The Nuisance of Denoising fMRI

James Kent (@SudoNeuroSci)

Dr. Michelle Voss

HBC Lab

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

Ellie + Vader Talk

(image sources)

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

Technical

Academic

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

fMRI Data is Noisy

fMRI = functional Magnetic Resonance Imaging

Noise Sources

  • Thermal
  • Head Motion
  • Spin History
  • Heartbeat
  • Heart Rate Variability
  • Respiration
  • Respiration Variability
  • CO2 Variability
  • Spontaneous Activity
  • HRF Model Error
  • Scanner instability

(image sources)

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

Independent Components Analysis (Temporal)

True Sources

Observed Sources

ICA Sources

(image sources)

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

Independent Components Analysis (Spatial)




 



 



 

(image sources)

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

Example Time Series

(Left Motor)

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

Independent Components Analysis (fMRI)

Handbook of Functional MRI Data Analysis 1st Edition

Spatial: 100,000+ data points

Time:

~100+ data points

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

Automatic Removal of Motion Artifacts (AROMA)

(Pruim et al. 2015)

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

Signal & Noise Components

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

Example Voxel Correlation

No Noise Removal

"Aggressive" Removal

"Non-Aggressive" Removal

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

Review: ICA-AROMA

  • Finds independent sources in data
  • Identifies noise sources
  • Is Aggressive or Non-Aggressive better?

(image sources)

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

Code Contribution: ICA-AROMA

FMRIPREP

2017-05-02

2017-05-[03-10]

2017-05-26

- 

2017-06-28

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

FMRIPrep: The Paper

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

FMRIPrep: My Next Contribution

  • Currently - "Non-Aggressive" denoising THEN use other confounds
  • Future - "Non-Aggressive" denoising SIMULTANEOUS with other confounds.
Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

The Problem:

Modular Preprocessing

fmriprep:

pull request

2018-02-02

fmriprep:

issue 817

2018-02-03

fmriprep:

issue 1049

2018-04-11

twitter:

the tipping point

2018-09-04

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

The Problem:

Modular Preprocessing

Favorite Drink?

Favorite Activity?

Favorite Treat?

(image sources)

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

The Problem:

Modular Preprocessing

Original Data

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

The Problem:

Modular Preprocessing

  1. y' = ' y ~ confound1 '
  2. y_resid = y - y'
  3. y'' = ' y_resid ~ confound2 '
  4. y'_resid = y_resid - y''
Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

The Problem:

Modular Preprocessing

  1. y' = ' y ~ confound2 '
  2. y_resid = y - y'
  3. y'' = ' y_resid ~ confound1 '
  4. y'_resid = y_resid - y''
Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

The Problem:

Modular Preprocessing

  1. y' = ' y ~ confound1 + confound2'
  2. y_resid = y - y'
Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

Review:

Modular Preprocessing

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk
  • Don't Do Modular Preprocessing!
    • Model everything you want to model
  • Online Collaborations = Fun + Informative

Takeaways

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk
  • Understanding data is hard
  • There are great communities online
  • You can get recognized for your contributions 
  • Collaborations make science better!

Thanks!

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk

FMRIPREP

Dr. Chris Gorgolewski

&

Dr. Chris Markiewicz

HBC Lab

  • Dr. Michelle Voss
  • Matt Sodoma
  • Adriana Rivera-Dompenciel

Thesis Committee

  • Dr. Michelle Voss
  • Dr. Jatin Vaidya
  • Dr. Vince Magnotta
  • Dr. Jan Wessel
  • Dr. Eliot Hazeltine

Questions?

Slides: slides.com/jameskent/2018-seminar
Code: github.com/jdkent/2018-Seminar-Talk
Made with Slides.com