High performance/throughput computing workflow for a neuro-imaging application: Provenance and approaches



  • We describe a high performance/throughput computing approach for a full-brain bootstrapped analysis of Diffusion Tensor Imaging (DTI), with a targeted goal of robustly differentiating individuals with Parkinson's Disease (PD) from healthy adults without PD. Individual brains vary substantially in size and shape, and may even vary structurally (particularly in the case of brain disease). This variability poses significant challenges in extracting diagnostically relevant information from Magnetic Resonance (MR) imaging as brain structures in raw images are typically very poorly aligned. Moreover, these misalignments are poorly captured by simple alignment procedures (such as whole image 12-parameter affine procedures). Nonlinear warping procedures that are computationally expensive are often required. Subsequent to warping, intensive statistical bootstrapping procedures (also computationally expensive) may further be required for some purposes, such as generating classifiers. We show that distributing the preprocessing of the images using a compute cluster and running multiple preprocessings in parallel can substantially reduced the time required for the images to be ready for quality control and the final bootstrapped analysis. The proposed pipeline was very effective developing classifiers for individual prediction that are resilient in the face of inter-subject variability, while reducing the time required for the analysis from a few months or years to a few weeks.
  • Authors

    Digital Object Identifier (doi)

    International Standard Book Number (isbn) 13

  • 9783319639154
  • Start Page

  • 245
  • End Page

  • 256