sequences and protocols, with comparable CBF maps to conventional methods even with an 83% reduction in scan time for stroke patient data.

Conclusions:

• The developed DL method effectively denoises ASL data, even with limited training datasets.
• The technique shows significant potential for accelerating ASL acquisition and enhancing overall image quality.
• This advancement promises to improve the clinical utility of ASL imaging for various neurological applications.