Rep2Rep-Shuffle: Shuffled Repetition-to-Repetition Learning for Noise-Adaptive Self-Supervised MRI Denoising

MRI applications with low SNR, such as non-proton imaging and low-field imaging, often require multiple signal averages to achieve adequate image quality at the cost of longer scan times. This work proposes Rep2Rep-Shuffle, a self-supervised learning framework that improves MRI denoising performance in repeated-acquisition imaging with limited training datasets.

Rep2Rep-Shuffle extends our previously proposed repetition-to-repetition learning (Rep2Rep)-based self-supervised denoising approach with a new shuffled repetition pairing strategy. Instead of using a fixed split between training input and target repetitions, different repetition combinations are randomly generated during training epochs for each dataset, which increases the effective training space when training datasets are limited. In addition, the framework also extends from slice-wise 2D to 3D denoising by exploiting correlations across all spatial dimensions. The proposed method was tested in vivo using sodium MRI and further validated using synthetic noisy MRI datasets. Different training strategies for both 2D and 3D denoising were compared.

In vivo sodium MRI results demonstrated improved denoising and better preservation of anatomical structures with 3D Rep2Rep-Shuffle compared with fixed pairing (Rep2Rep) and 2D denoising. Reader evaluation further confirmed that 3D Rep2Rep-Shuffle received higher scores for different assessment categories. In synthetic experiments, 3D Rep2Rep-Shuffle achieved the best performance among self-supervised methods and approached the performance of supervised learning.

Rep2Rep-Shuffle outperforms Rep2Rep for MRI denoising by increasing training diversity through shuffled repetition pairing and by leveraging volumetric spatial correlations. The method is particularly suited for low-SNR MRI applications that acquire multiple repetitions but have limited training datasets.