Denoising ASL Images Using Distribution Remapping-Based Deep Learning

Motivation: ASL images often have low SNR. Improving their SNR can enhance the practical utility of ASL in various brain imaging applications. Goal(s): To develop a deep learning-based ASL denoising method with very limited training data available. Approach: To generate diverse sets of ASL training images for deep-learning denoiser for special ASL sequences, we employed Schrödinger-Bridge-based networks to map the large public ASL datasets to match both intensity and noise distributions of target ASL sequence. The training data was used to train deep-learning models for ASL denoising. Results: The proposed method was evaluated using simulation and in vivo experimental data, showing excellent denoising performance. Impact: Our proposed method addresses the issue of limited target ASL training datasets for deep learning-based ASL denoising and demonstrates excellent denoising performance. It can be generalized for the practical utility in both research and clinical applications.