Open-Source Pipeline for Super-Resolution of Fetal NHP Diffusion MRI

NOTE: Currently under development.

This framework reconstructs high-resolution, motion-corrected 3D diffusion volumes from multiple low-resolution, motion-corrupted 2D stacks of fetal non-human primate (NHP) MRI data. It integrates state-of-the-art tools for distortion correction, motion correction, slice-to-volume reconstruction (SVR), and higher-order diffusion modeling to produce high-quality data suitable for advanced analysis, such as Fiber Orientation Distribution (FOD) for tractography.

Pipeline Overview

The pipeline is organized into three main stages:

1. Preprocessing: Each individual 2D diffusion stack is independently processed. This includes denoising, Gibbs ringing correction, brain extraction, bias field correction, and comprehensive correction for eddy currents, subject motion, and susceptibility-induced distortions using FSL eddy with either TOPUP- or T2-based field map estimation.

2. Super-Resolution & Alignment:

   • All preprocessed stacks are rigidly registered to a common reference stack.
   • Slice-to-Volume Reconstruction (SVR) is used to generate a single high-resolution, motion-corrected b=0 volume.
   • The SVR volume is aligned to a high-resolution T2-weighted anatomical template.
   • Finally, the full diffusion-weighted signal is reconstructed into the T2 template space at high resolution using an iterative method that models the signal with Spherical Harmonics.

3. Diffusion Modelling: The final high-resolution 4D DWI volume is used to:

   • Fit a diffusion tensor model to derive metrics like Fractional Anisotropy (FA) and Mean Diffusivity (MD).
   • Estimate Fiber Orientation Distributions (FODs) using Constrained Spherical Deconvolution (CSD) to resolve complex fiber crossings.

Prerequisites & Dependencies

This pipeline relies on several external software packages. You must have them installed and available in your system's PATH.

• FSL (v6.0 or later): For flirt, eddy, topup, fnirt, and other utilities.
• MRtrix3: For denoising, Gibbs correction, tensor and FOD modeling (dwidenoise, mrdegibbs, dwi2tensor, dwi2fod, etc.).
• ANTs: For N4 bias field correction and mask propagation (N4BiasFieldCorrection, antsApplyTransforms).
• Singularity / Apptainer: Required to run containerized versions of MIRTK and SVRTK.
  • MIRTK (mirtk.sif): Used for converting transformation formats.
  • SVRTK (svrtk.sif): The core toolkit for slice-to-volume reconstruction (mirtk reconstruct, mirtk reconstructDWI).
• Python 3: With libraries such as nibabel and numpy.

The pipeline assumes that brain masks have been generated beforehand. You can find the weights for our nnU-Net model here.

Data Structure

The pipeline is designed to work with data organized in a BIDS-like structure.

/path/to/project/
├── rawdata/
│ └── sub-<ID>/
|     └── ses-<ID>/
|           ├── dwi/
|           |   ├── sub-..._acq-ax_run-1_dwi.nii.gz
|           |   ├── sub-..._acq-ax_run-1_dwi.bval
|           |   ├── sub-..._acq-ax_run-1_dwi.bvec
|           |   ├── sub-..._acq-ax_run-1_dwi.json
|           |   └──... (other acquisitions/runs) ...
|           └── fmap/
|               ├── sub-..._acq-ax_run-1_epi.nii.gz
|               ├── sub-..._acq-ax_run-1_epi.json
|               └── ...
│
└── derivatives/
    ├── T2_recon/ (High-resolution T2s and masks)
    |   └── sub-<ID>/
    |       └── ses-<ID>/
    |           ├── T2_recon.nii.gz
    |           └── T2_recon_mask.nii.gz
    |
    ├── manual_masks/ (brain masks for DWI stacks)
    |    └── sub-<ID>/
    |        └── ses-<ID>/
    |            ├── acq-ax_run-1/
    |            │   └── mask.nii.gz
    |            └── ... (masks for each acq) ...
    |
    └── reconstruction/ (Output of this pipeline)
        └── sub-<ID>/
            └── ses-<ID>/
                ├── 01_prepared_stacks/
                ├── 02_preprocessed_stacks/
                ... (and so on) ...

Usage

The entire pipeline is orchestrated by the master script 00_run_pipeline.sh.

1. Clone the Repository:

   git clone https://github.com/MecaLab/Babofet_DWI.git
   cd Babofet_DWI

2. Configure the Master Script: Open 00_run_pipeline.sh and edit the USER CONFIGURATION section:

   • Set the correct paths for BIDS_ROOT_DIR and the Singularity image files (SVRTK_SIF_PATH, MIRTK_SIF_PATH).
   • Adjust reconstruction parameters (SVR_RESOLUTION, DWI_RECON_RESOLUTION, etc.) as needed.
   • Ensure the logic for finding the T2_TEMPLATE_IMAGE points to your high-resolution anatomical data.

3. Run the Pipeline: Execute the master script from the command line, providing the subject and session IDs as arguments:

   ./00_run_pipeline.sh <subject_id> <session_id>

   For example:

   ./00_run_pipeline.sh Aziza ses-01

   The script will create an output directory (derivatives/reconstruction/...) and run all processing steps sequentially.

Pipeline Stages in Detail

The pipeline is composed of modular scripts, each performing a specific task:

• 01_prepare_data.sh: Copies raw DWI data from the BIDS rawdata directory into the derivatives tree for processing.
• 02_preprocess_stacks.sh: The main per-stack preprocessing script. It performs denoising, degibbsing, distortion correction, bias correction, and eddy correction.
• 03_register_stacks.sh: Registers all preprocessed b=1000 mean images to a designated reference stack using FSL flirt.
• 04a_reconstruct_b0.sh: Uses mirtk reconstruct to create a single high-resolution, motion-corrected b=0 volume from all stacks.
• 04b_reconstruct_b1000.sh: (Optional but recommended) Reconstructs a high-resolution mean b=1000 image, which can be useful for registration.
• 05_align_to_t2.sh: Aligns the reconstructed b=0 volume to the subject's high-resolution T2 anatomical image. This defines the final space for the DWI reconstruction.
• 06_reconstruct_dwi.sh: The core reconstruction step. Uses mirtk reconstructDWI to reconstruct the full diffusion signal in the T2 template space.
• 07_fit_tensor.sh: Fits the diffusion tensor and CSD models to the final 4D DWI volume using MRtrix3.
• 08_mask_propagation.sh: Propagates anatomical tissue segmentations from the T2 space into the final DWI space for quantitative analysis.
• 99_quality_control.sh: Generates a series of "lightbox" images (e.g., FA, MD, color-FA maps) for quality control of the final outputs.