Functional Biomedical Informatics Research Network (FBIRN) UK E-Science Workshop

advertisement
Functional Biomedical
Informatics Research
Network (FBIRN)
David Keator
FBIRN NeuroInformatics Working Group Chair
University of California, Irvine
UK E-Science Workshop
October 2, 2006
FBIRN Goals
 Develop multi-site functional neuroimaging
tools.
 Develop the capability to analyze, as a single
data set, data acquired from multiple sites
using tools developed from multiple sites
 Develop a federated data management
system to support these multi-site fMRI
studies
Schizophrenia as a model
 Schizophrenia
• Brain illness with subtle structural and
functional changes
• Confounds of age/illness duration/treatment
effects
• Active area of imaging research with many
competing theories and approaches
 Progress hampered by
• Inconsistent data & lack of replications
• Noncomparable imaging techniques
• Small and diverse patient populations
 This applies to many areas of
study
FMRI variance and correction
How big a problem is it?
Traveling Humans Study (Phase I)
HIPAA
HIPAA
HIPAA
HIPAA
HIPAA
Subjects traveled around the country to
be scanned at all FBIRN sites
Unique dataset: Subject x site interactions
can be measured for the first time
Measuring Subject and Site Effects
ROI – Top 10% of Activated Voxels
Variance Source
Auditory
Hand
Visual
Subject
18.8
18.3
21.8
Site
43.0
21.0
43.8
Day
0.0
0.0
0.1
Run
0.4
0.1
0.1
Subject X Site
3.6
14.6
10.5
Subject X Site+
20.7
35.2
20.0
Residual
1.5
4.2
1.5
Calibration Keys
 Pre-data collection scanner QA
 Data collection
• Effect of different stimuli and response devices
 Experimental design:
• Cognitive task sensitivity, robustness, reliability
• Non-neuronal measures of BOLD signal
 Post-data collection corrections
• Use of the breatholding (BH), smoothness, others
 Data sharing across sites
• Data format issues, orientations, provenance, etc.
• Data analysis issues
Phantom QA methods
ave
nave
SFNR
Phantom QA methods
Time Series Stability
Informatics Tools for Tracking Scanners
https://portal.nbirn.net/BIRN/cgi-bin/DataGrid/MRIs/publicMRIStability.cgi
Scanner characterization
Friedman et al., Neuroimage, In press
Calibration methods: Smoothing
Smoothing to a common level reduces intersite effects
SMOOTHED
UNSMOOTHED
Site: HARV
MINN
IOWA
NMEX
.200
.075
-.075
-.200
Calibration Methods: Non-neuronal measures
of BOLD signal
 Scaling by the breathhold response increases group
effect size
Group
Differences:
Uncorrected
Corrected
Differences between young and old subjects in an fMRI task, before
and after correcting for BOLD differences in a separate breathhold
task
Thomason et al., Human Brain Mapping, 2006
Calibration methods: Grey Matter SFNR
Calibrating BOLD signal by SFNR in grey matter
reduces intersite effects
Intersite effects in 1.5T scanners in eight ROIs, before and after
correction
Friedman et al., Neuroimage, in press
Impact of Calibration Methods
Cohen’s f
ANOVA Observed Effect Size
Remaining
challenge
None
Smooth
Smooth, BH
Calibrate
Smooth, BH
Calibrate, BH Screen
Phase I Traveling Calibration Subject Dataset
Tools delivered since FBIRN inception in 2003
 Datasets
• Phase I, 10-site traveling subjects dataset
• MRI QA datasets
 Data collection
• Phantoms MRI calibration and QA protocol
• E-prime programs used in data collection
 Data management:
• XCEDE XML schema
 Tools for reading and writing XCEDE XML image wrappers
•
•
•
•
Standardized, general purpose directory hierarchy
Data management and “upload” scripts
Human Imaging Database schema/Graphical User Interface Code (HID)
Clinical Assessment Layout Manager/General Assessment Manager Engine
(CALM)
• FSL Image Processing Scripts (FIPS)
http://www.nbirn.net/downloads/index.shtm
Informatics/Analysis Infrastructure
Result Images and XML
wrapper in Data Grid
FMRI Images
•Automated image upload to
Data Grid/HID for sharing
fMRI Scanner
Proprietary Scanner
Formats,
DICOM, NIFTI
Data Grid
(Local)
HID(s)
(Local)
Clinical Data
•Computer aided scale input via
clinical data entry interface
FIPS: FSL Image
Processing Scripts
FIPS
Results
Results with standard descriptions
in HID (i.e. data provenance)
Multi-Site User Query
•Data storage
conventions
•XML data description
(XCEDE)
•Image and clinical data
held locally
•De-identification for
public sharing
Architecture Overview
Subject
Management
CALM/GAME
Assessments
Multi-Site
Query
XCEDE
Services
Clinical /
Demographics
Study
Protocols
Study
Data
Web Application
Core
Hierarchy
Schema
Oracle
Database
PostgreSQL
File System
Data
Data Grid
HID: Community Database Development
UCSD: Lead Development
UCI: Lead Development
MGH: Oracle Specifics
UCLA: PostgreSQL Specifics
UNM: Clinical Measures
UIowa: Performance
Duke: XML
Data Entry: HID+GUI
Clinical Assessment Layout Manager
(CALM)
Data Management: CALM & GAME
Multi-Database Query
Detailed Query Information
FBIRN Data Federation
UMN
HID
p2
p1
p2
p1
p2
Stanford p1
HID
UCLA
p2
HID
UCI
HID
UI
HID
p1
p2
UCSD
HID
p1
UNM
HID
= Data Integration Environment
= PostgreSQL test site
= Phase 1 / Phase 2 data
p1
p2
Duke: 67
BWH: 22
MGH: 12
UCLA: 54
UCSD: 6
UCI: 71
UNM: 57
UI: 64
UMN: 57
Yale: 66
p1
MGH
BWH HID
HID
Yale
p2 HID
p2
p1
419 Subject
Visits
3174 Subject
Assessments
p2
p1
Duke
HID
Phase II Study: Image Data Volume
 21,038 raw image files per subject
9.0
FBIRN Shared Data Files
Number of Files (Millions)
8.5
 2.4 GB of raw image data per
subject
 25 GB to 40 GB of processed
image data per subject (depending
on hypotheses tested)
8.0
7.5
7.0
6.5
6.0
5.5
5.0
July 05 Aug 05 Sept 05 Oct 05 Nov 05 Dec 05 Jan 06
 10 million slices of functional imaging data in Phase II
 7 Terabytes of image data for all of the Phase II analyses
(conservative estimate of 25 GB/subject)
Automated Weekly Data Management
XML Based Meta-Data Format
 The XML-Based Clinical and Experimental Data
Exchange (XCEDE) XML schema provides an
extensive metadata hierarchy for describing and
documenting research and clinical human imaging
studies.
XCEDE XML-Aware Components Image Format Independence
INPUTS
DICOM
Analyze7.5
add XML
PROCESSING
STEP
GE Pfile
add XML add XML
XML-Aware
Applications
(e.g., FIPS,
BIRN QA)
Format X
add XML
•Readable by
humans
•Expandable
•More efficient and
less prone to
errors
OUTPUTS
FIPS Analysis,
QA plots, etc.
SPM XCEDE XML Toolbox
 SPM toolbox to capture the
results from activation maps
using the XML activation
schema
 The toolbox supports both
SPM99 and SPM2 statistical
structures
 The toolbox has been used to
capture PET and fMRI
analysis results and the
associated analysis model
specifications.
Keator et. al., NeuroInformatics 2006.
FSL Image Processing Scripts (FIPS)
 Front end interface for automated image processing
and analysis
Personnel Time Savings
 Open source
Without FIPS
With FIPS
200
subjects
5 subjects
0
200
400
600
Hours
800
1000
Query Atlas
The Query Atlas project provides a
platform for the development of
scientific understanding of
experimental research results in the
context of neuroanatomy.
Interactive 3D and Web-based
knowledge management tool that is
coupled to the BIRN federated
database and links to other data
management systems from
neuroscience and other fields.
BIRN Tools Download
 http://www.nbirn.net/downloads/index.shtm
Download