[Paper Review] Patch2Self: Denoising Diffusion MRI with Self-Supervised Learning

Shreyas Fadnavis, Joshua Batson+1 more|arXiv (Cornell University)|Nov 2, 2020

Advanced Neuroimaging Techniques and ApplicationsMedicine46 references40 citations

TL;DR

Patch2Self denoises diffusion MRI data using a self-supervised, patch-based regression across volumes, achieving improved microstructure modeling and tractography without a noise model, outperforming Marchenko-Pastur PCA on real and simulated data.

ABSTRACT

Diffusion-weighted magnetic resonance imaging (DWI) is the only noninvasive method for quantifying microstructure and reconstructing white-matter pathways in the living human brain. Fluctuations from multiple sources create significant additive noise in DWI data which must be suppressed before subsequent microstructure analysis. We introduce a self-supervised learning method for denoising DWI data, Patch2Self, which uses the entire volume to learn a full-rank locally linear denoiser for that volume. By taking advantage of the oversampled q-space of DWI data, Patch2Self can separate structure from noise without requiring an explicit model for either. We demonstrate the effectiveness of Patch2Self via quantitative and qualitative improvements in microstructure modeling, tracking (via fiber bundle coherency) and model estimation relative to other unsupervised methods on real and simulated data.

Motivation & Objective

• Motivate denoising of diffusion-weighted MRI (DWI) data to improve downstream microstructure analysis and tractography.
• Develop a self-supervised denoising framework that does not require an explicit noise model.
• Leverage 4D DWI structure (multiple volumes) to separate signal from noise via local patch-based regression.
• Show that a simple linear regressor within Patch2Self can outperform state-of-the-art unsupervised denoising methods.

Proposed method

• Construct p-neighborhood patches around each voxel in every 3D volume of the 4D DWI data.
• Form a flattened feature matrix of size m × (p^3 × n) for m voxels and n volumes.
• Hold out the target volume j and train a J-invariant regressor Φ_J on Y_{*,*,-j} to predict Y_{*,0,j} via linear regression.
• Denoise by applying the learned Φ_J to the held-out p-neighborhoods to obtain denoised volume X̂_{*,*,*,j} across all volumes.
• Justify J-invariance via independence of noise across volumes, ensuring denoising preserves the true signal.
• Note that the regressor can be linear or nonlinear; linear regression offered comparable performance with faster training.

Experimental results

Research questions

• RQ1 Can self-supervised, patch-based regression denoise DWI without requiring a noise model?
• RQ2 Does Patch2Self improve downstream microstructure modeling (DTI, CSD) and tractography compared to existing unsupervised methods?
• RQ3 What is the impact of patch size and regressor choice on denoising performance across different acquisition schemes?
• RQ4 Is a simple linear regressor sufficient for effective denoising in Patch2Self?

Key findings

• Patch2Self yields visually coherent denoised outputs without introducing anatomical artifacts.
• In tractography, Patch2Self reduces incoherent streamlines and improves fiber bundle coherency compared to Marchenko-Pastur denoising.
• Patch2Self improves goodness-of-fit (R^2) for both DTI and CSD models relative to noisy and MP-PCA denoising, with substantial gains (e.g., across voxel locations CC and CSO).
• For DKI, Patch2Self reduces degeneracies in parameter estimation compared to noisy and MP-PCA denoising.
• Simulated data show substantial performance gains (higher R^2, lower RMSE) in low-to-moderate SNR ranges (5–20) over MP-PCA, with Patch2Self improving as SNR increases.