thesis defence copy1

THESIS DEFENCE COPY1THESIS DEFENCE COPY1sonilshrivastava22

SLIDE 1SLIDE 11

Insights into fMR data usingMachine Learning

Presented By:

SonilShrivastava(MT2014117)

Supervisor :

Prof Neelam Sinha

//www.emaze.com


SLIDE 2SLIDE 22

Introduction to fMRIIntroduction to fMRI◆

ObjectiveObjective◆

Dataset DescriptionDataset Description◆

Introduction to RFEIntroduction to RFE◆

Experimental DescriptionExperimental Description::

◆

RFE ImplementationRFE Implementation : :

•

StarPlus DataStarPlus Data◦

Probid DataProbid Data◦

◦

•

◦

OutlineROI AnalysisROI Analysis◦

Across SubjectsAcross SubjectsAnalysis -- 7 ROIsAnalysis -- 7 ROIs

◦


◦

ConclusionConclusion◆

ReferencesReferences◆

//www.emaze.com


SLIDE 3SLIDE 33

Introduction To fMRIIntroduction To fMRI◆

Objective◆

Dataset Description◆

Introduction -- RFE◆

Experimental Description:

◆

RFE Implementation :

•

StarPlus Data◦

Probid Data◦

◦

•

◦

OutlineROI Analysis◦

Subjects Analysis-- 7 ROIs

◦


◦

Conclusion◆

References◆

//www.emaze.com


SLIDE 4SLIDE 44

Introduction To fMRI

Acquisition of the Blood Oxygen LevelDependent (BOLD) in fMRI

fMRI -- functional Magnetic Resonanceimaging

◆

Provides information about thefunctioning of the human brain

◆

◆

Voxels and ROIs (Region of Interest).◆

voxels intensity --> f(x,y,z,t).◆

//www.emaze.com


SLIDE 5SLIDE 55

Introduction -- fMRI◆

ObjectiveObjective◆




◆


•

StarPlus Data◦

Probid Data◦

◦

•

◦



◦


◦

Conclusion◆

References◆

//www.emaze.com


SLIDE 6SLIDE 66

Objective

Mapping of the brain corresponding to a

cognitive state.

◆

Time stamp analysis.◆

Region of activation analysis.◆

Classification analysis across all the

subjects.

◆

//www.emaze.com


SLIDE 7SLIDE 77


Objective◆




◆


•

StarPlus Data◦

Probid Data◦

◦

•

◦



◦


◦

Conclusion◆

References◆

//www.emaze.com


SLIDE 8SLIDE 88

Dataset Description

StarPlus Data

2 cognitive states(Picture and Sentence).◆

54 trials, each trials of 27 seconds.◆

//www.emaze.com


SLIDE 9SLIDE 99

Dataset Description

StarPlus Data

Data is dimension of 64x64x8◆

5 subjects◆

//www.emaze.com


SLIDE 10SLIDE 1010

Dataset Description

Probid Data

3 cognitive states (Pleasant , Unpleasant andNeutral).

◆

5 subjects.◆

Dimension of 79x95x69, collected over 121time points.

◆

//www.emaze.com


SLIDE 11SLIDE 1111

Voxel Elimination ?

Data is high dimensional.◆

Minimum number of features 4500 forstarPlus.

◆

Maximum number of features 219727219727 forprobid.

◆

Solution --> RFE (Recursive featureElimination)

◆

//www.emaze.com


SLIDE 12SLIDE 1212

Sample Data in fMRI

Voxel 1 Voxel 2 Voxel 3 Voxel N Labelvalue value value value Class 1

value value value value Class 1



//www.emaze.com


SLIDE 13SLIDE 1313


Objective◆


Introduction To RFEIntroduction To RFE◆


◆


•

StarPlus Data◦

Probid Data◦

◦

•

◦



◦


◦

Conclusion◆

References◆

//www.emaze.com


SLIDE 14SLIDE 1414

Introduction To RFE

Multivariate feature extraction algorithm.◆

Recursively eliminates irrelevant features.◆

SVM classifier --> for removing irrelevant

voxels.

◆

//www.emaze.com


SLIDE 15SLIDE 1515

Introduction To RFEFlow Chart

//www.emaze.com


SLIDE 16SLIDE 1616

Introduction To RFE

Algorithm

//www.emaze.com


SLIDE 17SLIDE 1717


Objective◆




◆


•

StarPlus DataStarPlus Data◦

Probid Data◦

◦

•

◦



◦


◦

Conclusion◆

References◆

//www.emaze.com


SLIDE 18SLIDE 1818

RFE implementation on starPlus

Picture Vs Sentence Classfication

Analysis for each subject◆

Each task -- 8 seconds -- 16 images◆

Number of features --> 16 * number of16 * number ofvoxels voxels

◆

40 rows for picture and 40 rows for◆

//www.emaze.com


SLIDE 19SLIDE 1919

RFE implementation on starPlusResults

Feature Extraction LevelFeature Extraction Level

Clas

sific

atio

n Ac

cura

cyCl

assi

ficat

ion

Accu

racy

//www.emaze.com


SLIDE 20SLIDE 2020


Results


Clas

sific

atio

n Ac

cura

cyCl

assi

ficat

ion

Accu

racy

Results

//www.emaze.com


SLIDE 21SLIDE 2121


Results


Clas

sific

atio

n Ac

cura

cyCl

assi

ficat

ion

Accu

racy

Conclusion : Conclusion : As the feature extraction levels increases, accuracy also increases

Results

//www.emaze.com


SLIDE 22SLIDE 2222


Objective◆




◆


•

StarPlus Data◦

Probid DataProbid Data◦

◦

•

◦



◦


◦

Conclusion◆

References◆

//www.emaze.com


SLIDE 23SLIDE 2323

RFE implementation on Probid

Pleasant Vs Unpleasant classification

5 subjects.◆

50 rows for each of the task --> total 100rows.

◆

Number of voxels : 219727.◆

//www.emaze.com


SLIDE 24SLIDE 2424

RFE on Pleasant Vs UnPleasant Results


Clas

sific

atio

n Ac

cura

cyCl

assi

ficat

ion

Accu

racy

//www.emaze.com


SLIDE 25SLIDE 2525


Clas

sific

atio

n Ac

cura

cyCl

assi

ficat

ion

Accu

racy

RFE on Pleasant Vs UnPleasant


//www.emaze.com


SLIDE 26SLIDE 2626

RFE on Pleasant Vs Neutral


Clas

sific

atio

n Ac

cura

cyCl

assi

ficat

ion

Accu

racy


//www.emaze.com


SLIDE 27SLIDE 2727


Objective◆




◆


•

StarPlus Data◦

Probid Data◦

◦

•

◦



◦


◦

Conclusion◆

References◆

//www.emaze.com


SLIDE 28SLIDE 2828

RFE Experiment for correctness

Picture vs Sentence StarPlus Data.◆

Removed relevant voxels. ◆

Accuracy is decreased as feature extraction

level increases.

◆

//www.emaze.com


SLIDE 29SLIDE 2929

Reverse RFE Experiment for correctness

Conclusion: More the relevant percentage of voxels removed, moreConclusion: More the relevant percentage of voxels removed, moresharply accuracy decreased.sharply accuracy decreased.

//www.emaze.com


SLIDE 30SLIDE 3030


Objective◆




◆


•

StarPlus Data◦

Probid Data◦

◦

•

◦



◦


◦

Conclusion◆

References◆

//www.emaze.com


SLIDE 31SLIDE 3131

Time stamps analysis

Important Time Stamps analysis.

Objective

//www.emaze.com


SLIDE 32SLIDE 3232


Picture Vs Sentence data.◆

Analysis is done for individul subjects◆

Voxels from 7 ROIs ◆

Accuracy measured for each time stamp◆

Procedure

//www.emaze.com


SLIDE 33SLIDE 3333

Results

Time StampsTime Stamps

Clas

sific

atio

n Ac

cura

cyCl

assi

ficat

ion

Accu

racy


//www.emaze.com


SLIDE 34SLIDE 3434

Voxels from 7-14 timestamps (3.5 sec to 7 sec)are more discriminating.

◆

This conforms to theconcept of HRF(Heodynamic ResponseFunction )

◆

DiscussionTime stamps analysis

//www.emaze.com


SLIDE 35SLIDE 3535

Results

SubjectsSubjects

Clas

sific

atio

n Ac

cura

cyCl

assi

ficat

ion

Accu

racy


//www.emaze.com


SLIDE 36SLIDE 3636


Objective◆




◆


•

StarPlus Data◦

Probid Data◦

◦

•

◦

OutlineROI AnalysisROI Analysis◦


◦


◦

Conclusion◆

References◆

//www.emaze.com


SLIDE 37SLIDE 3737

ROI analysis

Important Region of interest Analysis

Objective

//www.emaze.com


SLIDE 38SLIDE 3838

ROI analysis

Picture Vs Sentence data.◆

Voxels from 7 ROIs.◆

Analysis -- individual subjects.◆

Most discriminating Voxels are found fromRFE.

◆

Distribution of these voxels are analyzedacross the ROIs.

◆

Procedure

//www.emaze.com


SLIDE 39SLIDE 3939

Results


Clas

sific

atio

n Ac

cura

cyCl

assi

ficat

ion

Accu

racy

ROI analysis

//www.emaze.com


SLIDE 40SLIDE 4040

3 ROIs (CALC, LIPL, LIPS) voxels are mostdiscriminating.

◆

70 % of the discriminating voxels are inthis 3 regions.

◆

DiscussionROI analysis

//www.emaze.com


SLIDE 41SLIDE 4141


Objective◆




◆


•

StarPlus Data◦

Probid Data◦

◦

•

◦



◦


◦

Conclusion◆

References◆

//www.emaze.com


SLIDE 42SLIDE 4242

Across all subjects.◆

Picture Vs Sentence data classification usingRFE.

◆

Purpose of this analysis, to find class forgeneralized test data.

◆

Mean value across ROIs is taken.◆

Procedure Across Subjects analysis -- 7 ROIs

//www.emaze.com


SLIDE 43SLIDE 4343

Results


Across Subjects analysis -- 7 ROIs

Clas

sific

atio

n Ac

cura

cyCl

assi

ficat

ion

Accu

racy

//www.emaze.com


SLIDE 44SLIDE 4444

Data is not classified accurately from 7ROIs voxels, across all the subjects.

◆

Some irrelevant features needs toremoved

◆

Discussion


//www.emaze.com


SLIDE 45SLIDE 4545


Objective◆




◆


•

StarPlus Data◦

Probid Data◦

◦

•

◦



◦


◦

Conclusion◆

References◆

//www.emaze.com


SLIDE 46SLIDE 4646

Only 3 ROI's (CALC, LIPS ,LIPL) voxels taken.◆

Procedure


//www.emaze.com


SLIDE 47SLIDE 4747

Results


Clas

sific

atio

n Ac

cura

cyCl

assi

ficat

ion

Accu

racy

Subjects analysis -- 3 ROIs

//www.emaze.com


SLIDE 48SLIDE 4848

Results Across Subjects analysis -- 3 ROIs

Feature extraction levels, voxelsFeature extraction levels, voxels

Clas

sific

atio

n Ac

cura

cyCl

assi

ficat

ion

Accu

racy

//www.emaze.com


SLIDE 49SLIDE 4949

Important voxels --> 3 ROIs (CALC, LIPS, LIPL)◆

Discussion


//www.emaze.com


SLIDE 50SLIDE 5050


Objective◆




◆


•

StarPlus Data◦

Probid Data◦

◦

•

◦



◦


◦

ConclusionConclusion◆

References◆

//www.emaze.com


SLIDE 51SLIDE 5151

Conclusion

Machine learning algorithm can besuccessfully utilized for fMR dataanalysis.

◆

Brain mapping corresponding tocognitive states established.

◆

3 Important Regions are CALC, LIPS ,LIPL in Picture vs Sentence analysis.

◆

//www.emaze.com


SLIDE 52SLIDE 5252


Objective◆




◆


•

StarPlus Data◦

Probid Data◦

◦

•

◦



◦


◦

Conclusion◆

ReferencesReferences◆

//www.emaze.com


SLIDE 53SLIDE 5353

References

De Martino, Federico and Valente, Giancarlo and

Staeren, Noel and Ashburner, John and Goebel, Rainer

and Formisano, Elia. Combining multivariate voxelCombining multivariate voxel

selection and support vector machines for mapping andselection and support vector machines for mapping and

classification of fMRI spatial patternsclassification of fMRI spatial patterns. Neuroimage,

43(1), 2008

◆

Mitchell, Tom M and Hutchinson, Rebecca and

Niculescu, Radu S and Pereira, Francisco and Wang,

Xuerui and Just, Marcel and Newman, Sharlene. LearningLearning

◆

//www.emaze.com


SLIDE 54SLIDE 5454

References

Weblink. http://www.cs.cmu.edu/afs/cs.cmu.edu/project

/theo-81/www/ StarPlus Data.StarPlus Data.

◆

Weblink. http://www.brainmap.co.uk/ PROBID data.PROBID data.◆

//www.emaze.com

thesis defence copy1

Documents