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

To develop an effective deep learning (DL)-based method to denoise arterial spin labeling (ASL) data.

Conventional DL-based ASL denoising methods often suffer from overfitting and poor generalization when training data are limited. The proposed method overcame this problem using two strategies: (i) perform data augmentation to create large training data and (ii) denoise in-distribution and out-of-distribution components of the target perfusion-weighted image separately. Specifically, Image-to-Image Schrödinger Bridge (I²SB)-based distribution remapping transforms were applied to the large public ASL datasets so that their intensity distribution matched that of the data to be denoised. U-Net-based DL denoisers were trained on the remapped data to capture in-distribution features. High-SNR outputs from the DL-denoiser were incorporated into a Bayesian model to reconstruct the out-of-distribution features with sparsity constraints, generating denoised images for cerebral blood flow (CBF) quantification.

Simulation studies highlighted the importance of distribution remapping for effective data augmentation in limited-data scenarios. Both simulation and in vivo experiments showed that the proposed method outperformed state-of-the-art approaches, achieving an average SNR improvement of approximately 7 dB. Evaluations on multiple datasets confirmed robust and generalizable performance across different ASL sequences and imaging protocols. To demonstrate clinical potential, our method was applied to denoising stroke patient data (using only one-sixth of total averages with ˜83% reduction in scan time) and produced comparable CBF maps to the conventional ASL method.

The proposed method enables effective ASL denoising with limited training data. It has the potential to accelerate ASL acquisition, enhance image quality, and improve clinical utility.