FSL

Preprocessing MRI images with FSL

All structural MRI data were preprocessed using a combination of neuroimaging tools including FSL (version 6.0.7.7), ANTs (version 2.2.0), and dcm2niix (version v1.0.20240202).

Prior to preprocessing, input files were converted to standardized NIFTI_GZ format using dcm2niix for DICOM files and fslchfiletype for ANALYZE format files. The preprocessing pipeline consisted of several sequential steps to ensure data quality and standardization across multiple datasets.

Overview of the preprocessing pipeline:

Fig SI: Flowchart of the full FSL-based preprocessing pipeline.

Step-by-step preprocessing:

  1. File Conversion & Reorientation
    • DICOM to NIFTI using dcm2niix
    • ANALYZE to NIFTI using fslchfiletype
    • Reoriented to standard space using fslreorient2std
T1w image after reorientation to standard.
  1. Bias Field Correction
    • Applied using N4BiasFieldCorrection from ANTs
Image after ANTs bias field correction.
  1. Robust FOV Cropping & Brain Extraction
    • Used robustfov to crop neck/lower head
    • Skull stripping with bet2, intensity threshold = 0.3
Left: Cropped image using `robustfov`; Right: Skull-stripped image using BET.
  1. Registration to MNI Space
    • Linear registration using FLIRT
    • DOF = 6 and 12
    • Templates: MNI152 (1mm and 2mm)
Linear registrations to MNI152 space using FLIRT with different DOFs and resolutions.
  1. Tissue Segmentation
    • Performed using FAST
    • Segmented into 3 tissue types: CSF, GM, WM
  2. White Matter Mask Creation
    • Thresholded WM partial volume map at 0.9
White matter partial volume map and the binarized WM mask after thresholding.
  1. White Stripe Intensity Normalization
    • Binary WM mask applied to T1w image
    • Mean WM intensity computed using fslstats
    • T1w image normalized using fslmaths
Final normalized T1w image using white stripe normalization.

Additional Notes

Other normalization techniques including min-max normalization and histogram matching were explored. However, white stripe normalization was ultimately selected due to its biological interpretability—the histogram peaks corresponding to CSF, WM, and GM were consistent across datasets and visually preserved tissue contrast with minimal artifacts.

Hardware & Execution Environment
The pipeline was parallelized and executed on an HPC cluster with the following specs:

  • CPU: Intel® Xeon® Gold 6240 (2.60GHz, dual CPUs)
  • Memory: 192 GB RAM

This allowed robust and scalable preprocessing across large neuroimaging datasets.

References