Related Experiment Video
Updated: Apr 12, 2026
Identification of Disease-related Spatial Covariance Patterns using Neuroimaging Data
Published on: June 26, 2013
Reproducibility of neuroimaging analyses across operating systems
Tristan Glatard1, Lindsay B Lewis2, Rafael Ferreira da Silva3
1McConnell Brain Imaging Centre, Montreal Neurological Institute, McGill University Montreal, QC, Canada ; Centre National de la Recherche Scientifique, University of Lyon, INSERM, CREATIS Villeurbanne, France.
Neuroimaging analysis pipelines produce varying results across different computing platforms. This study quantifies discrepancies in brain tissue classification, fMRI, and cortical thickness extraction, highlighting the need for improved numerical precision in pipelines.
Area of Science:
- Neuroimaging analysis
- Computational neuroscience
- Scientific reproducibility
Background:
- Neuroimaging pipelines are susceptible to variations based on the computational environment.
- Differences in results can impact the reliability of neuroimaging research.
- Standardization is crucial for reproducible scientific findings.
Purpose of the Study:
- To quantify the impact of computing platforms on neuroimaging pipeline outputs.
- To identify the causes of discrepancies in brain tissue classification, fMRI, and cortical thickness (CT) extraction.
- To propose solutions for enhancing the reproducibility of neuroimaging analyses.
Main Methods:
- Utilized three major neuroimaging packages: FSL, Freesurfer, and CIVET.
- Tested pipelines on different GNU/Linux operating system versions.
- Employed library and system call interception to identify sources of variation.
- Quantified differences using Dice coefficients for tissue classification and assessed fMRI and CT extraction variations.
Main Results:
- Significant differences observed in subcortical classification (Dice coefficients down to 0.59), fMRI analysis (ICA discrepancies in one-third of subjects due to motion correction), and cortical thickness extraction.
- Simple pipelines like brain extraction showed minimal impact, while complex analyses accumulated errors.
- Variations stem from evolving single-precision floating-point arithmetic implementations in operating systems.
Conclusions:
- Computing platform variations introduce significant variability in complex neuroimaging analyses.
- Single-precision floating-point arithmetic and evolving OS implementations are key contributors to these discrepancies.
- Improving numerical precision and reviewing pipeline stability are essential steps toward reproducible neuroimaging research.
