Self-supervised learning for denoising of multidimensional MRI data.

Area of Science:

• Medical Imaging
• Machine Learning
• Signal Processing

Background:

• Quantitative MRI requires high signal-to-noise ratio (SNR) for accurate tissue parameter estimation.
• Noise in MRI data complicates complex, non-linear signal model fitting, impacting quantification.
• Acquiring ground truth clean MRI data for training is often impractical.

Purpose of the Study:

• To develop a rapid denoising framework for high-dimensional MRI data.
• To implement a self-supervised learning scheme that eliminates the need for clean reference images.
• To enhance the SNR and quantification accuracy of MRI data.

Main Methods:

• A deep learning framework exploiting redundancy in multidimensional MRI data was proposed.
• A self-supervised model was trained using only noisy images, addressing the lack of clean data.
• The framework was validated on simulated MTC-MRF and in vivo DWI datasets.

Main Results:

• The proposed method significantly outperformed existing techniques (BM3D, tMPPCA, Patch2self) in denoising.
• Performance improvements were observed across various noise levels and distributions.
• Denoising led to more accurate quantitative results from the processed MRI data.

Conclusions:

• The MD-S2S denoising technique effectively enhances multidimensional MRI data.
• This approach can be applied to diverse MRI datasets, improving quantification accuracy.
• The framework offers a promising solution for robust quantitative MRI.