Educational session: Executing BIDS-Apps on large datasets and the *NiPreps* framework

name: title
layout: true
class: center
---
layout: false
count: false

.middle.center[
# Executing BIDS-Apps on large datasets and the *NiPreps* framework

### Oscar Esteban
#### CHUV | Lausanne University Hospital

###### [www.nipreps.org/assets/OHBM2022](https://www.nipreps.org/assets/OHBM2022)
]

---
layout: false
count: false

.middle.center[
# Executing BIDS-Apps on large datasets and the *NiPreps* framework

### Oscar Esteban
#### CHUV | Lausanne University Hospital

###### [www.nipreps.org/assets/OHBM2022](https://www.nipreps.org/assets/OHBM2022)
]

---
name: newsection
layout: true

.perma-sidebar[

<a rel="license" href="http://creativecommons.org/licenses/by/4.0/"><img alt="Creative Commons License" style="border-width:0; height: 20px; padding-top: 6px;" src="https://i.creativecommons.org/l/by/4.0/88x31.png" /></a>
 Executing BIDS-Apps on large datasets and the NiPreps framework

]

---

# Outline

.boxed-content[
.distribute.large[
* Managing large datasets - Datalad 101

* BIDS, and why BIDS?

* BIDS-Apps

* The case of fMRIPrep

* The NeuroImaging PREProcessing toolS (NiPreps)
]
]

---

# Datalad 101

.boxed-content[
.distribute.large[
* Datalad (amongst many features) allows versioning of data

* Datalad is based on Git and git-annex

* It is critical to maintain large datasets with versioning

* Enables access to lots of open datasets

* More: [datalad.org](https://datalad.org)

]
]

---

# BIDS - [bids.neuroimaging.io](https://bids.neuroimaging.io)

Specification: [bids-specification.readthedocs.io](https://bids-specification.readthedocs.io)

---

# Example dataset: AOMIC-PIOP2

<a href="https://openneuro.org/datasets/ds002790/versions/2.0.0">
<img src="../aomic-openneuro.png" width="70%" /> 
https://openneuro.org/datasets/ds002790/versions/2.0.0
</a>

---

# Example dataset: AOMIC-PIOP2

<a href="https://openneuro.org/datasets/ds002790/versions/2.0.0">
<img src="../aomic-openneuro-bids-valid.png" width="90%" /> 
https://openneuro.org/datasets/ds002790/versions/2.0.0
</a>

---

# Example dataset: AOMIC-PIOP2

<a href="https://openneuro.org/datasets/ds002790/versions/2.0.0">
<img src="../aomic-openneuro-files.png" width="45%" /> 
https://openneuro.org/datasets/ds002790/versions/2.0.0
</a>

---

# Example dataset: AOMIC-PIOP2

---

# AOMIC-PIOP2 (`datalad get`)

---

# BIDS-Apps

.boxed-content[
.distribute.large[
* Uniform I/O
    * Read in BIDS
    * Write out BIDS-Derivatives
    * Command-line interface specification:
      ```Shell
      bids-app-name /data/ds002790 /data/ds002790/derivatives/bids-app_1.0.0 participant
      ```

* Containerization: Docker & Singularity

* Index of apps: https://bids-apps.neuroimaging.io/apps/
]
]

---

# BIDS-Apps: subject-wise parallelization

<a href="https://doi.org/10.1371/journal.pcbi.1005209">
<img src="../journal.pcbi.1005209.g002.png" width="90%" /> 
(Gorgolewski et al., 2017)
</a>

---

<a href="https://doi.org/10.1038/s41592-018-0235-4">
<img src="../fmriprep-workflow-final.svg" width="73%" /> 
</a>
fMRIPrep (<a href="https://doi.org/10.1038/s41592-018-0235-4">Esteban et al., 2019</a>)

---

# Executing with Docker

.boxed-content[
.distribute.large[
* Docker tips & guidelines: https://nipreps.org/apps/docker

* Running the example dataset:
  ```Shell
  $ docker run -ti --rm \
        -v /data/datasets/ds002790:/data:ro \
        -v /data/derivatives/ds002790:/derivatives \
        nipreps/fmriprep:21.0.2 \
        /data /derivatives/fmriprep-21.0.2 \
        participant
  ```
]
]

---

# Important considerations

.boxed-content[
.middle[
* Interim results (“work directory”):
 * Preserve them if you will need to re-run the data (**caching**)
 * Make it a fast filesystem, if possible (e.g., `/tmp/` on an SSD system)
 * It will require a lot of space (10 GB per subject and task) - which sets a conflict with the previous item.
* Docker / Singularity:
 * FreeSurfer license file - **mount into containers!**
 * (docker only) write as user to preempt outputs have root ownership
* Memory and parallelization:
 * The more parallel, the more memory consumption
 * Recommendation: run single-subject processes (BIDS-Apps)
 * WATCH OUT FOR RACE CONDITIONS!
 * Find a good balance between ``--omp-nthreads`` and ``--nprocs``
* Inputs:
 * BIDS Validation: dataset MUST be valid
 * Tip for large databases: use ``--bids-database-dir``
 * If using datalad, data should have been pulled down with datalad get

]
]

---

# Executing with Docker

.large[
```Shell
$ docker run -ti --rm \
    -v /data/datasets/ds002790:/data:ro \
    -v /data/derivatives/ds002790:/derivatives \
    -v $HOME/.cache/pybids:/cache/bids \
    -v $HOME/tmp/fmriprep:/work \
    -u $( id -u ):$( id -g ) \
    nipreps/fmriprep:21.0.2 \
    /data /derivatives/fmriprep-21.0.2 \
    participant \
    --participant-label 0021 \
    --omp-nthreads 8 --nprocs 16 \
    --work /work/sub-0021 -vv --skip-bids-validation \
    --bids-database-dir /cache/bids/ds002790
```
]

---

# Running *fMRIPrep* with Docker

---

## The individual report

<video controls="controls" width="70%"
 name="Video Name" src="../fmriprep-report.mov"></video>

---

# Run-to-run reproducibility

.boxed-content[
.distribute.large[
*(... or the lack thereof)*

* *fMRIPrep* does not guarantee full run reproducibility with `--omp-nthreads` larger than 1
    * Which means, **disabling within-node paralellization**.

* *fMRIPrep* writes out a config file
    * stores **seeds** for all random number generators it has access to
]
]

---

# "Analysis-grade" data

.larger[
The *NeuroImaging PREProcessing toolS* (*[NiPreps](https://nipreps.org).org*) augment scanners to produce *analysis-grade* data (= **directly consumable by analyses**)
]

.pull-left[

***Analysis-grade* data** is an analogy to the concept of "*sushi-grade (or [sashimi-grade](https://en.wikipedia.org/wiki/Sashimi)) fish*" in that both are:

.large[**minimally preprocessed**,]

and

.large[**safe to consume** directly.]
]

.pull-right[
<img align="right" style='margin-right: 50px' src="https://1.bp.blogspot.com/-Osh4H4WXka0/WlMJmVgkZTI/AAAAAAAAEMY/GynUzSomJ-EBiyqv2m-maiOyKSM7SOmNACLcBGAs/s400/yellowfin%2Btuna%2Bsteaks%2Bnutrition.jpg" />
]

---

<img src="../nipreps-chart.png" width="63%" /> 
NiPreps (<a href="https://doi.org/10.31219/osf.io/ujxp6">Esteban et al., 2020</a>)

---

template: title
layout: false
.middle[

<img src="https://github.com/oesteban/fmriprep/raw/f4c7a9804be26c912b24ef4dccba54bdd72fa1fd/docs/_static/fmriprep-21.0.0.svg" width="95%" />

]

---

template: newsection
layout: false

# *TemplateFlow* | Archive

<img src="../torw2020/assets/templateflow-datatypes.png" width="78%" /> 
(<a href="https://doi.org/10.1101/2021.02.10.430678">Ciric et al., 2022</a>)

---

# *TemplateFlow* | Client

---

.boxed-content[
.distribute[
### *NiPreps* is a framework for the development of preprocessing workflows

* Principled design, with BIDS as a strategic component
* Leveraging existing, widely used software
* Using NiPype as a foundation

### Why preprocessing?

* We propose to consider preprocessing as part of the image acquisition and reconstruction
* When setting the boundaries that way, it seems sensible to pursue some standardization in the preprocessing:
  * Less experimental degrees of freedom for the researcher
  * Researchers can focus on the analysis
  * More homogeneous data at the output (e.g., for machine learning)
* How:
  * Transparency is key to success: individual reports and documentation (open source is implicit).
  * Best engineering practices (e.g., containers and CI/CD)

### Challenges

* Testing / Validation!
]
]

---

# Conclusion

<img src="https://media.springernature.com/full/springer-static/image/art%3A10.1038%2Fs41596-020-0327-3/MediaObjects/41596_2020_327_Fig1_HTML.png?as=png" width="90%" /> 
(<a href="https://doi.org/10.1038/s41596-020-0327-3">Esteban et al., 2020</a>)

---

.boxed-content[
.middle.center[
# Thanks!

### Questions?

### Join the NiPreps Hackathon!
]
]