Skip to content
Exploration of Montecarlo Arithmetic analysis for evaluating and correcting the stability of tools in neuroimaging
Jupyter Notebook HTML
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
paper @ 5f971b3

Comparing Noise Simulation Models for Evaluating Stability of Modeling in Neuroimaging

Gregory Kiar1,2, Tristan Glatard3

11McGill BME, 2MNI, 3Concordia CS


  • The numerical reproducibility of neuroimaging analyses is challenged by observations comparing results obtained with small data perturbations.
  • In this paper, we measure the compare the uncertainty of neuroimaging pipelines using various forms of stability analysis:
    • one-voxel (epsilon) perturbations,
    • Monte-Carlo arithmetic analyses,
    • (time permitting) Operating system
  • We evaluate the stability of structural connectome generation by performing simulations on two modelling + tracing algorithms commonly used in diffusion MRI:
    • Dipy 6-component tensor and EuDX (deterministic tracing)
    • Dipy ODF and probabilistic tracing.


  • Diffusion MRI NKI-RS dataset
  • Preprocessing done with FSL defaults and not evaulated here


  • Lower-order processing:
    • 6 component tensor model ("known" condition)
    • EuDX deterministic tracing
  • Higher-order processing:
    • ODF
    • Probalistic tracing

Noise injection

  • One-voxel noise:
    • Various noise strengths and injection locations
  • MCA
    • Recompiling cython libs with Verificarlo
    • (time permitting) Recompile blas+lapack with Verificarlo
    • Test both recommended precision bits (24, 53)
  • OS
    • Centos 5
    • Centos 6
    • Ubuntu 16

Prospective Results

  • Figure 1
    • Violin plots of an output norm for each noise setting and both session- and subject-differences
  • Figure 2
    • Compare distributions and mean results for multiple repeitions of equiv. 1-voxel noise and MCA executions
  • Figure 3
    • Comparison of computational efficiency for 1-voxel and MCA methods


You can’t perform that action at this time.