Automated Pipeline for Processing and Analysis of Clinical Fetal Diffusion MRI

Pipeline Overview

The pipeline is organized into a series of sequential steps, each handled by a dedicated script. The workflow is as follows:

1. Data Preparation (01_prepare_data.sh): Organizes raw BIDS-like dMRI data into the processing directory structure.
2. Preprocessing (02_preprocessing.sh):
   • Denoising using Marchenko-Pastur PCA (dwidenoise).
   • Brain extraction using a pre-trained deep learning model (Fetal-BET).
   • N4 bias field correction (N4BiasFieldCorrection).
   • Motion and eddy current correction using FSL's eddy with slice-to-volume correction. The script robustly tries several outlier detection thresholds to ensure convergence.
3. Registration (03_registration.sh):
   • Selects the closest age-matched template from the spatiotemporal dHCP fetal diffusion atlas.
   • Performs affine registration (flirt) followed by high-dimensional non-linear registration (antsRegistration - SyN) to align the atlas to the subject's native space.
   • Propagates an extended set of anatomical parcellations to the subject's native space for regional analysis.
4. Diffusion Tensor Modeling (04_tensor_fitting.sh):
   • Reorients b-vectors according to the registration transform.
   • Fits the diffusion tensor model and calculates metrics (FA, MD, AD, RD, CFA) in two spaces:
     1. Subject's native space (for native-space analysis).
     2. Standard atlas space (for group comparisons and visual QC).
5. Quality Control (99_qc.sh):
   • Extracts quantitative QC metrics (e.g., motion parameters, outlier slices).
   • Generates visual reports (lightbox images) of the resulting tensor maps (FA, MD, CFA) for easy inspection.

Dependencies

This pipeline relies on several external neuroimaging software packages. Please ensure they are installed and available in your system's PATH.

• FSL: Used for eddy, flirt, and various image manipulation tools.

• MRtrix3: Used for denoising, file conversion, and diffusion tensor modeling.

• ANTs (Advanced Normalization Tools): Used for N4 bias field correction and non-linear registration.

• Python 3: With a configured environment (e.g., Conda) including:

  • Nibabel
  • Numpy
  • PyTorch (for the Fetal-BET brain extraction tool)

• Fetal-BET: The brain extraction tool and its pre-trained model must be available.

The main execution script (00_run_pipeline.sh) is configured to load these dependencies using module and conda. You may need to adapt these lines to match your system's configuration.

Data Structure

Input Data

The pipeline expects input data to be organized in a BIDS-like structure. The main script takes the subject, session, and series ID as arguments.

/path/to/rawdata/
└── <SUBJECT_ID>
    └── <SESSION_ID>
        └── <SERIE_ID>
            ├── dwi.nii.gz
            ├── dwi.bval
            ├── dwi.bvec
            └── dwi.json

Output Data

All processed data and intermediate files are saved in a derivatives directory, structured as follows:

/path/to/derivatives
    └── <SUBJECT_ID>
        └── <SESSION_ID>
            └── <SERIE_ID>
                ├── 01_prepared_data
                ├── 02_preprocessed_data
                ├── 03_registration
                ├── 04_tensor
                └── 99_QC

• 02_preprocessed_data: Contains the motion-corrected DWI (dwi_eddycorr.nii.gz) and brain mask.
• 03_registration: Contains registration transforms and warped parcellations .
• 04_tensor: Contains the final diffusion metrics in both native and template space.
• 99_QC: Contains QC reports, including lightbox images of the tensor maps.

Atlas and Parcellations

This pipeline relies on the dHCP spatiotemporal fetal diffusion atlas. You must download it and provide the correct path in the 03_registration.sh script (ATLAS_DIR).

A key contribution of this work is the use of an extended set of parcellations, which have been pre-registered to the dHCP atlas. These include:

• CRL2025 Fetal Atlas Parcellations
• dHCP Neonatal Atlas Parcellations
• dHCP Neonatal Multi Channel White Matter Parcellations

These extended parcellations are available in the parcellation.zip file.

The pipeline automatically warps these parcellations into each subject's native space, enabling detailed region-of-interest (ROI) analysis.

Usage

1. Configuration:

   • Clone this repository.
   • Update the paths in 00_run_pipeline.sh (e.g., ROOT_DIR, DERIVATIVES_DIR) to match your environment.
   • Update the atlas paths in 03_registration.sh (ATLAS_DIR, PARCELLATIONS_DIR).
   • Ensure the dependency loading commands (module load, conda activate) in 00_run_pipeline.sh are correct for your system.

2. Execution: The pipeline is launched using the main wrapper script 00_run_pipeline.sh. It requires three arguments: SUBJECT_ID, SESSION_ID, and SERIE_ID.

   bash 00_run_pipeline.sh <SUBJECT_ID> <SESSION_ID> <SERIE_ID>

   Example:

   bash 00_run_pipeline.sh sub-0001 ses-0024 13

   This command will execute all steps from data preparation to final QC report generation for the specified scan.