A High-Throughput Pipeline Identifies Robust Connectomes But Troublesome Variability

Presented by Eric Bridgeford

Follow the slides: ericwb.me/lectures/ohbm/ohbm_ndmg.html

Multimodal MR Imaging (M3R) is Commonplace

  • Multimodal MRI (M3R): dMRI, fMRI, sMRI
  • Connectome: network of connections in the brain
  • Many open-access datasets:
    • Consortium for Reliability and Reproducibility
    • Human Connectome Project
    • Healthy Brain Network
  • No brain imaging biomarkers for clinical psychiatry

Processing and Analysis of Connectomes is Challenging

  • sample sizes of single studies are relatively small
  • data are heterogeneous within and across studies
  • existing pipelines are customized per dataset
  • lack of generative statistical network models


⇒ many failures to replicate

A Principled Approach for Connectomics

Principles

NDMG Processes dMRI & fMRI in Parallel

Pipeline

Discriminable Across 25+ Studies and 6000+ Scans

  • Discriminability: non-parametric generalization of ICC
  • $D = \mathbb{P}(x_{i, i'} < x_{i, j})$
  • Are within-subject connectomes more similar than across-subject connectomes?
Discriminability

What is Mega-Analysis?

  • Multiple Group Level Analysis, or mega-analysis, is when we have several unique groups we want to aggregate data over
  • Cohort = 6000+ connectomes across 25+ studies

Ipsilateral Connections Are Stronger than Contralateral in all Diffusion Connectomes

Hemisphere

Magnitudes of Effects Exhibit Large Heterogeneity Within and Across Sites

XSite Pvals

Conditioning on Basic Demographics and Study Fails to Eliminate Heterogeneity

Conditional Pvals

NDMG: An Efficient Pipeline for Scalable Connectomics

  • ndmg pipeline is open-source, parallel: ~1 hr/scan
  • Docker, singularity, BIDS App
  • >6000 of connectomes online at neurodata.io
  • extensive QA suite produced with each connectome

ndmg establishes a need for improved cross-site replicability

  • every dataset is highly discriminable
  • certain properties preserved across all scans
  • considerable heterogeneity within and across studies
  • conditioning on study and demographics fails to mitigate heterogeneity

Potential next Steps

  • data analysis improvements
    • probabilistic tractography
    • multiscale atlases
  • data acquisition improvements
    • deep phenotyping
    • data acquisition harmonization

Links



Acknowledgements

Joshua T. Vogelstein, Greg Kiar, Randal Burns, Xi-Nian Zuo, Vince Calhoun, Sephira Ryman, Rex Jung, Daniel Marguiles, Vikram Chandrashekhar, Disa Mehembere, Will Gray Roncal, Brian Caffo, Carey Priebe, Cameron Craddock, Michael Milham DARPA {XDATA, SIMPLEX, GRAPHS}; NSF {NeuroNex}; NIH; Kavli

Questions?