Denoise

denoise - Diffusion-Weighted Image Noise Reduction

Part of the micaflow processing pipeline for neuroimaging data.

This module denoises diffusion-weighted images (DWI) using the Patch2Self algorithm, which leverages redundant information across diffusion gradients to remove noise without requiring additional reference scans. Patch2Self is a self-supervised learning approach that improves image quality and enhances subsequent diffusion analyses by removing random noise while preserving anatomical structures and signal characteristics.

The Patch2Self Algorithm:

Patch2Self is a denoising method that: - Uses the redundancy in diffusion MRI data across gradient directions - Learns to predict each volume from surrounding volumes in q-space - Requires no external training data or noise model - Preserves anatomical edges and tissue boundaries - Improves SNR without introducing spatial blurring

Features:

Self-supervised learning approach requiring no separate reference data
Adapts to the unique noise characteristics of each dataset
Preserves anatomical structure while effectively removing noise
Compatible with standard diffusion MRI acquisition protocols
Improves subsequent analyses such as fiber tracking and diffusion metrics
Optional separate b0 volume denoising
Uses Ordinary Least Squares (OLS) regression model
Intensity shifting to ensure positive values

Command-Line Usage:

# Standard denoising (recommended for most cases) micaflow denoise

–input <path/to/dwi.nii.gz> –bval <path/to/dwi.bval> –bvec <path/to/dwi.bvec> –output <path/to/denoised_dwi.nii.gz>

# With separate b0 denoising (experimental) micaflow denoise

–input <path/to/dwi.nii.gz> –bval <path/to/dwi.bval> –bvec <path/to/dwi.bvec> –output <path/to/denoised_dwi.nii.gz> –b0-denoise

# With Gibbs ringing removal micaflow denoise

–input <path/to/dwi.nii.gz> –bval <path/to/dwi.bval> –bvec <path/to/dwi.bvec> –output <path/to/denoised_dwi.nii.gz> –gibbs

Python API Usage:

>>> from micaflow.scripts.denoise import run_denoise
>>>
>>> # Standard denoising
>>> output = run_denoise(
...     moving="raw_dwi.nii.gz",
...     moving_bval="dwi.bval",
...     moving_bvec="dwi.bvec",
...     output="denoised_dwi.nii.gz",
...     b0_denoising=False  # Default
... )
>>>
>>> # With b0 denoising
>>> output = run_denoise(
...     moving="raw_dwi.nii.gz",
...     moving_bval="dwi.bval",
...     moving_bvec="dwi.bvec",
...     output="denoised_dwi.nii.gz",
...     b0_denoising=True
... )
>>>
>>> # With Gibbs ringing removal
>>> output = run_denoise(
...     moving="raw_dwi.nii.gz",
...     moving_bval="dwi.bval",
...     moving_bvec="dwi.bvec",
...     output="denoised_dwi.nii.gz",
...     gibbs=True
... )

Pipeline Integration:

Denoising is typically the FIRST step in diffusion preprocessing: 1. Denoising (denoise) ← You are here 2. Motion/eddy current correction (motion_correction) 3. Susceptibility distortion correction (apply_SDC) 4. Bias field correction (bias_correction) 5. Brain extraction (bet) 6. DTI metric calculation (compute_fa_md)

Exit Codes:

0 : Success - denoising completed 1 : Error - invalid inputs, file not found, or processing failure

Technical Notes:

Patch2Self uses a self-supervised regression approach
Model type: Ordinary Least Squares (OLS) regression
B0 threshold: 50 s/mm² (volumes below this considered b0)
Intensity shifting: Enabled (ensures positive values)
Negative value clipping: Disabled (preserves signal characteristics)
Processing time: ~1-2 minutes for typical datasets (varies with # volumes)
Recommended to denoise BEFORE motion correction for best results
B0 denoising is experimental; standard mode excludes b0 from denoising

References:

Fadnavis S, Batson J, Garyfallidis E. Patch2Self: Denoising Diffusion MRI with Self-Supervised Learning. Advances in Neural Information Processing Systems. 2020;33:16293-16303.
Fadnavis S, Farooq H, Theaud G, et al. Patch2Self denoising of diffusion MRI in the CLEAR test-retest dataset. Medical Imaging 2021: Image Processing. 2021;11596:115962D. doi:10.1117/12.2582043

Command Line Usage

micaflow denoise [options]

Source Code

View the source code: GitHub Repository

Description

This script denoises diffusion-weighted images (DWI) using the Patch2Self: algorithm, which leverages redundant information across diffusion gradients to remove noise without requiring additional reference scans.

Full Help

╔════════════════════════════════════════════════════════════════╗
║                     DWI IMAGE DENOISING                        ║
╚════════════════════════════════════════════════════════════════╝

This script denoises diffusion-weighted images (DWI) using the Patch2Self
algorithm, which leverages redundant information across diffusion gradients
to remove noise without requiring additional reference scans.

────────────────────────── USAGE ──────────────────────────
  micaflow denoise [options]

─────────────────── REQUIRED ARGUMENTS ───────────────────
  --input     : Path to the input DWI image (.nii.gz)
  --bval      : Path to the b-values file (.bval)
  --bvec      : Path to the b-vectors file (.bvec)
  --output    : Output path for the denoised image (.nii.gz)

─────────────────── OPTIONAL ARGUMENTS ───────────────────
  --b0-denoise: Denoise b0 volumes separately (default: False)
               Experimental - not recommended for most cases
  --gibbs     : Apply Gibbs ringing removal after denoising
  --threads   : Number of threads for Gibbs removal (default: 1)

──────────────────── EXAMPLE USAGE ──────────────────────

# Standard denoising (recommended)
micaflow denoise \
  --input raw_dwi.nii.gz \
  --bval dwi.bval \
  --bvec dwi.bvec \
  --output denoised_dwi.nii.gz

# With b0 denoising (experimental)
micaflow denoise \
  --input raw_dwi.nii.gz \
  --bval dwi.bval \
  --bvec dwi.bvec \
  --output denoised_dwi.nii.gz \
  --b0-denoise

# With Gibbs ringing removal
micaflow denoise \
  --input raw_dwi.nii.gz \
  --bval dwi.bval \
  --bvec dwi.bvec \
  --output denoised_dwi.nii.gz \
  --gibbs

────────────────── PATCH2SELF ALGORITHM ──────────────────

How it works:
• Uses redundancy across diffusion gradient directions
• Predicts each volume from neighboring volumes in q-space
• Self-supervised learning (no external training data needed)
• Preserves edges and anatomical structures

Benefits:
• Better visualization of white matter tracts
• More accurate DTI fitting and tractography
• Reduced bias in diffusion metrics (FA, MD)
• No spatial blurring (unlike traditional filters)

────────────────────────── NOTES ─────────────────────────
• Denoising should be the FIRST step in preprocessing pipeline
• Apply BEFORE motion correction for best results
• Uses Ordinary Least Squares (OLS) regression model
• B0 threshold: 50 s/mm² (volumes below this are b0)
• Intensity shifting enabled to ensure positive values
• Processing time: ~1-2 minutes (depends on # volumes)
• Standard mode excludes b0 volumes from denoising (recommended)
• --b0-denoise flag enables experimental b0 denoising
• Output preserves original image geometry and orientation

─────────────── PIPELINE POSITION ───────────────────────
1. Denoising (denoise)           ← You are here
2. Motion correction (motion_correction)
3. Distortion correction (apply_SDC)
4. Bias field correction (bias_correction)
5. Brain extraction (bet)
6. DTI metrics (compute_fa_md)

────────────────────── EXIT CODES ───────────────────────
0 : Success - denoising completed
1 : Error - invalid inputs, file not found, or processing failure

───────────────── COMMON ISSUES ─────────────────────────
Issue: Denoising takes very long
Solution: Typical processing is 2-5 minutes; longer suggests high resolution

Issue: Output looks over-smoothed
Solution: Patch2Self preserves edges; check if input already denoised

Issue: "Out of memory" error
Solution: Reduce image resolution or process on machine with more RAM