3. Getting lost in your data
• MRI has been used to study
the human brain for over 20
years.
• Despite similarities in
experimental designs and data
types each researcher tends to
organize and describe their
data in their own way.
http://www.nature.com/news/brain-imaging -fmri-2-0-1.10365
4. Getting lost in your data
Heterogeneity in data description practices causes:
• problems in sharing data (even within the same lab),
• unnecessary manual metadata input when running
processing pipelines,
• no way to automatically validate completeness of a
given dataset,
• difficulties in combining data from multi-center studies.
5. Brain Imaging Data Structure
Brain Imaging Data Structure (BIDS) is a new
way for standardizing, describing and
organizing results of a human neuroimaging
experiment.
6. Who is it for?
1. Lab PIs. It will make handing over one dataset from one
student/postdocto another easy.
2. Workflow developers. It’s easier to write pipelines expecting
a particular file organization.
3. Database curators. Accepting one dataset format will make
curation easier.
7. Principles behind BIDS
1. Adoption is crucial.
2. Don’t reinvent the wheel.
3. Some meta data is better than no metadata
4. Don’t rely on external software (databases) or
complicated file formats (RDF).
5. Aim to capture 80% of experiments but give the
remaining 20% space to extend the standard.
8. Implementation
1. Some metadata is encoded in the folder structure.
2. Some metadata is replicated in the file name for simplicity.
3. Use of tab separated files for tabular data.
4. Use of NIFTI files for imaging data.
5. Use of JSON files for dictionary type metadata.
6. Use of legacy text file formats for b vectors/values and
physiological data.
7. Make certain folder hierarchy levels optional for simplicity.
8. Allows for arbitrary files not covered by the spec to be
included in any way the researchers deem appropriate.
9. Why TSV?
1. Simple text format with wide software support.
2. Strings with commas do not need to be escaped by
quotation marks.
10. Why NiFTI?
Pros:
1. Widest support from software packages.
2. Designed for neuroimaging.
Cons:
1. Poor metadata support.
2. Memory mapped random access to compressed
NifTI is hard to implement.
11. Why JSON?
1. Simple text (you can use notepad to edit).
2. Wide support from different programming languages.
3. Simpler than XML, but almost as powerful.
4. Extensible with linked data.
12. BIDS features
1. Handles multiple sessions and runs
2. Supports sparse acquisition (via slice timing)
3. Supports contiguous acquisition covariates (breathing, cardiac
etc.)
4. Supports multiple field map formats
5. Supports multiple types of anatomical scans
6. Supports function MRI: both task based and resting state.
7. Supports diffusions data (together with corresponding bvec, bval
files)
8. Supports behavioral variables on the level of subjects
(demographics), sessions, and runs.
17. Keys to success
1. Make the community involved in the design process.
2. Provide a good validation tool (browser based!).
3. Build tools/workflows/pipelines that make adopting BIDS
worthwhile (AA, Nipype, C-PAC etc.)
4. Get support from databases (LORIS, COINS, SciTran,
OpenfMRI, XNAT, etc.)
23. Convincing people to share
data is hard
1. Publication as an incentive (data papers – Gorgolewski et al.
2013)
2. Sharing only statistical derivatives (NeuroVault – Gorgolewski
et al. 2014)
25. Convincing people to share
data is hard
1. Publication as an incentive (data papers – Gorgolewski et al.
2013)
2. Sharing only statistical derivatives (NeuroVault – Gorgolewski
et al. 2014)
3. Journal policies (see PloS One, F1000Research Scientific Data)
26. Data sharing fears
1. Fear of being scooped
2. Fear of someone finding a mistake
3. Misconceptions about the ownership of the
data
27.
28. Stanford | Center for Reproducible
Neurscience
Analyzing for reproducibility
reproducibility.stanford.edu
• Automated quality control reporting
• Data analysis service
• Using cutting edge, robust and well tested methods
• Leveraging supercomputer power not accessible to
most labs
• Quantify reproducibility by out of sample prediction
estimates
• “Glass box” – in depth documentation describing all data
analysis steps
29. Stanford | Center for Reproducible
Neurscience
Analyzing for reproducibility
reproducibility.stanford.edu
• The service is completely free of charge
• Under one condition: the data will be publicly available
after a grace period
30. Stanford | Center for Reproducible
Neurscience
Analyzing for reproducibility
reproducibility.stanford.edu
• CRN will:
• Make more data publicly available
• Improve access to best methods and algorithms
(including yours!)
• Enable automatic data exploration and hypothesis
generation
• Foster the culture of looking at out of sample
predictions and effect sizes