Deep Learning Image Denoising for In-Vivo Low-Field MRI Using Test-Time Training

Motivation: Low-field MRI, while portable and cost-effective, suffers from low SNR, limiting its clinical use. Goal(s): To investigate self-supervised test-time training as a fine-tuning method for low-field MRI denoising models, particularly for in-vivo data with different contrasts. Approach: A denoising model was pretrained with simulated low-field T2-weighted MRI data and further refined using self-supervised test-time training on in-vivo data. Model performance was assessed with and without test-time training across T2- and PD-weighted data. Results: The proposed fine-tuning by self-supervised test-time training provides the best compromise between denoising performance and preservation of structural details. Impact: Noise-free reference data for supervised training of low-field MRI denoising models does not exist. Using supervised pretraining on simulated data combined with self-supervised test-time training narrows the performance gap in low-field MRI denoising models when training and testing data differs.