Large-scale functional MRI study on a production grid

https://doi.org/10.1016/j.future.2009.07.014Get rights and content

Abstract

Functional magnetic resonance imaging (fMRI) analysis is usually carried out with standard software packages (e.g., FSL and SPM) implementing the General Linear Model (GLM) computation. Yet, the validity of an analysis may still largely depend on the parameterization of those tools, which has, however, received little attention from researchers. In this paper we study the influence of three of those parameters, namely (i) the size of the spatial smoothing kernel, (ii) the hemodynamic response function delay and (iii) the degrees of freedom of the fMRI-to-anatomical scan registration. In addition, two different values of acquisition parameters (echo times) are compared. The study is performed on a data set of 11 subjects, sweeping a significant range of parameters. It involves almost one CPU year and produces 1.4 Terabytes of data. Thanks to a grid deployment of the FSL FEAT application, this compute and data intensive problem can be handled and the execution time is reduced to less than a week. Results suggest that optimal parameter values for detecting activation in the amygdalae deviate from the default typically adopted in such studies. Moreover, robust results indicate no significant difference between brain activation maps obtained with the two echo times.

Introduction

Functional magnetic resonance imaging (fMRI [1]) is a noninvasive method for detecting brain activation that is now applied extensively in neuroscience, neurosurgical planning and drug research. fMRI detects changes in oxyhaemoglobin/deoxyhaemoglobin ratio resulting from increased local perfusion in the brain following a rise in neural activity, the so-called blood oxygenation level dependent (BOLD) contrast. Functional MR data can be acquired rapidly and spatial resolution is high, so that large data sets are generated. fMRI data is usually analysed with software packages such as fMRIB Software Library (FSL) [2] and Statistical Parametric Mapping software (SPM) [3]. Although the user interface of these packages conceals much of the complexity of the image analysis process, the choice of parameters still plays an important role in fMRI analysis. Most researchers, however, perform the analysis using standard (default) parameter settings without questioning the role of these parameters, mostly due to practical reasons. The comparison of results obtained for different parameter settings requires a large amount of computing resources for analysis and data storage, but such advanced IT infrastructures are normally not available in a typical neuroscience environment. On the other hand, open grids [4] hold the promise to provide such high capacity computational resources in a shared and distributed model, and could be used to perform such parameter studies. Although the feasibility of grids has been demonstrated for several applications in medical imaging, their adoption in practice is still challenging and reports illustrating successful stories that go beyond the demonstrator level are still scarce.

In this paper we present a practical example of a parameter study in fMRI in which we varied three different parameters in a standard data pre-processing procedure and General Linear Model (GLM) analysis. We used for this example an emotional-provoking task which is known to activate the amygdalae: a brain area mainly involved in the processing of emotional reactions and memory. We also investigate the effect of one image acquisition parameter, namely the echo time. The analysis is performed on a production grid infrastructure, enabling this compute and data intensive problem to be tackled within a reasonable amount of time. This example illustrates the usefulness of such methodological experiments that employ massive computing resources to investigate the optimal parameter settings and the robustness of results obtained with fMRI.

The paper is organised as follows. Section 2 presents details of the fMRI problem and the designed parameter sweep experiment. The implementation of this experiment on a grid infrastructure is presented in Section 3. fMRI and application performance results are presented and discussed in Sections 4.1 fMRI results, 4.2 Grid performance. Section 6 presents conclusions of this study.

Section snippets

fMRI parameter study

In this study the acquired data consists of a functional MRI scan containing time series of 3D volumes, a high resolution MRI T1-weighted scan used for anatomical reference, and text files containing timing information about the adopted stimulus. The data is analysed with FSL, version 3.3, using the fMRI Expert Analysis Tool (FEAT), version 5.63. Each fMRI data set is first individually submitted to first-level analysis, which calculates brain activation maps as a result of a pipeline of image

Grid implementation

The designed experiment involves large computation effort. The individual analysis is the most compute intensive one due to pre-processing, registration and linear model computation. It takes about 45 min in a regular desktop. The group analysis is the most data intensive one, since it needs to access all the individual analysis results for averaging (see summary in Table 1). Parameter sweeps are costly; in particular, when the parameter space has several dimensions (three in this study),

fMRI results

Fig. 5 illustrates the mean Z-score indicating activation within the amygdalae using different degrees of freedom D for the fMRI-to-anatomical scan registration. No significant activation differences were observed, a pattern that is also observed for other smoothing kernel sizes. In particular, note that a registration with D=3 (translation only) produced similar results as obtained with D=9 and D=12. Given that both scans belong to the same subject and have been acquired during a single

Related work

To our knowledge, this is the first study to investigate the effects of varying multiple parameters during spatial pre-processing and statistical modelling of functional MRI data. Previous studies have addressed single parameters only, for example adding temporal and dispersion derivatives to the canonical hemodynamic response function to account for minor time shifts of the BOLD response [22], or adjusting spatial smoothing size to correct for between-scanner differences in multicenter studies 

Conclusion

In this paper, the benefit of performing parameter sweeps for methodological fMRI studies has been established by assessing amygdalae activation for different analysis parameter choices. Detecting activation in this brain area is in general cumbersome, emphasising the need of a good choice of parameters to obtain reliable results. Our initial results indicate that the optimal values deviate from the default parameters that are typically used in this type of analysis in FSL FEAT. In the future

Acknowledgements

We thank P.T. de Boer and K. Boulebiar for the development on the VBrowser front-end and plug-ins. We also warmly thank all the people operating behind mailing lists [email protected], [email protected]. This work was carried out in the VL-e project, supported by a BSIK grant from the Dutch Ministry of Education, Culture and Science (OC&W) and is part of the ICT innovation program of the Ministry of Economic Affairs (EZ).

Tristan Glatard obtained a Ph.D. in grid computing applied to medical image analysis from the University of Nice Sophia-Antipolis, France in 2007. He spent a year as a post-doc at the University of Amsterdam, Netherlands. He is now a researcher at Creatis-LRMN, Lyon, France.

References (30)

  • S. Erberich et al.

    funcLAB/G

    NeuroImage

    (2007)
  • W. Sudholt et al.

    Application of grid computing to parameter sweeps and optimizations in molecular modeling

    Future Generation Computer Systems

    (2005)
  • S. Smith

    Overview of fMRI analysis

    The British Journal of Radiology

    (2004)
  • K. Friston

    Statistical parametric mapping and other analysis of functional imaging data

  • I. Foster et al.

    The anatomy of the Grid: Enabling scalable virtual organizations

    International Journal of Supercomputer Applications

    (2001)

    • Cited by (7)

    View all citing articles on Scopus

    Tristan Glatard obtained a Ph.D. in grid computing applied to medical image analysis from the University of Nice Sophia-Antipolis, France in 2007. He spent a year as a post-doc at the University of Amsterdam, Netherlands. He is now a researcher at Creatis-LRMN, Lyon, France.

    Remi S. Soleman received the M.Sc. degree in clinical neuropsychology from the VU university Amsterdam, the Netherlands in 2008. He participated in several research projects with a focus on neuroimaging techniques, such as functional MRI (fMRI) and electroencephalography (EEG). Currently he is a Ph.D. student at the VU university medical center, Amsterdam, the Netherlands. His research interests include hormones, cognition and fMRI.

    Dick J. Veltman. After his graduation in Medicine (Free University, 1983, in Amsterdam), Dick was educated as a Psychiatrist/Psychotherapist at the Provinciaal Ziekenhuis Santpoort and the Valeriuskliniek Amsterdam. From 1989 he had worked at the psychiatric policlinic of the Valeriuskliniek and did experimental research on the role of psychological vs. somatic factors in panic attacks. In 1995 he finished his dissertation on this subject (supervisors: prof. R. van Dyck and prof. M.A. van den Hout).

    From 1997, he worked off and on as a visiting research fellow at the Functional Imaging Laboratory of the Wellcome Department of Cognitive Neurology in London (current director: prof. K.J. Friston). From 1997 to 2004 he was second to the PET-center of the Free University medical center VUmc (director: prof. A.A. Lammertsma), working on neuro-activation studies. In 2004, he became director of the methodological Brain Imaging program from the research school neurosciences at the VUmc.

    In 2005, he was appointed stragic professor neuro-imaging with a focus on anxiety disorders and depression at the Free University, and Extraordinary Professor neuro-imaging with a focus on addiction at the University of Amsterdam.

    Dick has been (co)author of more than 50 international and 30 Dutch scientific publications and chapters in books. Today, he is supervising about 20 Ph.D. students.

    Aart J. Nederveen graduated cum laude in Applied Physics at the Technical University of Delft in 1998. He then started working as a Ph.D. student at the department of Radiotherapy of the University Medical Center Utrecht on a project on image guided radiotherapy of prostate cancer, resulting in a Ph.D. thesis entitled Image guided position verification for intensity modulated radiotherapy of prostate cancer in 2002. In 2002 he started working as a postdoc at the NMR department of the Bijvoet Center in Utrecht. Here he participated in a European project (NMRQUAL) on the validation of protein structure determination by using NMR and contributed to the design of databases for structural genomics. He carried out calculations for structure determination of 500+ proteins on the Condor computer clusters of the Department of Computer Sciences at the University of Madison-Wisconsin. He also initiated a research project on the simulation of the internal dynamics and NMR relaxation parameters of ubiquitin in solution in the microsecond timescale. Calculations were carried out on the TERAS supercomputer at the national computer center SARA. In 2005 he joined the department of Radiology of the AMC in Amsterdam and was appointed as research coordinator of the 3T MRI research facility of the AMC/UvA. His main research interests focus on advanced MRI brain imaging (spin labeling, fMRI) and applications of grid computing within a clinical framework. Aart Nederveen has authored and co-authored over 30 peer-reviewed publications in the field of medical physics and bioinformatics and supervises several Ph.D. students.

    Sílvia D. Olabarriaga is assistant professor at the Academic Medical Center of the University of Amsterdam, NL, where she leads the e-BioScience group at the Bioinformatics Laboratory. She obtained a Ph.D. in Computer Science at the University of Amsterdam in 1999 with a thesis on the topic of medical imaging. Prior to joining the AMC in 2007, Dr. Olabarriaga held various positions in research, education and industry, including Post Doctoral fellowships at the University of Amsterdam and the University Medical Center Utrecht (Netherlands); assistant professorship at the Federal University of Rio Grande do Sul (Brazil); researcher at INESC Lisbon (Portugal) and software engineer at EDISA (Brazil). Since 2005 she participates in various Dutch e-Science initiatives for life sciences applications. Her current research interests include the design, development and deployment of grid-enabled infrastructures to enable and enhance biomedical research, in particular to facilitate the interaction between researchers and large distributed systems.

    View full text
    Copyright © 2009 Elsevier B.V. All rights reserved.