Motivation: Developing denoising deep learning models that can generalize effectively across diverse MRI test distributions, overcoming limitations posed by data scarcity in training datasets. Goal(s): To determine if joint supervised and self-supervised learning can outperform traditional fully supervised (FS) training on Out-Of-Distribution (OOD) MRI image denoising problems. Approach: We employ a joint training methodology that combines both FS and self-supervised contrastive loss terms for MRI image denoising. This approach is evaluated for its robustness to several OOD test datasets. Results: Our joint training approach outperforms FS learning on OOD image denoising problems at low noise levels; however, its efficacy diminishes at higher noise levels. Impact: This research demonstrates improved denoising performance on low-noise OOD MRI data, addressing a key challenge in generalizing to diverse imaging conditions with limited training data.
Zaki et al. (Tue,) studied this question.