on the large scale of studying dynamics with meg: lessons learned from the human connectome project
TRANSCRIPT
On the large scale of studying dynamics with MEG: Lessons learned from the Human Connectome Project
Robert Oostenveldtogether with the HCP MEG team
Donders Institute, Radboud University, Nijmegen, NLKarolinska Institute, Stockholm, SE
AcknowledgementsLinda Larson-PriorAvi SnyderStefania Della PennaVittorio PizzellaGian-Luca RomaniLaura MazzettiFrancesco de PasqualeFred PriorRichard Bucholz
Francesco di PompeoGiorgos MichalareasJan-Mathijs Schoffelen
Tracy NolanMatthew KelseyAbbas Babajani-FeremiJeff StoutTera Kiser
David van Essen and ~100 researchers in the WU-Minn HCP consortium
All subjects that participated
Human Connectome Project
2010: start of project2012: start of acquisition2015: complete MEG release2015: first large outreach workshop (100 people)2016: second workshop (>100 people)2017: complete MRI release2017: third workshop
Human Connectome Project - Goals IStudy a large population:
1,200 healthy adults300 twin pairs and their non-twin siblings
Use cutting-edge neuroimaging methods3T Skyra MRI, customized gradient7T MRI (UMinn, 200 subjects)MEG/EEG (SLU, 100 subjects)dMRI/tractographyResting-fMRI & rMEGTask-fMRI & tMEG
Extensive behavioral testing Blood samples for genotyping
Nelson et al. (Neuron, 2010)
Human Connectome Project - Goals IIMake raw+processed data freely availableProvide a user-friendly informatics platform
Data mining, discovery scienceRelate brain connectivity to individual
capabilitiesLink to heritability, genetics
Provide a baseline for future studies of brain disorders
Autism, schizophrenia, ADHD, etc.Neurosurgical planning and interpretation
Paradigms with MEG – subset of fMRI tasks
3x 5 mins Resting state0 and 2-back Working Memory taskStory/Math taskMotor task (with EMG)
Overview of MEG analysis
Adding dynamics to the Human Connectome Project with MEG. Neuroimage (2013).
PREPROCESSING
CONNECTIVITY
TEMPORAL AND SPECTRAL ESTIMATES
SOURCE ESTIMATION
Overview of MEG analysis
PREPROCESSING
RESTING STATE DATA TASK DATA
CONNECTIVITY
Adding dynamics to the Human Connectome Project with MEG. Neuroimage (2013).
ANAT
OMY
PREPROCESSING
RESTING STATE DATA
TASK DATA
CONNECTIVITY
ConnectomeDBfunctional data in
shared anatomical representation tabular data ?new CIFTI data format building on NIFTI and GIFTI
HCP data organisation (>20TB, now 900 subjects)
subjectid/unprocessed/MEGsubjectid/unprocessed/3T/…subjectid/unprocessed/7T/…
subjectid/MEG/anatomysubjectid/MEG/Restinsubjectid/MEG/Motortsubjectid/MEG/WrkMemsubjectid/MEG/StoryM
subjectid/T1w/…subjectid/Diffusion/…subjectid/MNINonlinear/…
RnoisePnoiseRestin1Restin2Restin3WrkMem1…
datacheckbaddataicaclassrmegpreprocicamnebfblpenvicaimagcoh…
resultsquality controlfigures
consistent naming scheme and organization of data
Processing – over and over again
MEG data acquired from 2012-2014~10 pilot subjects, ~120 study subjects
Pipeline development from 2012-2015
WashU HCP centre compute cluster & storageCompiled MATLAB code with explicit release
versions (0.1, 0.2, … 1.0, 2.0, 2.1, 3.0)Each result file complemented with automatic
“log” file with version detailsVerification of version details prior to release
proper data management, instead of ad-hoc processing
MEG analysis strategies
Currently released results represent the state-of-the-art (at the time of data processing) from the MEG perspective
Comparisons with other HCP results should shed light on the validity of the assumptions in the MEG analyses
Application of new analysis techniques will certainly improve the connectivity patterns
Why sharing of both data and software?
Do analysis on our (MEG/fMRI/DTI) connectomes
Different choices for parameters in pipelinesDifferent source estimation methods
models are based on imperfect/unknown assumptions Different connectivity estimation methods
sensitivity to different features and better robustnessMore elaborate channel level analyses
certain electrophysiological features do not require spatial interpretation
Reanalysis in other software packagesSee also poster Mo-008, HCP with MNE-Python, Engeman et al.
Beyond HCP
Existing Neuroimaging (fMRI) databases
Primary and minimally processed dataGeneral: fMRIDC [RIP], 1000FC, openfMRITargeted: ADNI, NKI-Rockland; ConnectomeDB, CCF
Extensively processed (published) resultsCoordinate-based: Brainmap, SumsDB [RIP], NeurosythStatistical maps: Neurovault
MEG databases
Wakeman and Hensonraw data from 16 subjects (subset from 19)
Open MEG Archive (OMEGA)raw data from 127 subjects
Human Connectome Projectraw and processed data from 95 subjects
(subset from 1200)Others
…
Data sharing is not easy
Resting state data is easier to share than functional task data
Examples available that demonstrate how to share corresponding anatomical data
Brain Imaging Data Structure: draft now available for MEG (BIDS-MEG, see poster Mo-P011)
http://bids.neuroimaging.io
Data sharing is not easy
Dealing with very rich human dataMEG electrophysiologyMR imagingselect populationgeneticsbehavioural testsblood tests
Solutionde-identify all datade-face the imaging datashare coregistered dataprovide age in 5-year binsshare sensitive data only under special agreement
Data sharing is not easy
Data use agreements may have to applyNot all data may be available (e.g. coreg)
“Free as in free speech, not as in free beer” (Richard Stalman)
Wikipedia: The English adjective free is commonly used in one of two meanings: "for zero price" (gratis) and "with little or no restriction" (libre). This ambiguity of free can cause issues where the distinction is important, as it often is in dealing with laws concerning the use of information, such as copyright and patents.
Analysis pipeline sharing is not easy
Dependencies on external components, compute platforms, commercial software, …
Requires structuring and documenting codeAllowing others to use your code also allows them
to test novel ideas or extend your code
HCP focus on reproducibility – all code shared!
Lessons learned from HCP
There are no short term incentives in making MEG research more reproducible
The community needs to take long-term responsibility in pushing the field forward, solidifying methods and translating them to basic and clinical applications
Sharing (of results, data and pipelines) beyond the primary publications is a vital step forward
Resting-state MEG connectivity studies are among the easiest to take this step forward
Lessons learned from HCP
Sharing of expertise is indispensible
You may have data, we have expertise and we are willing to share!
[email protected] contact the HCP team
On the large scale of studying dynamics with MEG: Lessons learned from the Human Connectome Project
Robert Oostenveldtogether with the HCP MEG team
Donders Institute, Radboud University, Nijmegen, NLKarolinska Institute, Stockholm, SE
http://www.humanconnectome.org