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.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
code
figs
paper @ 5f971b3
.gitignore
.gitmodules
LICENSE
README.md

README.md

Comparing Noise Simulation Models for Evaluating Stability of Modeling in Neuroimaging

Gregory Kiar1,2, Tristan Glatard3

11McGill BME, 2MNI, 3Concordia CS

Introduction

  • 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.

Methods

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

Modeling

  • 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

References

You can’t perform that action at this time.