Spatial and Temporal Features to Distinguish Noise in fMRI

Vanessa Sochat

Overview

• Background

• Methods • Evaluation

• Results • Conclusion

• Paperification

Questions for the Peanut Gallery

1. Standard developed by 3 or 4 researchers – ideas?

2. Stronger paper if “released” with an actual tool?

3. Other feedback?

Paperize

Background

• Goal: diagnosis and subtyping of neuropsychiatric disorder from functional brain networks

• Landscape: exploding publicly available data call for data-driven, automatic methods.

• ICA: derives independent signal: networks and noise

• Current Task: separate noise from networks

Background

Overview of Method

Methods

time

DATA

135 Spatial + 111 Temporal Features

• 53 resting BOLD functional

magnetic resonance imaging

data-sets

• 24 Healthy Control / 29

Schizophrenia

Preprocess Functional

MRI Data

motion correction

segmentation

normalization

smoothing

Extract Components with

ICA

N = 1518 components:

component = spatial maps + timecourse

patterns of noise and brain networks

Define Features of

Components

Create Standards of

Component Types

Machine Learning to Predict

Component Type Based on

Features

Creating the “Gold Standard”

Methods

1. Select data and save labels

2. Select next / previous

3. Label component

4. Spatial map

5. Temporal timecourse

6. Distribution of signal (FFT)

For the following component types:

Methods

N = 818 (700) N=46 (1472) N=40 (1478

N=67 (1451) N=48 (1470)

A B C

D E A.Noise

B.Eyeballs

C.Head motion

D.White matter

E.Parieto-occipital

F.cortex

ALL NOISE EYEBALLS HEAD MOTION

WHITE MATTER PARIETO OCCIPITAL CORTEX

Percent total activation skull

% activation voxels LR symmetric

Power band 0.0 to 0.008 Hz

% total activation MNI152 all

edges

Avg distance btw 10 local max

Hpsd bin 2 freq0 to pi 0.038312

Percent total activation in CSF

Four lag auto correlation

Hpsd bin 3 freq0 to pi 0.076624

three lag auto correlation

Percent total activation in CSF

Percent total activation MNI152 edges

kurtosis measure outlier-prone ts

Caudate R

Percent total activation skull

Caudate L

Percent activation in eyeballs

Paracentral Lobule L

Spatial Entropy of IC distribution

Percent total activation spinal cord

Percent total activation ventricles

Percent total activation in WM

Caudate R

Cingulum Post R

Caudate L

Thalamus L, Thalamus R

Max cluster size 10 local max region

growing thresh 2.5 < .5 overlap

power band 0.05 to 0.1 Hz

Cingulum Post L

Occipital Sup R

Occipital Mid R

Parietal Sup L

Percentage activation voxels LR

symmetric

Occipital Mid L

Occipital Sup L

Parietal Inf L

Parietal Sup R

Occipital Inf L

Parietal Inf R

LASSO most highly weighted features predict component types

Results

Percent activation in eyeballs

Spatial Entropy of IC distribution

Avg distance btw 10 local max

Skewness of IC distribution

% activation voxels LR symmetric

Rectus L

Olfactory R

Percent total activation in WM

Perfect total activation in GM

Evaluation of classifiers is very good!

Evaluation

Lasso L1 constrained linear regression selects features to distinguish component types (N=1518)

with cross validation accuracies of .8689, .9834, .9808, .9675, and .9695 respectively.

A B C D E

A. Noise

B. Eyeballs

C. Head motion

D. White matter

E. Parieto-occipital

cortex

Conclusion

Noisy components can be computationally defined using

spatial and temporal features to allow for automatic filtration

of large functional MRI databases, an improvement over

current manual methods.

Selected features provide researchers with semantic

understanding of component types.

This method can be extended to identify features of

functional brain networks.

Conclusion

Goals for Writing into Paper:

• Journal Choice: NeuroImage

• Time: before end of Spring Quarter (before Quals)

• My first paper!

Paperize

Feedback and Problems

•“Vanessa Standard” is not good enough

• Need to demonstrate utility

Paperize

How I'm Fixing it

• Separating SZ and HC data

• Build model with HC, test on SZ

• New datasets: • Age and gender matched

• Different institution, same scanner • Different institution, different scanner • Standard developed by 3 or 4 researchers

• Original SZ model successful, or model built with subset of new data successful on second portion.

Paperize

My progress so far:

1. Separating SZ and HC data

2. Build model with HC 3. Test on SZ

4. New dataset: Different institution, same scanner 5. Different institution, different scanner 6. Standard developed by 3 or 4 researchers 7. Original SZ model successful, or model built with

subset of new data successful on second portion.

Paperize

Thank you! :)

n x m n x n n x m

INDEPENDENT COMPONENT ANALYSIS