Source-linked AI summary
Patch2Self: Denoising Diffusion MRI with Self-Supervised Learning
Shreyas Fadnavis, Joshua Batson, Eleftherios Garyfallidis
TL;DR
DWI noise limits microstructure modeling, tractography, and related analyses, while existing unsupervised denoisers can require noise-model specification. Patch2Self uses self-supervised, locally linear prediction across acquisition volumes, and outperforms Marchenko-Pastur denoising on subsequent analyses in real and simulated data.
Problem
DWI noise limits visual, microstructure, and tractography analyses, while existing denoising methods can require explicit or implicit noise-model information.
Method
Patch2Self learns locally linear mappings from spatial patches in other acquisition volumes to denoise each held-out DWI volume using self-supervised J-invariant regression.
Results
Patch2Self outperforms the state-of-the-art Marchenko-Pastur method on downstream microstructure modeling, tractography, and other analyses.
Takeaways & Limitations
Patch2Self provides a denoising approach intended to improve downstream DWI analysis without requiring explicit noise-model selection or calibration.
Takeaways & Limitations
The J-invariance guarantee assumes that noise is independent across acquisition volumes.
Abstract
from arXiv · showhide
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.
1 Introduction
DWI enables tissue-microstructure measurement and white-matter reconstruction, but substantial, variable noise limits downstream analysis. Patch2Self addresses this by denoising without selecting or calibrating an explicit noise model.
- 1 Introduction: DWI noise confounds visual inspection, microstructure modeling, and tractography, making denoising necessary before anatomical inference.
- 1 Introduction: Many gradient directions are acquired because voxels can contain multiple tissue types, creating repeated representations that also support denoising.
- 1 Introduction: Existing DWI denoisers include image-based filters and local low-rank methods, with Marchenko-Pastur PCA described as the state-of-the-art unsupervised approach.
- 1 Introduction: Marchenko-Pastur PCA and related methods require explicit or implicit noise-model information for denoising.
- 1 Introduction: Patch2Self removes the need to select or calibrate a noise model while denoising DWI data through self-supervised learning.The method is evaluated against other unsupervised approaches on real and simulated data.
2 Self-Supervised Local Low Rank Approximation
Patch2Self uses local spatial patches from other DWI volumes to predict each held-out volume in a self-supervised regression scheme. Its J-invariance relies on independent noise across volumes, while linear regression offers a fast practical implementation.
- 2.1.2 Patches and Local Matrix Approximations: Patch2Self learns a global estimator per 3D volume from local patches sampled across the remaining volumes, combining local and non-local information.
- 2.2 Denoising via Self-Supervised Local Approximations: Patch2Self trains on local patches from volumes excluding v_j and predicts the center voxels of held-out volume v_j to construct denoised 4D DWI data.Each volume is processed by extracting neighborhoods, fitting a regressor, predicting held-out voxels, and unraveling the outputs.
- 2.2 Denoising via Self-Supervised Local Approximations: The method converts each 4D DWI dataset into m samples with p^3 × n patch features by extracting p-neighbourhoods around every voxel.The resulting features combine local spatial neighborhoods across all acquisition volumes.
- 2.2 Denoising via Self-Supervised Local Approximations: J-invariance makes the self-supervised loss align with ground-truth denoising loss when noise is independent across acquisition volumes and predictions exclude the target volume.The method is constructed so each predictor uses signal from volumes other than the one being denoised.
- 2.1.2 Patches and Local Matrix Approximations: Linear regression performs comparably to more sophisticated regressors while training faster, although the framework supports linear, regularized, and nonlinear models.
- 2.1.2 Patches and Local Matrix Approximations: With fewer acquisition volumes, a larger patch radius is required, while the RMSE difference between patch radii decreases as volume count increases.
3 Evaluation on Real Data
On real DWI datasets, Patch2Self produced visually coherent denoised outputs, reduced spurious tractography, and improved downstream microstructure model fitting relative to noisy and Marchenko-Pastur data.
- 3.1 Evaluation on in-vivo data: Patch2Self produced more visually coherent outputs than Marchenko-Pastur across PPMI, Stanford HARDI, and Sherbrooke 3-Shell datasets without anatomical features in residual maps.The datasets included single-shell, high-angular-resolution, and multi-shell acquisition schemes.
- 3.2 Effect on Tractography: Patch2Self further reduced spurious tracts after Marchenko-Pastur denoising, yielding a cleaner optic-radiation fiber-bundle representation.Fiber Bundle Coherency maps showed fewer incoherent streamlines after denoising.
- 3.2 Effect on Tractography: Patch2Self alleviated more degeneracies in mean-kurtosis estimation and produced more consistent downstream DTI and CSD model fitting than Marchenko-Pastur.Degeneracies appeared as black voxels, while box plots quantified R2 improvement over noisy data.
- 3.3 Impacts on Microstructure Model Fitting: Patch2Self achieved higher cross-validated R2 than Marchenko-Pastur by 22% in corpus callosum and 65% in centrum semiovale.The comparison used CSD and DTI fits at single-fiber and crossing-fiber voxel locations.
4 Evaluation on Simulated Data
On simulated DWI with SNRs from 5 to 30, Patch2Self was evaluated against noisy and Marchenko-Pastur data using ground-truth comparisons. It suppressed more noise and consistently improved as SNR increased.
- 4 Evaluation on Simulated Data: The simulation used noise-free whole-brain DWI with b-values of 0, 1000, and 2000 s/mm2 plus realistic multi-channel coil noise.The simulated acquisition included 2 b0 volumes and 30 directions at each nonzero b-value.
- 4 Evaluation on Simulated Data: Patch2Self consistently outperformed Marchenko-Pastur on simulated data across realistic SNRs, evaluated using MSE and R2 metrics.The strongest gains occurred in the SNR 5–20 range common for in-vivo imaging.
- 4 Evaluation on Simulated Data: Patch2Self produced visibly cleaner simulated outputs than Marchenko-Pastur at low SNRs and showed increasing performance with higher SNR.Figure 6 compared noisy, Marchenko-Pastur, and Patch2Self outputs against ground truth.
5 Conclusions
The paper presents Patch2Self as a DWI denoising method designed to improve downstream analysis of low-SNR data. It reports better performance than Marchenko-Pastur on microstructure modeling, tractography, and related analyses.
- 5 Conclusions: Patch2Self outperformed the unsupervised Marchenko-Pastur method on downstream microstructure modeling, tractography, and other analyses.The authors propose incorporating an efficient, unit-tested implementation into DIPY.
Broader Impacts
The broader impacts include potential benefits for medical imaging and self-supervised learning, alongside societal risks from improved extraction of information from poor measurements.
- Broader Impacts: Better denoising could enable higher-quality images with fewer or shorter acquisitions and support advanced acquisition schemes, biomarkers, and visualization of small structures.The stated medical-imaging impacts include potential clinical applications such as spinal-cord visualization.
- Broader Impacts: Patch2Self’s self-supervised signal extraction could also expand surveillance capabilities by enabling identities, movements, or disease status to be obtained from farther away and at lower cost.This is presented as a societal risk associated with extracting signal from poor measurements.