SDnDTI: Self-supervised deep learning-based denoising for diffusion tensor MRI.

Area of Science:

• Neuroimaging
• Medical Physics
• Machine Learning

Background:

• Diffusion tensor magnetic resonance imaging (DTI) maps brain microstructure and white matter tracts.
• Image noise in DTI reduces accuracy and necessitates longer acquisition times.
• Supervised deep learning denoising requires high-SNR data, limiting practical use.

Purpose of the Study:

• To develop a self-supervised deep learning method (SDnDTI) for denoising DTI data.
• To eliminate the need for additional high-SNR data in DTI denoising.
• To improve the feasibility and application of deep learning for DTI analysis.

Main Methods:

• SDnDTI utilizes multi-directional DTI data, creating repetitions with identical contrasts but varying noise.
• A 3D convolutional neural network (CNN) denoises individual repetitions using an averaged higher-SNR target.
• Averaging CNN-denoised images enhances the overall signal-to-noise ratio (SNR).

Main Results:

• SDnDTI demonstrated high efficacy in denoising DTI data, validated on Human Connectome Project datasets.
• The method preserved image sharpness and textural details, outperforming conventional algorithms.
• Results were comparable to supervised learning methods, showing significant improvement over raw data.

Conclusions:

• SDnDTI offers a practical and effective self-supervised approach for DTI denoising.
• The method leverages diffusion MRI physics to enhance CNN-based denoising feasibility.
• SDnDTI has potential for accelerated DTI acquisition and improved brain microstructure mapping.