Click to edit Master title style

Pattern Recognition for Neuroimaging Toolbox: PRoNTo

Jessica Schrouff

PRNI 2018

June 14th

NUS, Singapore

Click to edit Master title styleOutline

• PRoNTo’s goals and history

• PRoNTo for users General framework

Modules

• PRoNTo for developers Implementation

Scripting examples

• PRoNTo v3

• Hands-on session with Fabio Ferreira

6/14/2018 J. Schrouff - PRNI 2018 2

Click to edit Master title styleAims

6/14/2018 J. Schrouff - PRNI 2018 3

PRoNToMachine Learning

communityNeuroscience and

Clinical Neuroscience communities

sMRI

fMRI

Click to edit Master title stylePRoNTo’s history

• Pascal Harvest project: August 2011, v1.0

• Team from different backgrounds and institutions

• Releases since: v1.1, v2.0, v2.1, v3.0 to come

6/14/2018 J. Schrouff - PRNI 2018 4

Click to edit Master title stylePRoNTo features

• Matlab based

• Open-source

• Interfaced: GUI and batch

• SPM compatible

• Multiple classification and regression algorithms

• Includes developments from the team member’s work

Available at: http://www.mlnl.cs.ucl.ac.uk/pronto/

6/14/2018 J. Schrouff - PRNI 2018 5

http://www.mlnl.cs.ucl.ac.uk/pronto/

Click to edit Master title styleQuestions it addresses

• Classification Can we use brain scans to diagnose psychiatric or neurological disorders?

Can we decode from the brain scans information about the stimuli being processed?

• Regression Can we predict clinical scores from brain scans?

6/14/2018 J. Schrouff - PRNI 2018 6

Click to edit Master title style

PRoNTo for users

Jessica Schrouff

PRNI 2018

June 14th

NUS, Singapore

Click to edit Master title styleFramework

6/14/2018 J. Schrouff - PRNI 2018 8

WeightsRun modelSpecify model

Feature set

Data & Design

• Groups• Subjects/scans• Modalities• Conditions

• Extract features• Build kernel(s)• Detrend• Scaling

• Targets/classes• Cross-Validation• Hyper-parameter

optimization• Machines • Data operations

• Estimate model• Assess performance• Assess significance

• Feature weights• Kernel contribution

DataDesign1st level mask

Atlas2nd level mask (ROI)

PRT.mat

Review Design Review Kernel Display performance Display weightsReview Model, CV

Click to edit Master title styleData & Design

• User inputs:

Groups

Subjects

Modalities

Design

Covariates

Targets

Masks

6/14/2018 J. Schrouff - PRNI 2018 9

PRT.mat

Data & Design

DataDesign1st level mask

Review Design

Click to edit Master title styleFeature set

• Reads through all images in modality

• Selects features for kernel(s)

• v2: combines kernels

6/14/2018 J. Schrouff - PRNI 2018 10

Feature set

Atlas2nd level mask (ROI)

Review Kernel

Click to edit Master title styleSpecify model

• Define the question:

feature set

classes / samples

machine and parameters

cross-validation (inner & outer)

data operations

6/14/2018 J. Schrouff - PRNI 2018 11

Specify model

Review Model, CV

Click to edit Master title styleMachines

• Classification:

SVM (LIBSVM)

Gaussian Processes (GPML)

L1-MKL (simpleMKL)

• Regression:

RVR

KRR

Gaussian Processes (GPML)

L1-MKL (simpleMKL)

6/14/2018 J. Schrouff - PRNI 2018 12

Click to edit Master title style

4/14/2018 J. Schrouff - Course 2018 13

Standard approaches:• LOSO• LOBO• LORO• LOSCO• k-fold CV

Cross-validation

Flexible CV schemes allowed

Click to edit Master title styleRun model

• Estimates the model

• Batch: runs permutations

6/14/2018 J. Schrouff - PRNI 2018 14

Run model

Display performance

Measuresofperformanceforregression

Correla*on

Coefficientofdetermina*on

MeanSquaredError(MSE)

NormalisedMSE

Click to edit Master title styleCompute weights

• Build voxel weights

• Maps back to original image

• ROI contributions:

average from atlas (post-hoc)

kernel contribution (MKL)

6/14/2018 J. Schrouff - PRNI 2018 15

Click to edit Master title styleThank you!

Questions?

6/14/2018 J. Schrouff - PRNI 2018 16

Click to edit Master title style

PRoNTo for developers

Jessica Schrouff

PRNI 2018

June 14th

NUS, Singapore

Click to edit Master title stylePRoNTo for developers

• Tool box: use PRoNTo as a basis for developments

• Avoids to re-code e.g. cross-validation when new data/model

• Established framework: less prone to errors, double dipping

6/14/2018 J. Schrouff - PRNI 2018 18

Click to edit Master title stylePRoNTo architecture

• Modular

• User interfaces

• Core functions

• All written in Matlab with occasional mex interfaces

6/14/2018 J. Schrouff - PRNI 2018 19

WeightsRun model

Specify model

Feature set

Data & Design

Click to edit Master title stylePRoNTo architecture

6/14/2018 J. Schrouff - PRNI 2018 20

GUI Batch

Core functions

Click to edit Master title stylePRT structure

• Similar to SPM structure

• Saved in folder as ‘PRT.mat’

• PRT has one or 2 fields per module

• Load and dig:

6/14/2018 J. Schrouff - PRNI 2018 21

Click to edit Master title styleCore functions

• One function per module

• Ideal for scripting:

• Not for Data & Design step

• Best for running models

6/14/2018 J. Schrouff - PRNI 2018 22

Click to edit Master title styleScripting example - 1

• Simulated data: one feature set per SNR and sparsity

• 19 * 15 = 285 models

• Power using PRoNTo:Home script

Batch to create feature set

Load PRT

Use a pre-defined model as basis

Replace the feature set

Run and extract performance

6/14/2018 J. Schrouff - PRNI 2018 23

Click to edit Master title styleScripting example - 2

• Permutations can be lengthy

• Can run code on cluster

• Different strategies: one job per xx permutations

using parfor

spmd

Batch or prt_permutations

Careful with graphical/workspace output

6/14/2018 J. Schrouff - PRNI 2018 24

Click to edit Master title stylePRoNTo machines

• PRoNTo interfaces libraries

6/14/2018 J. Schrouff - PRNI 2018 25

Wrapper functions Third-party libraries

Click to edit Master title styleNew machine

• Can interface your classification/regression algorithm

6/14/2018 J. Schrouff - PRNI 2018 26

esvr_train.mesvr_predict.m

Re-arrange inputs/outputs and arguments

Click to edit Master title styleNew machine: further considerations

• labels: 1 for class 1, 2 for class 2, …

• Hyper-parameter optimization: careful with string arguments

• Use ‘custom machine’

• Compiled machine?

• Weight computation: Default: linear kernel

If not: need to specify weight function

6/14/2018 J. Schrouff - PRNI 2018 27

Click to edit Master title styleContributing

• Yes!

• Next releases in Github

• Need to interface: GUI and batch

• Backwards compatibility

• Matlab backwards compatibility and toolbox codes

• Coming soon: developer’s manual

6/14/2018 J. Schrouff - PRNI 2018 28

Click to edit Master title styleThank you!

Questions?

6/14/2018 J. Schrouff - PRNI 2018 29

Click to edit Master title style

PRoNTo v3.0

Jessica Schrouff

PRNI 2018

June 14th

NUS, Singapore

Click to edit Master title styleNew input formats

• SPM MEEG

6/14/2018 J. Schrouff - PRNI 2018 31

Click to edit Master title styleNew input formats

• .mat

6/14/2018 J. Schrouff - PRNI 2018 32

Click to edit Master title styleNew functionalities

• Non-kernel methods

• Combining different data formats (e.g. EEG and fMRI)

• Multi-Task Learning

• Test model module: share PRT and weight maps

• and much more …

6/14/2018 J. Schrouff - PRNI 2018 33

Click to edit Master title styleThank you!

Suggestions?

6/14/2018 J. Schrouff - PRNI 2018 34

Click to edit Master title style

Hands-on session

Jessica Schrouff

PRNI 2018

June 14th

NUS, Singapore

Click to edit Master title styleDemo

• OASIS data set

• 50 demented – 50 non-demented

• Confounds: age and gender (boolean)

• Two features extracted: grey and white matter

• Goal: classify demented from non-demented

6/14/2018 J. Schrouff - PRNI 2018 36