London, UK21 May 2012

Janaina Mourao-Miranda,Machine Learning and Neuroimaging Lab,University College London, UK

Pattern Recognition for Neuroimaging Toolbox

Outline• Motivation• PRoNTo Framework• Future Developments

PRoNToMachine Learning

communityNeuroscience and

Clinical Neuroscience communities

sMRI

fMRI

time

Advantages of Pattern Recognition Analysis

Accounts for the spatial correlation of the data (multivariate)• fMRI data are multivariate by nature.• Can yield greater sensitivity than conventional analysis.

Enable classification/prediction of individual subjects/scans• ‘Mind-reading’ or decoding applications• Clinical applications

•Does the pattern of activation in brain regions A, B and C encode information about a variable of interest?

•Can we classify groups of subjects (e.g. patients and healthy controls) based on brain scans?

•Which features lead to the best discrimination between a group of patients and a group of controls?

•Can we predict continuous measures (e.g. age, performance, etc) from brain scans?

•etc

Questions investigated with pattern recognition analysis

Conventional neuroimaging studies (encoding):

- Investigate which regions of the brain are involved in a perceptual or cognitive task.

- Probe for individual voxels whether the average activity during one condition is significantly different from the average activity during a base line condition.

- Mapping: cognitive state -> fMRI data

Pattern Recognition (‘mind-reading’) approaches (decoding):

- Address the problem of classifying the cognitive state of a subject based on the fMRI data.

- Based on a training set identify a distributed pattern of activation that discriminates two cognitive states and uses this pattern to make prediction for new data.

- Mapping: fMRI data -> cognitive state

77

OUR GOAL“develop a toolbox based on machine learning

techniques for the analysis of neuroimaging data”

PRoNTo

BUT “free”, matlab based, compatible with SPM,

easy to use (with GUI),multiple modalities (fMRI/sMRI/PET/betas),

various machines, modular code, easy to contribute

Existing softwareFreely available packages for machine learning modeling of neuroimaging data: • 3dsvm plugin for AFNI (LaConte et al., 2005)• the Matlab MVPA toolbox for fMRI data (Detre et al., 2006)• PyMVPA (Hanke et al., 2009)• PROBID (http://www.brainmap.co.uk/probid.htm)

• 3dsvm, the Matlab MVPA toolbox for fMRI data and PyMVPA require (advanced) programming skills, can not be directly integrated into the main neuroimaging analysis pipelines, e.g. SPM (http://www.fil.ion.ucl.ac.uk/spm/).

• PROBID is optimized for groups classification (i.e. patients vs. healthy controls) and therefore does not easily enables single subject analysis or flexible cross-validation framework.

Pascal Harvest Project• “The main idea behind Harvest is: put together in a room a

team for long enough to produce an innovative software for a real application. PASCAL2 will pick up the bills”.

Requirements:• Piece of software as their main objective• Training component.• International team.

Title: PRoNTo (Pattern Recognition for Neuroimaging Toolbox)

Coordinator: Dr. Janaina Mourao-Miranda

Participants: Dr. Christophe Phillips (Cyclotron Research Centre, University of Liège,

Belgium)Dr. John Ashburner (Wellcome Trust Centre for Neuroimaging, UCL)Dr. Jane Rondina (Department of Neuroimaging, KCL)Dr. Andre Marquand (Department of Neuroimaging, KCL)Dr. Maria Joao Rosa (Wellcome Trust Centre for Neuroimaging, UCL)Dr. Jonas Richiardi (Ecole Polytechnique Fédérale de Lausanne,

Switzerland)Ms. Jessica Schrouff (Cyclotron Research Centre, University of Liège,

Belgium)Dr. Carton Chu (National Institute of Mental Health (NIMH), NIH, USA )

Hosting site: UCL, Computer Science Department, London, UK Website: http://www.mlnl.cs.ucl.ac.uk/pronto/

Toolbox based on pattern recognition techniques for the analysis of neuroimaging data:

- “free” (MATLAB based) - easy to use (with GUIs) - easy to contribute (modular

code) - multiple modalities:

- fMRI/sMRI/PET/betas- compatible with SPM

For usersUser Interface

Easy to useClick buttons

Matlab Batch

Very efficientCan be saved and copiedCompatible with SPM

Functions

User-specific analysis

More programming skills required

1313

User point of viewUser Interface Matlab Batch

For developersUser Interface

Specific GUIBatch system

Script

Machine learning

FeaturesKernelModel

TrainingValidation

Machines

Classification (SVM, GPC, RF)

Regression (KRR,RVR)W

rapp

er

1515

Developer point of view

prt_machine

Structure containing:-Data/Kernel-Labels-EtcStructure containing:-Function name (machine)-Arguments

Structure containing:-Predictions-Coefficients/Weights-etc

prt_machine_krrprt_machine_gpml

prt_machine_svm_bin

Machine Library (classification and regression models)

Structure containing:-Data/Kernel-Labels-Etc

Structure containing:-Predictions-Coefficients/Weights

Future developments

• improved fMRI data handling (detrending & hrf estimation)

• feature selection (GP based, RFE,…)

• more machines (provided by Machine Learning community)

• etc

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

CreditsThe development of PRoNTo was possible with the financial and logistic support of: - PASCAL Harvest Programme (http://www.pascal-network.org/) - the Department of Computer Science, University College London (http://www.cs.ucl.ac.uk); - the Wellcome Trust; - PASCAL2 (http://www.pascal-network.org/); - the Fonds de la Recherche Scientifique-FNRS, Belgium (http://www.fnrs.be); - The Foundation for Science and Technology, Portugal (http://www.fct.pt); - Swiss National Science Foundation (PP00P2-123438) and Center for Biomedical Imaging (CIBM) of the EPFL and Universities and Hospitals of Lausanne and Geneva.

PRoNTo a team work!