þÿ

[ ]

1st INCF Workshop on NeuroImaging Database Integration

August 30-31, 2007
International Neuroinformatics Coordinating Facility Secretariat
Stockholm, Sweden

Authors

Lars Forsberg and Per Roland

Scientific Organizer

Per Roland, Karolinska Institutet, Stockholm, Sweden

Workshop Participants

Katrin Amunts, Forschungzentrum Jülich GmbH, Jülich, Germany
Jean-Pierre Changeux, Institut Pasteur, Paris, France
Rodney Douglas, UNI ETH, Zurich, Switzerland
David C Van Essen, Washington University, St. Louis, USA
Lars Forsberg, Karolinska Institutet, Stockholm, Sweden (Rapporteur)
Jesper Fredriksson, KTH, Stockholm, Sweden
Albert H Gjedde, Aarhus University Hospital, Aarhus, Denmark
Kazuhisa Ichikawa, Kanazawa Institute of Technology, Kanazawa, Japan
David N Kennedy, Massachusetts Gen Hospital, Charlestown, USA
Torkel Klingberg, Karolinska University Hospital, Stockholm, Sweden
Per Roland, Karolinska Institutet, Stockholm, Sweden
Ulla Ruotsalainen, Tampere University of Technology, Tampere, Finland
Ryoji Suzuki, Kanazawa Institute of Technology, Kanazawa, Japan
Jack Van Horn, University of California Los Angeles, Los Angeles, USA
Karl Zilles, Düsseldorf University, Düsseldorf, Germany

Supported by the EU Special Support Action INCF, the INCF Central Fund and the Swedish
Foundation for Strategic Research

Nature Precedings : doi:10.1038/npre.2008.1781.1 : Posted 8 Apr 2008

[ ]

1. Executive Summary

The goal of this workshop was to map existing neuroimaging databases, particularly those

containing primary data, and identify mechanisms that could facilitate integrated use of such

databases, including interconnections between databases and data sharing.

1.1 Interconnections

The workshop group recommended that INCF should promote federation of databases by

coordinating and integrating neuroimaging resources. Such a federation can be achieved by

creating a portal from which different databases can interconnect to each other. The group rec-

ommended that interconnections between databases should be facilitated through the design

and development of an INCF portal. This portal should give users access to all the databases

through one interface, and also enable the users to go from one database to another and search

for similar contents. Still to be determined is how to compare data between different databases

through modeling.

1.2 Data sharing

A federation can also be achieved through data sharing between two or more databases, where

the data from one database is integrated into another database. With this integration, data can

be preprocessed and presented through the context of other databases. The group thought that

INCF should promote data sharing and recommended that this should start with a test case:

data sharing between the two fMRI databases fMRIDC and NeuroGenerator. The knowledge

and tools developed in this project should be reused for future data-sharing projects.

With the knowledge gained from this project, INCF should help to advance the issue of de-

scribing neuroimaging data. This can be achieved by collaborating with major journals to de-

velop guidelines and standards for reporting metadata related to experimental methods. Fur-

thermore, INCF should encourage investigators to submit data and tools to a clearinghouse.

1.3 Future possibilities

The workshop group recommended that INCF should attempt to integrate databases of dif-

ferent modalities, including structural MRI, EEG, MEG, microstructural, (e.g., cytoarchitec-

tonic) and receptor databases. Anyone should be able to access the databases, not only the

member states of INCF. User login and password should not be required. The search features

for the different databases should be documented on the portal with tutorials and examples.

Nature Precedings : doi:10.1038/npre.2008.1781.1 : Posted 8 Apr 2008

[ ]

2. Introduction

The very first functional neuroimaging database with online

access on the Internet was created in 199 [Fox, Lancaster,

199]. Known as BrainMap, it currently contains 00 experi-

ments. Mosaic was still the most commonly used web browser

at the time BrainMap was created and the size of hard drives

was measured in MB rather than GB. Thus, given the infancy of

the World Wide Web, the most realistic content to store in this

database was the stereotaxic coordinates that were published

in the scientific papers. Since most studies publish their results

in the Talairach coordinate system, the results from different

studies should be comparable. An MRI-based stereotactic atlas

from 0 subjects was published in 199, to be followed by

the MNI 0 template in 199 [Evans et. al. 199, Evans et. al.

199]. However, the number of different stereotaxic standard

brains continues to increase and although they should all be in

the Talairach coordinate system, their coordinate systems are

all significantly different from the Talairach system and from

one another. Therefore, to compare studies with different stan-

dard brains, a transformation needs to be applied directly to

the coordinates [Brett et. al. 00; Lancaster et al., 007] or

through a surface-based atlas as an intermediate [Van Essen,

00].

A functional ROI (region of interest) in a functional experi-

ment may contain thousands of coordinates, which describe its

extension and form in D space. Publications often only report

the statistical peak coordinate of the ROI, together with its size,

but the extension and form of the ROI are not published. Par-

ticularly when the ROI is actually multiple functional regions

merged together in the thresholding of the statistical image,

the peak coordinate from the most significant ROI is typically

the only information reported. If a database also covered the

form of the functional ROI, it would have to store the D im-

age format which describes the full extension of the functional

region.

In 1999, the European Computerised Human Brain Database

(ECHBD) endeavored to store the findings from a functional

study as D images, rather than just the peak coordinates [Ro-

land et. al. 1999]. It also contained cytoarchitectonic measure-

ments in the same stereotaxic space as the functional images.

The database was unable to accept data from different standard

templates, so the users were asked to change the format of their

data into the ECHBD standard brain format before submitting

it to the database. In practice, it was impossible to ensure that

the data was correctly transformed, and the data would be diffi-

cult to compare. Additionally, different experiments were pro-

cessed in different ways, with varying spatial filter sizes, dif-

ferent thresholding of statistical images, etc. One way to create

a homogenous statistical database is thus to analyze the data

through the same processing pipeline, using the same methods

and the same standard brain. This would require the users to

submit the raw data and enough experimental metadata to ana-

lyze the study in a standardized pipeline.

NeuroGenerator is the successor to ECHBD, and it has ad-

dressed the problem of heterogeneous datasets by collecting

the raw data and processing it through a common processing

workflow [Roland et al, 2001]. The result is a statistical data-

base where the user can select which filter size and threshold

level to work with. All the data has been transformed into the

same standard brain using the same transformation method.

The project was created in 2000 and the first database was sent

out to users in 00. Due to the size of the raw data, Neuro-

Generator only allows people to access the statistical data. This

access is via an open-source D visualization and query tool,

available for Linux and Mac OS X. The database currently

contains 7 studies from 9 subjects as well as cytoarchitec-

tonic probabilistic maps for anatomical reference [Amunts and

Zilles, 00].

Nature Precedings : doi:10.1038/npre.2008.1781.1 : Posted 8 Apr 2008

[ 7 ]

The fMRI Data Center (fMRIDC) is a public repository of

peer-reviewed fMRI studies [Van Horn, et. al. 001]. This

repository was established in 1999. It currently contains 1

complete neuroimaging studies representing thousands of indi-

vidual subjects, and hundreds of thousand of individual fMRI

and MR structural volumes, as well as accompanying meta-

data from published research articles. Users may request study

data and it will be delivered to them on media and via digital

download in cases where the overall size of the data is not pro-

hibitive. The fMRIDC thereby allows the researcher not only

to replicate the findings of the original study, but also to apply

other methods and potentially obtain new or alternative find-

ings, and even use the data in training and education.

The number of studies represented in NeuroGenerator and

fMRIDC is much smaller than those in BrainMap, but the size

of these databases is much larger because each study can be

anywhere from 10-0 GB in size. BrainMap on the other hand,

strives to record the reported statistical local maxima of ac-

tivation from published articles along with the details of ex-

perimental design, thus facilitating meta-analysis across neu-

roimaging studies. The final derived statistical database from

NeuroGenerator is a few 100 MBs in size. Depending on the

meta-study, the researcher can benefit from all these different

kinds of databases.

This report summarizes the INCF workshop on Neuroimag-

ing Database Integration, whose goal is to map existing neu-

roimaging databases containing particularly primary data, e.g.,

NeuroGenerator and fMRIDC, and discuss the benefits and is-

sues regarding data sharing across databases. One major topic

was whether INCF should contribute to facilitating data shar-

ing between fMRIDC and NeuroGenerator.

3. Concepts

In order to get a common ground for the following discussions,

it is necessary to establish some basic concepts.

Integration

Integration of databases can either be to establish an interoper-

ability between them or to share data between databases.

Interoperability

Interoperability refers to the ability of different database sys-

tems to work together (interoperate) using an established pro-

tocol for interaction.

Data sharing

Data sharing can refer either to sharing data through a clearing-

house or to copying data between databases.

Database

A database is a data repository stored in a structured way, with

the possibility to query the repository for information.

Clearinghouse

A neuroimaging clearinghouse is a distribution center which

collects and distributes tools and data.

Pipeline

A processing pipeline is an ordered set of tools forming a di-

rected workflow used to process data.

Nature Precedings : doi:10.1038/npre.2008.1781.1 : Posted 8 Apr 2008

[ ]

4. Federation of Databases

The number of databases will continue to grow. Integrating

neuroimaging databases will be made possible through a fed-

eration of databases. This integration will make the data more

useful to researchers, allowing them, for example, to go from

one database resource to another and see which other databases

have common fields (e.g., anatomical regions). For this to be-

come a reality there must be a portal through which researchers

can access the different database resources and a way for them

to easily jump between the databases. Data sharing is an alter-

native example, that involves preprocessing the content from

one database and presenting it in a different way in another

database.

The participants at the neuroimaging workshop agreed that

data sharing is highly desirable, as it has the potential to ac-

celerate research progress in neuroimaging, analogous to the

dramatic advances that have occurred in genomics and other

areas of bioinformatics. However, there were some concerns

as to how to do this in practice while simultaneously maintain-

ing the quality. More specifically, how does one compare data

between different databases? There has to be a link between

different observations and theories through modeling. In this

process, it must be specified what kind of data to be compared

and confront the collection of data through modeling. We have

to understand what types of questions we are going to address

through integration of databases. In the case of data sharing,

what kind of data should be exchanged and should there be a

minimum description of neuroimaging data to facilitate such

data exchange? Should we be more precise as to what kind

of data to compare and exchange from a theoretical point of

view? How do we find the balance between simplicity of ac-

cess (which would result in broad acceptance and usage by the

community) and high quality?

5. Recommendations

5.1 A Test Case--Data Sharing Between

NeuroGenerator and fMRIDC

The workshop gave rise to many questions, which is why the

participants recommend that INCF attempt an example case

study as a demonstration of how the scientific community

might benefit from data sharing between two databases. It was

felt among the workshop participants that the data sharing be-

tween NeuroGenerator and fMRIDC could serve as such a test

case if justified by a well-defined proposal with realistic objec-

tives. The role of INCF should be that of an honest broker, to

facilitate fulfillment of the agreement. INCF can also provide

technical support and resources, as long as the development

can be reused in the future and if the process can be generic to

other databases.

To get started, some example datasets would be exchanged in

order to demonstrate the value of data sharing between Neu-

roGenerator and fMRIDC. During this process, quality control

of the data is important so as to maintain the integrity of each

of the databases.

There were some discussions about the metadata needed to get

a precise insight of the conditions in an fMRI experiment, as

well as the issue of minimum description of neuroimaging (for

instance as an XML of RDF standard). For the data sharing

between NeuroGenerator and fMRIDC, the decision was to not

wait for an XML standard. The two databases already have

their own description of fMRI studies. The challenge here is to

establish a suitable exchange protocol between the two data-

bases and to define suitable benchmarks as to what would con-

stitute successful data sharing and integration in this test case.

5.2 Future Development

The proposed, as well as other, test cases, should help in iden-

tifying what a minimum description requires, so that it could

be reused in future data sharing, including other databases such

as Function BIRN. For future data collection, it is highly desir-

able to include a minimum description of neuroimaging data.

INCF could help advance the issue of describing neuroimag-

ing data and encourage people to submit data to an appropri-

ate database. To get this started, it was suggested that INCF

should coordinate with major journals to help in identifying

Nature Precedings : doi:10.1038/npre.2008.1781.1 : Posted 8 Apr 2008

[ 9 ]

relevant requirements. One problem in the field today is the

lack of a standard way to present the methods of a study in a

scientific paper. This lack of standardization makes it difficult

for researchers to compare different studies. A minimum for-

mat for describing neuroimaging and the statistical analyses

will reduce uncertainties as to how the results were obtained,

thus making it more likely that another study will be able to

reproduce the results. This would also be a way to indicate to

journals that the study meets appropriate standards. The Neu-

roimaging Informatics Tools and Resources Clearinghouse

(NITRC nitrc.org) may be useful in this regard.

Demonstrated at the workshop was the LONI-pipeline, whose

front end is a graphical workflow of different analysis mod-

ules. The processing backend can either be the same computer

or a cluster of computers. The LONI-pipeline description can

be used to present the statistical methods used in the publica-

tions. If data were submitted to a clearinghouse together with

the necessary processing tools and a LONI-pipeline descrip-

tion, anyone could reproduce the results using the same data.

By sharing computing resources, people might be more willing

to use the LONI-pipeline system and finally submit the data in

a standard format to a clearinghouse. This would increase the

quality of the data.

In practical terms, it is not possible to define all the metadata

needed to describe the conditions of an experiment in detail.

The complexity of new experiments is increasing as new meth-

ods are being used to analyze the data. A minimum descrip-

tion is not intended to cover all methods and paradigms of all

experiments in the way a fully-featured ontology would, but it

should nevertheless attempt to describe a major fraction of a

major subset of published studies (as opposed to all aspects of

all studies). If journals supported a consistent minimum infor-

mation framework as sanctioned by INCF, authors would have

specific guidelines about what information should be included

in their submitted scientific manuscripts, which would help

to improve methods reporting. Furthermore, if this informa-

tion were present, then text-mining approaches applied to the

content of the published article would be enriched and better

able to identify similarities between studies, etc. Minimal in-

formation provides several advantages for improving scientific

communication and INCF should play a role in fostering the

development of MIAMI-like lists of domain-specific metadata

classes of these purposes.

The SumsDB database of structural and functional neuroim-

aging data was also demonstrated at the workshop. SumsDB

contains a diverse set of surface-based and volume-based data

from humans, monkeys and rodents. The human data includes

over 1000 stereotaxic coordinates from more than 00 pub-

lished studies, along with extensive experimental metadata.

This data could potentially be federated with the aforemen-

tioned coordinate data in BrainMap, thereby capitalizing on

the complementary visualization and analysis methods associ-

ated with the two databases. Additionally, the volume-based

data in SumsDB could potentially be federated with the data

and analysis resources of fMRIDC and NeuroGenerator.

Federation of databases can be described as either data sharing

or interconnectivity between the databases. Comprehensive in-

terconnectivity requires a portal from which one can access all

the databases. One suggestion was to have a common search

interface on the INCF website instead of choosing a specific

database. It should also be possible to migrate between data-

bases through the INCF portal or through other portals such

as the Neuroscience Information Framework (NIF). To facili-

tate this process, it was recommended that databases should be

evaluated by INCF and best practices should be established,

possibly through collaboration between INCF, NIH, EU and

Japan [Suzuki et al., 007]. Federation of databases should

eventually cover not only fMRI databases, but also structural

MRI, EEG, MEG, and other modalities such as a microstruc-

tural (e.g., cytoarchitectonic) or receptor databases [Amunts

and Zilles, 00; Eickhoff et al., 00; Zilles et al., 00].

All the participants agreed that anyone should be able to access

the databases, not only those from the member states of INCF.

It should not even require a user login to access and query the

databases. To help people understand how the databases could

be used, there should be tutorials on the portal describing the

search features.

All in all, a useful and successful database integration will ben-

efit researchers greatly, creating better ease of use and improv-

ing the current standards in the field.

Nature Precedings : doi:10.1038/npre.2008.1781.1 : Posted 8 Apr 2008

[ 10 ]

Appendix: Workshop Program

August 30:

09.00 - 09.30

Introduction and orientation (Bjaalie and Roland)

09.30 - 12.00

Scientific presentations and discussions

Karl Zilles

Brain maps for functional imaging: From histology to probabilistic maps

Katrin Amunts

Brain maps for functional imaging: From probabilistic maps to meta-analysis

Jean Pierre Changeux

Modeling access to consciousness and its consequences for brain imaging

Ryoji Suzuki

Neuroimaging study and platform in Japan: Overview

Kazuhisa Ichikawa

Neuroimaging study and platform in Japan: Linguistic brain functions, integrative

analysis, and NIMG-PF

12.00 - 13.00

Lunch

13.00 - 18.00

Scientific presentations and discussions

Jack van Horn

Is it time for a minimum data framework for neuroimaging reporting, exchange,

and archiving?

Ulla Ruotsalainen

Databases of functional brain PET images

David C van Essen

Mining structural and functional neuroimaging data using the SumsDB Database

and WebCaret visualization

Rodney Douglas

The challenge of EM image storage and analysis in large volume (m^3)

reconstructions

Albert H Gjedde

Functionally integrative neuroscience: Expanding the frontiers of brain function

David Kennedy

The neuroimaging informatics tools and resource clearinghouse: A new

knowledge environment for fMRI research

Lars Forsberg

Finding co-activation patterns with PCA: A meta-analysis study using the

NeuroGenerator database

Jesper Fredriksson

Mining the NeuroGenerator database

Torkel Klingberg

Imaging of brain development and models of brain development

19.00 -

Dinner and further discussion

August 31:

09.00 - 12.00

Discussions and draft report

12.00 - 13.00

Lunch

Each presentation was scheduled for 0 minutes, including questions.

Nature Precedings : doi:10.1038/npre.2008.1781.1 : Posted 8 Apr 2008

[ 11 ]

References

Amunts, K., Zilles, K.: Atlases of the human brain: Tools for

functional neuroimaging. In: Neuroanatomical tract-tracing

-moleculesn Neurons, Systems (Zaborszky, L., Wouter-

lood, F.G., Lanciago; J.L., eds), pp. -0 (00)

M. Brett and I. S. Johnsrude and A. M. Owen, "The problem of

functional localization in the human brain", Nature Reviews

Neuroscience, Vol , -9, March 00.

Eickhoff, S., Stephan, K.E., Mohleberg, H., Grefles, C., Fink,

G.R., Amunts, K., Zilles, K.: A new SPM toolbox for com-

bining probabilistic cytoarchitectonic maps and functional

imaging data. NeuroImage , 1-1 (00)

A. C. Evans and D. L. Collins and B. Milner, "An MRI-based

stereotactic atlas from 0 young normal subjects", Journal

Soc. Neurosci. Abstr. 1: 0, 199

A. C. Evans and D. L. Collins and S. R. Mills and E. D. Brown

and R. L. Kelly and T. M. Peters, "D statistical neuro-

anatomical models from 0 MRI volumes", Proc. IEEE-

Nuclear Science Symposium and Medical Imaging Confer-

ence, 11-117, 199.

P. T. Fox and J. L. Lancaster, "Neuroscience on the net", Sci-

ence , 99-99, 199.

Lancaster, J.L. and Tordesillas-Gutierrez D. and Martinez M.

and Salinas F. and Evans A. and Zilles K. and Mazziotta,

J.C. and Fox P.T. (007) Bias between MNI and Talairach

coordinates analyzed using the ICBM-1 brain template.

Hum Brain Mapp. Jan 1; [Epub ahead of print].

P. E. Roland and J. Fredriksson and P. Svensson and K. Amunts

and C. Cavada and R. Hari and A. Cowey and F. Crivello and

S. Geyer and G. Kostopoulos and B. Mazoyer and D. Pop-

pelwell and A. Schleicher and T. Schormann and M. Seppa

and H. Uylings and K. de Vos and K. Zilles, "ECHBD: A

database for functional-structural and functional-functional

relations in neuroimaging", Poster 1, Fifth International

Conference on Functional Mapping of the Human Brain,

1999.

P. E. Roland and G. Svensson and T. Lindeberg and T. Risch

and P. Baumann and A. Dehmel and J Fredriksson and H.

Halldórsson and L. Forsberg and J. Young and K. Zilles,

"A database generator for human brain imaging", Trends in

neurosciences. :10, -, Oct. 001.

Suzuki, R., Niki, K., Fujimaki, N., Masaki, S., Ichikawa, K.

and Usui, S., "Neuro-Imaging Platform for Neuoinformat-

ics", 1

International Conference on Neural Information

Processing (ICONIP007), Kitakyushu, 007, TMB-

D. C. Van Essen, A population-average, landmark- and surface-

based (PALS) atlas of human cerebral cortex. NeuroImage.

28, -, 00.

J. D. Van Horn and J. S. Grethe and P. Kostelec et. al, "The

functional magnetic resonance imaging data center (fM-

RIDC): the challenges and rewards of large-scale databas-

ing of neuroimaging studies", Philosophical Transactions of

the Royal Society of London Series B Biological Sciences,

vol. , no. 11, 1-19, 001.

Zilles, K., Schleicher, A., Palomero-Gallagher, N., Amunts, K.:

Quantitative analysis of cyto- and receptoarchitecture of the

human brain, pp7-0. In; Brain Mapping: The Methods,

edition (A.W.Toga and J.C.Mazziota, eds) Academin

Press (00)

Nature Precedings : doi:10.1038/npre.2008.1781.1 : Posted 8 Apr 2008