LAMAReg: Label Augmented Modality Agnostic Registration

We introduced a novel approach for more accurate registration between modalities. This python based workflow combines deep learning-based segmentation and numerical solutions (ANTs) to generate precise warpfields, even for modalities with low signal-to-noise ratio, signal dropout and strong geometric distortions, such as diffusion MRI and fMRI acquisitions. LAMAReg is implemented in a lightweight end-to-end T1/FLAIR/DWI pipeline as a seperate pip installable package (MICAFlow).

Overview

LAMAReg provides contrast-agnostic registration between different MRI modalities by using SynthSeg's brain parcellation to enable robust alignment between images with different contrasts (e.g., T1w to T2w, FLAIR to T1w, DWI to T1w).

This Python-based workflow combines deep learning-based segmentation (SynthSeg) and numerical optimization (ANTs) to generate precise warpfields, even for modalities with low signal-to-noise ratio, signal dropout, and strong geometric distortions, such as diffusion MRI and fMRI acquisitions.

!lamar_workflow

Installation

pip install -e .

Workflows

LAMAReg offers three main workflows and direct access to individual tools:

1. Full Registration Pipeline

Parcellate both input images, register them, and apply the transformation:

lamareg register [options]

2. Generate Warpfield Only

Create warpfields without applying them to the input image:

lamareg generate-warpfield [options]

3. Apply Existing Warpfield

Apply previously created warpfields to an input image:

lamareg apply-warpfield [options]

4. Direct Tool Access

Run individual components directly:

lamareg synthseg [options]      # Run SynthSeg brain parcellation
lamareg coregister [options]    # Run ANTs coregistration
lamareg apply-warp [options]    # Apply transformations
lamareg dice-compare [options]  # Calculate Dice similarity coefficient

Command-Line Arguments

Full Registration

Required Arguments:

• --moving PATH : Input image to be registered
• --fixed PATH : Reference image (target space)
• --output PATH : Output registered image

Optional Arguments:

• --moving-parc PATH : Path for moving image parcellation (temp file if omitted)
• --fixed-parc PATH : Path for fixed image parcellation (temp file if omitted)
• --registered-parc PATH : Path for registered parcellation (temp file if omitted)
• --affine PATH : Path for affine transform (temp file if omitted)
• --warpfield PATH : Path for primary warpfield (temp file if omitted)
• --inverse-warpfield PATH : Path for inverse primary warpfield
• --secondary-warpfield PATH : Output path for robust stage warpfield (requires --compose off)
• --inverse-secondary-warpfield PATH : Output path for inverse robust warpfield (requires --compose off)
• --compose : Compose primary and secondary warpfields into a single output warpfield
• --registration-method STR : Registration method (default: SyNRA)
• --threads N : threads (default: all cores)
• --qc-csv PATH : Path for QC Dice score CSV (required to enable QC; omit to skip automatically)
• --skip-fixed-parc : Skip fixed image parcellation if it already exists
• --skip-moving-parc : Skip moving image parcellation if it already exists
• --skip-qc : Skip quality control metrics calculation
• --disable-robust : Disable two-stage robust registration
• --verbose : Enable verbose output
• Providing exactly one warpfield (forward or inverse) without --compose is invalid.

ANTs Registration Parameters

When using coregister directly, additional ANts parameters are available:

• --verbose : Enable verbose output
• --grad-step FLOAT : Gradient step size (default: 0.2)
• --flow-sigma FLOAT : Smoothing for update field (default: 3)
• --total-sigma FLOAT : Smoothing for total field (default: 0)
• --aff-metric STR : Metric for affine stage (default: "mattes")
• --aff-sampling INT : Sampling parameter for affine metric (default: 32)
• --syn-metric STR : Metric for SyN stage (default: "mattes")
• --syn-sampling INT : Sampling parameter for SyN metric (default: 32)
• --reg-iterations STR : SyN iterations, comma-separated (e.g., "40,20,0")
• --aff-iterations STR : Affine iterations, comma-separated (e.g., "2100,1200,1200,10")
• --aff-shrink-factors STR : Affine shrink factors, comma-separated (e.g., "6,4,2,1")
• --aff-smoothing-sigmas STR : Affine smoothing sigmas, comma-separated (e.g., "3,2,1,0")
• --random-seed INT : Random seed for reproducibility
• --initial-affine-file PATH : Path to initial affine transform to use
• --initial-warp-file PATH : Path to initial warp field to use
• --interpolator STR : Interpolation method (default: "genericLabel")

Generate Warpfield

Same arguments as full registration, but without --output

Apply Warpfield

Required Arguments:

• --moving PATH : Input image to transform
• --fixed PATH : Reference space image
• --output PATH : Output registered image

Optional Arguments:

• --threads N : threads (default: All)
• --affine PATH : Path to affine transformation
• --warpfield PATH : Path to warp field
• --secondary-warpfield PATH : Path to secondary warp field

SynthSeg

Required Arguments:

• --i PATH : Input image
• --o PATH : Output segmentation

Optional Arguments:

• --parc : Output parcellation
• --cpu : Use CPU
• --threads N : Number of threads

Dice Compare

Required Arguments:

• --ref PATH : Path to reference parcellation image
• --reg PATH : Path to registered parcellation image
• --out PATH : Output CSV file path

Transform Order and Composition

Robust Mode (Two-Stage Registration)

When using robust mode (default), LAMAReg performs a two-stage registration:

1. Primary Registration: Parcellation-based registration (contrast-agnostic, coarse alignment)

   • Produces: primary_warp.nii.gz + affine.mat

2. Secondary Registration: Direct image registration using Stage 1 as initialization (fine-tuning)

   • Produces: secondary_warp.nii.gz (refinement)

3. Automatic Composition: Both warpfields are automatically composed into a single transform when --secondary-warpfield is not specified

   • Total transform = primary_warp ∘ secondary_warp

Important Note:

Warpfield composition results in small interpolation losses; leave --compose unset and capture both warpfields if you plan to reuse them separately.

Argument Parsing Logic

LAMAReg uses a subcommand-based CLI structure using the Python argparse library. Here's how it works:

1. Main Parser: Defines the global command structure with subparsers for each workflow.
2. Subparsers: Each workflow (register, generate-warpfield, etc.) has its own subparser with specific arguments.
3. Command Routing: The main() function routes commands to the appropriate functions in the lamar module.
4. Default Help: If no command is provided, the comprehensive help message is shown.
5. Unknown Arguments: For commands like synthseg, additional arguments are parsed from unknown_args to handle SynthSeg-specific options.
6. Direct Tool Access: Commands like coregister and apply-warp forward arguments to their respective modules.

All output files require explicit paths to ensure deterministic behavior and prevent accidental file overwrites.

Example Usage

Register DWI to T1w using example data:

lamareg register --moving example_data/sub-HC001_ses-02_space-dwi_desc-b0.nii.gz --fixed example_data/sub-HC001_ses-01_T1w.nii.gz \ 
  --output output/sub-001_dwi_in_T1w.nii.gz --moving-parc output/sub-001_dwi_parc.nii.gz \
  --fixed-parc output/sub-001_T1w_parc.nii.gz --registered-parc output/sub-001_dwi_reg_parc.nii.gz \
  --affine output/dwi_to_T1w_affine.mat --warpfield output/dwi_to_T1w_warp.nii.gz \
  --inverse-warpfield output/T1w_to_dwi_warp.nii.gz --threads 4 

Register without robust two-stage approach:

lamareg register --moving subject_flair.nii.gz --fixed subject_t1w.nii.gz \
  --output registered_flair.nii.gz --moving-parc flair_parcellation.nii.gz \
  --fixed-parc t1w_parcellation.nii.gz --affine flair_to_t1w_affine.mat \
  --warpfield flair_to_t1w_warp.nii.gz --disable-robust

Generate parcellations separately:

lamareg synthseg --i subject_t1w.nii.gz --o t1w_parcellation.nii.gz --parc
lamareg synthseg --i subject_flair.nii.gz --o flair_parcellation.nii.gz --parc

Register using existing parcellations:

lamareg register --moving subject_flair.nii.gz --fixed subject_t1w.nii.gz \
  --output registered_flair.nii.gz --moving-parc flair_parcellation.nii.gz \
  --fixed-parc t1w_parcellation.nii.gz --skip-fixed-parc --skip-moving-parc \
  --registered-parc registered_parcellation.nii.gz --affine flair_to_t1w_affine.mat \
  --warpfield flair_to_t1w_warp.nii.gz --inverse-warpfield t1w_to_flair_warp.nii.gz

Apply existing warpfield:

lamareg apply-warpfield --moving subject_flair.nii.gz --fixed subject_t1w.nii.gz \
  --output registered_flair.nii.gz --warpfield flair_to_t1w_warp.nii.gz \
  --affine flair_to_t1w_affine.mat

Evaluate registration quality:

lamareg dice-compare --ref reference_parcellation.nii.gz \
  --reg registered_parcellation.nii.gz --out dice_scores.csv

Working with Existing Parcellations

LAMAReg is designed to work with both new and existing parcellations:

1. Generate New Parcellations: If you provide paths to non-existing parcellation files, LAMAReg will generate them using SynthSeg.

2. Use Existing Parcellations: If the parcellation files already exist, LAMAReg will use them directly without regenerating them.

This flexibility allows you to:

• Process data end-to-end in a single command
• Pre-compute parcellations for reuse across multiple registrations
• Mix existing and new parcellations in your workflow

Technical Implementation

LAMAReg's registration approach consists of three main steps:

1. Brain Parcellation: SynthSeg generates contrast-agnostic parcellations of both the moving and fixed images.
2. Registration: ANTs registers the parcellations using the SyNRA method (rigid + affine + SyN).
3. Transformation Application: The resulting transformation is applied to the original moving image.

This approach enables accurate registration between images with different contrast properties where direct intensity-based registration might fail.

Robust Registration Mode

When not using the --disable-robust flag, LAMAReg performs a two-stage registration process:

1. First Stage: Register parcellations (contrast-agnostic approach)
2. Second Stage: Fine-tune with a second direct nonlinear registration using the first result as initialization

This two-stage approach can improve registration accuracy for cases where initial alignment is difficult, such as images with large geometric distortions or very different contrast mechanisms. The only reason to disable this is if you wanted to run your own second stage, or to speed up the runtime.

Directory Structure

LAMAReg/
├── setup.py
├── requirements.txt
├── README.md
├── lamareg/
│   ├── __init__.py
│   ├── cli.py
│   ├── scripts/
│   │   ├── lamar.py
│   │   ├── apply_warp.py
│   │   ├── coregister.py
│   │   ├── synthseg.py
│   │   └── dice_compare.py
│   ├── SynthSeg/
│   │   └── ... (SynthSeg code files)
│   └── ext/
│       ├── lab2im/
│       └── neuron/

Notes

• LAMAReg works with any MRI modality combination
• If parcellation files already exist, they will be used directly
• All output files need explicit paths to ensure deterministic behavior
• The transforms can be reused with the apply-warpfield command
• Use dice-compare to evaluate registration quality
• The robust mode performs a two-stage registration for improved accuracy:
  1. Register parcellations (contrast-agnostic)
  2. Fine-tune with a second direct registration using the first result as initialization
• For reproducible results, you can set a random seed when using the coregister command directly
• Temporary filenames are created automatically when optional outputs (parcellations, transforms) are omitted.
• Dice-based QC runs only when --qc-csv is provided; omitting it skips QC without needing --skip-qc.

References

1. Billot, Benjamin, et al. "Robust machine learning segmentation for large-scale analysis of heterogeneous clinical brain MRI datasets." Proceedings of the National Academy of Sciences 120.9 (2023): e2216399120.
2. Avants, Brian B., Nick Tustison, and Gang Song. "Advanced normalization tools (ANTS)." Insight j 2.365 (2009): 1-35.

License

This project is licensed under the CC-BY-NC License.

Contributors

• Ian Goodall-Halliwell
• Paul Bautin
• Nya Yazdi
• Kevin Du
• Raul R. Cruces