Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Kennesaw State UniversityDigitalCommons@Kennesaw State University
Master of Science in Computer Science Theses Department of Computer Science
Spring 6-6-2018
NEW ALGORITHMS FOR COMPRESSEDSENSING OF MRI: WTWTS, DWTS,WDWTSSrivarna Settisara Janney
Follow this and additional works at: https://digitalcommons.kennesaw.edu/cs_etd
Part of the Bioimaging and Biomedical Optics Commons, Biomedical Commons, and the SignalProcessing Commons
This Thesis is brought to you for free and open access by the Department of Computer Science at DigitalCommons@Kennesaw State University. It hasbeen accepted for inclusion in Master of Science in Computer Science Theses by an authorized administrator of DigitalCommons@Kennesaw StateUniversity. For more information, please contact [email protected].
Recommended CitationSettisara Janney, Srivarna, "NEW ALGORITHMS FOR COMPRESSED SENSING OF MRI: WTWTS, DWTS, WDWTS" (2018).Master of Science in Computer Science Theses. 11.https://digitalcommons.kennesaw.edu/cs_etd/11
1
NEW ALGORITHMS FOR COMPRESSED SENSING OF MRI:
WTWTS, DWTS, WDWTS
An Abstract of
A Thesis Presented to
The Faculty of Department of Computer Science
by
Srivarna Settisara Janney
In Partial Fulfillment
of Requirements for the Degree
Master of Science
Kennesaw State University
May 2018
2
ABSTRACT
Magnetic resonance imaging (MRI) is one of the most accurate imaging
techniques that can be used to detect several diseases, where other imaging
methodologies fail. MRI data takes a longer time to capture. This is a pain taking
process for the patients to remain still while the data is being captured. This is
also hard for the doctor as well because if the images are not captured correctly
then it will lead to wrong diagnoses of illness that might put the patients lives in
danger. Since long scanning time is one of most serious drawback of the MRI
modality, reducing acquisition time for MRI acquisition is a crucial challenge for
many imaging techniques. Compressed Sensing (CS) theory is an appealing
framework to address this issue since it provides theoretical guarantees on the
reconstruction of sparse signals while projection on a low dimensional linear
subspace. Further enhancements have extended the CS framework by performing
Variable Density Sampling (VDS) or using wavelet domain as sparsity basis
generator. Recent work in this approach considers parent-child relations in the
wavelet levels.
This paper further extends the prior approach by utilizing the entire
wavelet tree structure as an argument for coefficient correlation and also considers
the directionality of wavelet coefficients using Hybrid Directional Wavelets
(HDW). Incorporating coefficient thresholding in both wavelet tree structure as
well as directional wavelet tree structure, the experiments reveal higher Signal to
Noise ratio (SNR), Peak Signal to Noise ratio (PSNR) and lower Mean Square
Error (MSE) for the CS based image reconstruction approach. Exploiting the
sparsity of wavelet tree using the above-mentioned techniques achieves further
lessening for data needed for the reconstruction, while improving the
reconstruction result. These techniques are applied on a variety of images
including both MRI and non-MRI data. The results show the efficacy of our
techniques.
3
NEW ALGORITHMS FOR COMPRESSED SENSING OF MRI: WTWTS, DWTS, WDWTS
A Thesis Presented to
The Faculty of Department of Computer Science
by
Srivarna Settisara Janney
In Partial Fulfillment
of Requirements for the Degree
Master of Science
Advisors: Dr. Sumit Chakravarty and Dr. Chih-Cheng Hung
Kennesaw State University
May 2018
4
NEW ALGORITHMS FOR COMPRESSED SENSING OF MRI: WTWTS, DWTS, WDWTS
Approval Committee Members: Dr. Chih-Cheng Hung, Thesis Director and Advisor Department of Computer Science Kennesaw State University Dr. Sumit Chakravarty, Advisor Department of Electrical Engineering Kennesaw State University Dr. Craig Chin Department of Electrical Engineering Kennesaw State University Dr. Mingon Kang Department of Computer Science
Kennesaw State University
5
DEDICATION
This thesis is dedicated to my husband who has been my inspiration and my strongest supporter. I would also thank my parents, in-laws, brother & his family, and sister-in-law & her family.
6
PREFACE
The thesis is submitted in partial fulfillment of the requirements for the Master of Science degree program at Kennesaw State University, Kennesaw.
The research project has been conducted under the supervision of Dr. Sumit Chakravarty from Department of Electrical Engineering and Dr. Chih-Cheng Hung from Department of Computer Science, during the years 2017-2018. Financing for the work has been provided in the form of scholarship from the Graduate College of Kennesaw State University through the Graduate Research Assistantship.
7
ACKNOWLEDGEMENTS
I would like to express my gratitude to my advisors Dr. Sumit Chakravarty and Dr. Chih-Cheng Hung for their support, encouragement, and motivation through this entire thesis. Thank you for believing in me and leading me to the path of research. I would also like to thank my thesis committee members, Dr. Mingon Kang and Dr. Craig Chin for their insightful comments and valuable suggestions.
8
QUOTES
“Arise,Awake,andstopnotuntiltheGoalisachieved”
- SwamiVivekananda
“Theimportantthingistonotstopquestioning.Curiosityhasitsownreasonfor
existing.”-AlbertEinstein
“Weallneedpeoplewhowillgiveusfeedback.That’showweimprove.”-BillGates
9
TableofContents
Chapter1............................................................................................................141.1 Introduction............................................................................................................................................141.1.1 MagneticResonanceImaging(MRI).......................................................................................151.1.2 BasicConceptsofCompressedSensing...................................................................................16
Chapter2............................................................................................................212.1 Literaturesurveyandproblemstatement................................................................................212.1.1 CompressedSensingforMRI.......................................................................................................212.1.2 TraditionalCompressedSensingMethodologiesinMRI.................................................242.1.3 Drawbacks.........................................................................................................................................282.1.4 ProblemStatement.......................................................................................................................28
Chapter3............................................................................................................293.1 ProposedSolutions..............................................................................................................................293.1.1 WeightedThresholdWaveletTreeStructure(WTWTS)................................................303.1.2 DirectionalWaveletTreeStructure(DWTS).......................................................................323.1.3 WeightedDirectionalWaveletTreeStructure(WDWTS).............................................35
3.2 ComplexityAnalysis............................................................................................................................37
Chapter4............................................................................................................384.1 Experimentsandresults...................................................................................................................384.1.1 TestEnvironmentSpecification.................................................................................................384.1.2 DataSet................................................................................................................................................384.1.3 Hyper-Parameters...........................................................................................................................414.1.4 Parameters..........................................................................................................................................424.1.5 PreliminaryResults.........................................................................................................................424.1.6 ResultsandEvaluation..................................................................................................................524.1.7 ComparisonbetweenWTWTS,DWTS,WDWTS.................................................................58
Chapter5............................................................................................................59FutureWork.......................................................................................................................................................59
Chapter6............................................................................................................63Conclusion...........................................................................................................................................................63Bibliography.......................................................................................................................................................64Appendix..............................................................................................................................................................72ListofAlgorithms.............................................................................................................................................72
10
ListofFigures
Figure1.1:MRIscanning..............................................................................................................................14Figure1.2:MRIscanningmachine[2]....................................................................................................16Figure1.3:CompressedSensingvsImageCompressionFramework[3]...............................17Figure1.4:BasicCompressedSensingmatrix[3].............................................................................18Figure1.5:Comparisonbetween(a)l1issparseand(b)l2normsisnotsparse...............19Figure1.6:(a)kspacedata(b)kspacedatawithnoiselikepatterncausebyincoherence
........................................................................................................................................................................19Figure2.1:ImageReconstructionfromk-spacetopixelimage[40]........................................22Figure2.2:RepresentstheoriginalimageanditsreconstructedimageusingSVD
method.......................................................................................................................................................25Figure2.3:(a)OriginalImage(b)CSwithPCA,redcirclemeanstheareahashigherSNR
comparedtopreviousmethods.......................................................................................................26Figure2.4:(a)WaveletDecomposition(b)DecomposedImagedirection............................27Figure3.1:WaveletQuad-TreeStructure(a)AcardiacMRimage(b)Corresponding
WaveletTreeGroupofHierarchyGP-P-CStructure...............................................................30Figure3.2:Directionalfilterbankfrequencypartitioningusingeightdirections...............32Figure3.3:(a)Verticaldirectionalfilterbanks(b)Horizontaldirectionalfilterbanks...32Figure3.4:ExampleofDirectionalfilterbankonrepresentation..............................................33Figure3.5:EightDirectionalfilterbanksrepresentusingclassicBarbaraImage..............33Figure4.1:MRIData.......................................................................................................................................39Figure4.2:SamplingMask...........................................................................................................................39Figure4.3:(a)T1imageand(b)T2image...........................................................................................40Figure4.4:PortraitData...............................................................................................................................40Figure4.5:CrowdData..................................................................................................................................41Figure4.6:LandscapeData.........................................................................................................................41Figure4.7:VariationofnumberofzerotreesbasedonGP-P-CHierarchyforonebrain
image...........................................................................................................................................................43Figure4.8:VariationofnumberofzerotreebasedonGP-P-CHierarchyforT1images.43Figure4.9:VariationofnumberofzerotreesbasedonGP-P-ChierarchyforT2image.44Figure4.10:VariationofnumberofzerotreesbasedonGP-P-Chierarchylevelforentire
500brainimages....................................................................................................................................44Figure4.11:3DGraphofrangeofthresholdsetsagainstEntropyofdataatParent-Child
Level.............................................................................................................................................................45Figure4.12:ComparisonofdifferentthresholdingsetsagainstSNRfordifferent
resolutionsizesforBrainImages(RedlineindicatereferenceSNRofWatMRIalgorithm)..................................................................................................................................................46
Figure4.13:ComparisonofdifferentthresholdingsetsagainstSNRfordifferentresolutionsizesforHeartImages(RedlineindicatereferenceSNRofWatMRIalgorithm)..................................................................................................................................................46
Figure4.14:ComparisonofdifferentthresholdingsetsagainstSNRfordifferentresolutionsizesofShoulderImages(RedlineindicatereferenceSNRofWatMRIalgorithm)..................................................................................................................................................46
11
Figure4.15:ComparisonofdifferentthresholdingsetsagainstSNRfordifferentresolutionsizesforChestImages(RedlineindicatereferenceSNRofWatMRIalgorithm)..................................................................................................................................................47
Figure4.16:SNRcomparisonbetweenWatMRI(Greencolor)andWTWTSwithSet2(Redcolor)and3(Bluecolor)approaches.................................................................................47
Figure4.17:ComparisontheSNRforPortrait,CrowdandLandscapeimagesforthresholdingset2and3(RedlineindicatesreferenceSNRofWatMRIalgorithm)48
Figure4.18:Histogramof8DirectionsatParentandChildlevelforBrainImage.............48Figure4.19:Histogramof8DirectionsatParentandChildlevelforHeartImage.............49Figure4.20:Histogramof8DirectionsatParentandChildlevelforShoulderImage.....49Figure4.21:Histogramof8DirectionsatParentandChildlevelforChestImage.............50Figure4.22:ComparisonofPortraitimageagainstoriginaldatausingWTWTSand
WDWTSalgorithmswithchosenthresholdset........................................................................51Figure4.23:ComparisonofCrowdimageagainstoriginaldatausingWTWTSand
WDWTSalgorithmswithchosenthresholdset........................................................................51Figure4.24:ComparisonofLandscapeimageagainstoriginaldatausingWTWTSand
WDWTSalgorithmswithchosenthresholdset........................................................................52Figure4.25:Brainimagereconstructionusingthe3algorithmscomparedagainst
WatMRI.......................................................................................................................................................53Figure4.26:Heartimagereconstructionusingthe3algorithmscomparedagainst
WatMRI.......................................................................................................................................................53Figure4.27:Shoulderimagereconstructionusingthe3algorithmscomparedagainst
WatMRI.......................................................................................................................................................54Figure4.28:Chestimagereconstructionusingthe3algorithmscomparedagainst
WatMRI.......................................................................................................................................................54Figure4.29:Portraitimagereconstructionusingthe3algorithmscomparedagainst
WatMRI.......................................................................................................................................................55Figure4.30:Crowdimagereconstructionusingthe3algorithmscomparedagainst
WatMRI.......................................................................................................................................................55Figure4.31:Landscapeimagereconstructionusingthe3algorithmscomparedagainst
WatMRI.......................................................................................................................................................56Figure5.1:OverviewofDeepNeuralNetworkforCSinMRI.......................................................59Figure5.2:ReconstructionErrorflowchart........................................................................................60Figure5.3:Multi-SliceofBrainMRI[38]...............................................................................................60Figure5.4:3DBrainMRIdisplayingeachdimension[40]............................................................61Figure5.5:DynamicMRIrepresentedinTensor[46].....................................................................61Figure5.6:VisualizationofHigherOrderSingleValueDecomposition(HOSVD)(a)3D
Tensor(b)4DTensor[45].................................................................................................................62
12
ListofTables
Table3.1:PseudoCodeofWTWTS..........................................................................................................31Table3.2:PseudoCodeofDWTS..............................................................................................................34Table3.3:PseudoCodeofWDWTS..........................................................................................................36Table4.1:SNRandMSEcomparisononMRIdata............................................................................56Table4.2:CPUTimecomparisononMRIdata....................................................................................57Table4.3:SNRandMSEcomparisononNon-MRIdata..................................................................57Table4.4:CPUTimecomparisononNon-MRIdata..........................................................................58
13
ListofAcronyms
BMP:BasisMatchingPursuit.....................................................................................................................................22C:Child.....................................................................................................................................................................................29CS:CompressedSensing...............................................................................................................................................14DFB:DirectionalFilterBanks.....................................................................................................................................32DWTS:DirectionalWaveletTreeStructure....................................................................................................30,32FCSA:FastCompositeSplittingAlgorithm..........................................................................................................23FISTA:FastIterativeShrinkageThresholdingAlgorithm...........................................................................27GP:Grandparent.................................................................................................................................................................29HDW: Hybrid Directional Wavelets............................................................................................................................2MRI:MagneticResonanceImaging.........................................................................................................................14MSE: Mean Square Error..................................................................................................................................................2OMP:OrthogonalMatchingPursuit........................................................................................................................22P:Parent..................................................................................................................................................................................29PSNR: Peak Signal to Noise Ratio...............................................................................................................................2SNR: Signal to Noise Ratio.............................................................................................................................................2TVCMRI:TotalVariationCompressedSensingMRI......................................................................................23VDS: Variable Density Sampling.................................................................................................................................2WaTMRI:WaveletTreeMRI.......................................................................................................................................27WDWTS:WeightedDirectionalWaveletTreeStructure..........................................................................30,35WTWTS:WeightedThresholdWaveletTreeStructure....................................................................................30
14
Chapter1
1.1 Introduction
OneofthecorechallengesofMagneticResonanceImaging(MRI)
reconstructionisthesignificanttimedurationrequiredtoacquirethe
images.Oneoftheconstraintsisthatpatientshavetoremainstationaryfor
thedurationofscanthattypicallytakesmorethananhour.Another
constraintisthattheresolutionofthereconstructedimagesshouldbeof
significantlyhighgrade.InthecaseofMRI,considerableattentionhasbeen
paidtoCompressedSensing(CS)thatisknownasoneofthemostsuccessful
methodsforimagereconstructionwithsignificantlysparseamountof
coefficients.Figure1.1showstheMRIscanningprocessthatpatientshaveto
undergowhilegettingaMRIscan.
Figure1.1:MRIscanning
Initiallythepatientsaregivenaninjectioncontaining“Gadolinium”whichis
aheavymetaloralsocalledas“contrastagent”.Theywaitfor10-15minsfor
theinjectedcontenttospreadthroughouttheareathatneedtobescanned.
Patientisthenaskedtoliedownonaslidingtablethatentersthehuge
magneticcoildrumthatcapturesthescandatainspaceknownask-space
data.CompressedSensingofwaveletisusedtoaccuratelyreconstructthe
imagethatknownasMRIscanningimage.Thewholeprocesstypicallytakes
aroundtwohoursofpatient’stime.Ifthereisanymovementbythepatient,
15
theninaccuratereconstructiontakesplaceandprocessneedstoberepeated
forrecapture.Thisisachallengingsituationforbothpatientsandradiologist.
ThisresearchpaperdevelopedthreenewalgorithmsnamelyWeighted
ThresholdWaveletTreeStructure(WTWTS),DirectionalWaveletTreeStructure
(DWTS)andWeightedDirectionalWaveletTreeStructure(WDWTS),using
multipleMRIdataincludingBrain,Chest,Shoulder,andHearttypeofdatasets.
DevelopedalgorithmswerealsotestedonNon-MRIdata(Portrait,Landscape,
CrowdImages)togeneralizethealgorithmtoanyimagedata.
1.1.1 MagneticResonanceImaging(MRI)
MagneticResonanceImaging(MRI)scanisatechniqueusingby
radiologiststhatusesmagnetismofthehugemagneticcoildrumandalgorithms
areusedtoreconstructimageofbodystructures.Fortheprocedures,acontrast
agent,suchasgadolinium,whichisaheavymetal,isinjectedintothepatient’s
bodypartthatneedstoscan.TheMRIscannerisagiantcirculardrumofmagnet
andaslidingtableasshowninFigure1.2.Thepatientisplacedonamoveable
bedthatisinsertedintothemagnet.Themagnetcreatesastrongmagneticfield
thatalignstheprotonsofhydrogenatoms.Thisspinsthevariousprotonsofthe
body,andtheyproduceasignalthatisdetectedduringtheMRIscanner.A
computerprocessesthereceiverinformation,andaftercompressedsensing
usingwavelettoreconstructthefinalimage,knownasMRIisproduced.
TheimageandresolutionproducedbyMRIneedstobedetailedand
shoulddetecttinychangesofstructureswithinthebody.Theradiologistwill
ablebeaccuratelydiagnosediseasesorstructuralabnormalitiesbasedonthe
qualityofimagethatisreconstructedfromthemachine.Ifthepatientmoves
thenthedatawillnotbeaccuratelycapturedandtheradiologisthastoredothe
scanning.
16
Patientswhohaveanykindofmaterialsthatcontainmetalintheirbodies
willcreateaproblemduringscanning.Normallytheyareaskedtoremovethings
likejewelry,clipsetc.Mostoftheinternalmetalmaterialsthatareinsidethe
patient’sbodyaddedasapartofmedicalproceduressuchasartificialjoints,
pacemaker,metallicboneplates,orprostheticlimbsetc.coulddistortMRI.
Scanningcannotbeperformedonthesepatients.Alternativemethodswillbe
usedfordiagnosis[1].
Figure1.2:MRIscanningmachine[2]
Sincesomepatientscanbeclaustrophobicwhoexperiencefearofclosedplaces,
theradiologistmayrecommendamildsedativepriortotheMRIscanning.The
rotatingofmagneticcoilgenerateshugenoise,soearplugsmightbegivento
patientswhorequestthem.
1.1.2 BasicConceptsofCompressedSensing
17
CompressedSensingshouldnotbeconfusedwithtraditionalimage
compression.Figure1.3showstheframeworkofthetwoconceptswhereXisthe
originalinputsignalandXcisthereconstructedsignalwithsampleNnumberof
sampleswhileMrepresentsthesparsenumberofsamplethatisrequiredfor
reconstruction.
Figure1.3:CompressedSensingvsImageCompressionFramework[3]
Compressedsensing(alsoknownascompressivesensing,sparsesamplingor
compressivesampling,CS)isasignalprocessingtechniqueforefficiently
collectingsignalandaccuratelyreconstructingthesignal.Basedontheprinciple
ofsparsityofthesignalfound,thesesignalshavemuchfewersamplesthanthat
requiredbytheShannon-Nyquistsamplingtheorem[4],[5],[6],[7].
CShasgainedalotofpopularityvariousareaoverlastfewdecades.This
conceptwasintroducedbyCandes,Romberg,andTao[12],andDonoho[14].CS
helpstorecoverthesparsesignalsusingonlyfewsamplesincontrasttoNyquist
samplingtheory,wheresamplingofthesignalisperformedataratelargerthan
highestfrequencypresentinthesignal.Anoverviewofcompressedsensing
theorycanbefoundin[8],[9],[10],[11].
Consider𝑏tobetheactualsignalhavingdimensionsNx1where𝑏 ∈ ℝ.Φ
18
representsthesamplingmatrixwithMxNdimensionsandyisthefinal
compressedvectorwithdimensionMx1whichismuchsmallerthanN.i.eM<<N
𝑦 = Φ𝑏(1)
AccordingtoCompressedSensing,theequationlookslike
𝑦 = ΦΨ𝓍=A𝓍(2)
wherebisthenon-sparsesignalrepresentedas𝑏=Ψ𝓍,𝓍=sparse
representationintransformdomainΨ.A=ΦΨwhichrepresentsdimension
MxNsensingmatrixhavingksparsemeasurementrequiredforreconstructionof
image.Theaimistominimizecoefficientstogetsparsesignalcalledmin! .Figure
1.4showsthefundamentalconceptsinCSmatrixrepresentationthatisshownin
equation(2).
Figure1.4:BasicCompressedSensingmatrix[3]
TherearetwofundamentalconceptsinCompressedSensing.Firstone
issparsityofthesignal;thatis,thesignalneedstosparseintransformdomain.
Thesecondoneisincoherence;thatis,appliedthroughtheRestrictedIsometric
Property(RIP)thatissufficientforsparsesignals.
Theabovereconstructionproblemisnon-deterministicpolynomial-time
19
hard(NP-hard).Duetothenon-convexnatureof𝑙!minimization,theproblemis
illconditionedanddifficulttosolveasitrequiresanexhaustivesearchtofindthe
mostsparsesolution(i.e.combinatorialproblem)[16].Sincetheuniquesparse
solutioncanbefoundexactlyusing𝑙!minimization.𝑙!normisknownforleast
squaresolution,apenaltywillbeimposedonsmallnon-zerocoefficientswhere
ittendstoenergytolargenumberofentriesandtheresultisnotsparse[13].
ComparisonbetweenthenormsisasshowninFigure1.5andtheincoherenceis
showninFigure1.6.
(a)(b)
Figure1.5:Comparisonbetween(a)l1issparseand(b)l2normsisnotsparse
(a)(b)Figure1.6:(a)kspacedata(b)kspacedatawithnoiselikepatterncausebyincoherence
20
Inadditiontosparsity,theminimumacceptablenumberofsamplesfor
reconstructionalsodependsupontheincoherencebetweenΦ andΨ,suchthat
thereisnocorrelationbetweentherowsofΦandthecolumnsofΨ.Toachieve
this,“A”asexplainedintheequation(2)shouldsatisfytheRestrictedIsometry
Property(RIP)[17].
1.1.2.1 Disadvantages
Earlyiterationsmayfindinaccuratesampleestimates,howeverthis
methodwilldown-sampletheseatalaterstagetogivemoreweighttothe
smallernon-zerosignalestimates.Oneofthedisadvantagesistheneedfor
definingavalidstartingpoint,asaglobalminimummightnotbeobtainedevery
timeduetotheconcavityofthefunction.Anotherdisadvantageisthatthis
methodtendstouniformlypenalizetheimagegradientirrespectiveofthe
underlyingimagestructures.Thiscausesover-smoothingofedges,especially
thoseoflowcontrastregions,subsequentlyleadingtolossoflowcontrast
information[3].
1.1.2.2 Advantages
Theadvantagesofthismethodinclude:reductionofthesamplingratefor
sparsesignals;reconstructionoftheimagewhilebeingrobusttotheremovalof
noiseandotherartifacts;anduseofveryfewiterations.Thiscanalsohelpin
recoveringimageswithsparsegradients[3].
21
Chapter2
2.1 Literaturesurveyandproblemstatement
Thissectionisrelatedtopreviousresearchesthathavebeencarried
out.Eachsub-sectionshowcasestheidea,equationandthescopeof
improvementthatbeenbroughtaboutthroughthefindings.Theseconcepts
markabaselineforthisresearchisbuiltoversomeoftheseideas.
2.1.1 CompressedSensingforMRI
CompressedSensing(CS)isoneofthemostsuccessfulmethodsof
imagereconstructionwithsignificantlysparseamountofcoefficients.
AccordingtoCStheory[15][19],onlyO(K+Klogn)sampling
measurementsareenoughtorecoverK-sparsedatawithlengthn,the
probleminovercomingthechallengeassociatedwithimagingwithouta
significantdegradationinimagequality.Thekeyapproachofcompressed
sensingMRIapplicationsistoutilizeseparatedomains.AsMRIimagesare
formedinK-space(FourierDomain)itbecomestheprimarydomainfor
monitoringreconstructionerrorasshownintheFigure2.1.Ithasbeenwell
establishviatheresearchofLustigandDonoho[19]toimplementsparsity
thatneedstooperateinasparsifyingdomainlikewavelets.Finallyastheend
consumersoftheimagedata,theneedistovisualizetheresultinthespatial
sampledomain.
22
Figure2.1:ImageReconstructionfromk-spacetopixelimage[40]
Thenon-linearconjugategradient(NLCG[18])wasappliedtosolve
theunconstrainedproblemintheLagrangianformtogetequation(3):
𝑚𝑖𝑛!{ 𝐹 𝑥 = !!𝐴𝑥− 𝑏 !
! + 𝛽 𝜙! !} (3)
Where𝐴=samplingmatrixobtainedfrompartialFouriertransformof
𝑥=MxNimagetobereconstructed,𝛽=positiveparameter,b=original
signaldataand𝜙=sparsifyingdomain,
Equation(3)attemptstominimizethemeansquareerrorinthe
frequencydomainwhilechoosingthesparsestsetofcoefficientsinthe
waveletdomainwithL2norm.Thisoptimizationcanbeperformedvia
varioustechniqueslikeOMP(OrthogonalMatchingPursuit)[21]orBasis
MatchingPursuit(BMP)[22]algorithms.
SeveralstudieshavebeendoneinCS-MRIdomain.Muchoftheworkdeals
23
withmanipulatingtheconstraintterminequation(3).Techniquesinthe
gamutofsignalprocessingliketotalvariationsmoothening,opticalflowof
coefficients,[23]variabledensitysampling[24]andsingularvalue
decomposition[25],conjugategradient[26]andpartialFourier(RecPF)[57]
aresomeoftheproposedapproachesusedincurrentliterature.Thenext
sectionwillpresentrelatedworkonwhichourcurrentworkisbased.
Waveletshavebeenextensivelyusedasacompressivedomaininthe
fieldofCS.Apartfromthepreviouslylistedtechniques;thefollowing
methodsdeservespecialattention.In[24]amethodcalledtotalvariation
compressedsensingMRI(TVCMRI)ispresented.Useoftotalvariationhelps
regularizethespatialvariabilityofthewaveletcoefficientdomain.Itusesan
operator-splittingalgorithmtosolvetheproblemofreconstructioncalled
thefastcompositesplittingalgorithm,(FCSA)[53].Thisapproachextends
equation(3)withtheadditionaltermappendedattheendofequation(4).
𝑚𝑖𝑛!{ 𝐹 𝑥 = !!𝐴𝑥− 𝑏 !
! +𝛼 𝑥 !" + 𝛽 𝜙! !}(4)
where 𝛼 𝑥 !" = (∇!"! )! + (∇!"! )!!! (5)
α=positiveparameter,wheretheterminsidetherootissummation
oftheresultoffinitedifferenceoperatorsonthefirstandsecondcoordinates
respectively,𝛽=positiveparameter, ∇denotestheforwardfinitedifference
operatorforfirstandsecondco-ordinateswhichisrepresentedintermsof𝑥
as
∇!"! = 𝑥!!!,! − 𝑥!"
∇!"!= 𝑥!,!!! − 𝑥!"
SignaltoNoiseRatiomeasuresthequalityofreconstructedimageshowin
24
equation(6).
SNR = !!
!(6)
Whereµ!=averagerootmeanofsignaland𝜎isthestandarddeviation
2.1.2 TraditionalCompressedSensingMethodologiesinMRI
Therearefewwell-knownorconventionalcompressedsensing
algorithmsthatarecommonlyusedinMRI.Someofthemarementioned
below.
1) CSofMRIusingSingularValueDecomposition(SVD)
Accordingtoauthorsof[43],thereconstructedimageMcanbe
representedusingtheSingularvalueDecompositionformula
M = UΣV!(7)
whereUandVareorthogonalmatricesandΣisdiagonalmatrixwithnon-
zerosingularvalues.
Whenexquation(7)isusedininverseFourierdomainthenweget
Ψ m = U!∗mV!(8)
whereU!andV!areleftandrightunitarymatrixthatareabletoprovidea
sparserepresentationforthedesiredimageMwhenprojectedonU!andV!
fornnumberofsamples.Forbetterresults,matricesUandVareiterated
withSVDusingtheequation(3).Figure2.2representstheoriginalimage
anditsreconstructedimageusingSVDmethod.
25
Figure2.2:RepresentstheoriginalimageanditsreconstructedimageusingSVD
method.
InFigure2.2,Thetoprowshowstheprocessofgeneratingthesparsity
basisfromlefttoright:theimageI,theunitarymatrixU,thediagonal
matrixandunitarymatrixV.ThebottomrowshowstheresultofSVDwith
sparserepresentationofdesiredimagefromlefttoright:thesparseimage
M,theunitarymatrixU,thedesiredimageMandunitarymatrixV.
2) CSofMRIusingPrincipleComponentAnalysis(PCA)
In[24]amethodcalledtotalvariationcompressedsensingMRI
(TVCMRI)ispresented.Useoftotalvariationhelpsregularizethespatial
variabilityofthewaveletcoefficientdomain.Itusesanoperator-splitting
algorithmtosolvetheproblemofreconstructioncalledthefastcomposite
splittingalgorithm(FCSA)[53].
𝑚𝑖𝑛!{F m = ∥ Φ! m − y ∥!!+ βTV(m)+ λ ∥ Ψ m ∥!}(9)
Whereβinequation(9)representstheweightofthetotalvariation(TV)thatis
theaddedtermfordenoising.
AnothermethodforCSbasedonPCAanalysisinwhichtheweightofnorm
26
LIoftheimagewasadaptivelyupdatedateachiterationwithrespectto
reductionofdimensions.In[39]method,PCAadaptiveapproachastheweight
ofsparsity,theconstrainedoptimizationisconvertedintotheminimizationof
thefollowingequation:
𝑚𝑖𝑛!{F m = ∥ Φ! m − y ∥!!+ βTV m +∥ u×s×v ∥! ∥ Ψ m ∥!}(10)
Matricesu,sandv,arethePCAelements([u,s,v)=PCA(m)).
EfficiencyandaccuracyofPCAisclaimedtobe15%higherthantheSVD
method.ResultcomparisonisshowninFigure2.3.
(a)(b)Figure2.3:(a)OriginalImage(b)CSwithPCA,redcirclemeanstheareahashigherSNR
comparedtopreviousmethods
3) CSofMRIusingWaveletTreeStructure(WaTMRI)
Waveletsarefunctionsthatsatisfycertainmathematicalrequirements
andareusedinrepresentingdataorotherfunctions.Waveletanalysisis
usedtoapproximatefunctionsthatarecontainedneatlyinfinitedomains.
Waveletsarebestsuitedforapproximatingdatawithsharpdiscontinuities.
Waveletapplicationsareusedinareassuchasimagecompression,computer
27
vision,predictionofearthquakeprediction,etc.Whenmovedtofrequency
domainfromtimedomainthenDiscreteWaveletTransform(DWT)isused
andDiscreteDaubchiesWaveletTransformisusedtoobtainwavelet
coefficients.Thewaveletdecompositionconsistsoffourcomponentscalled
Approximation,Horizontal,VerticalandDiagonaldetailsasshowninthe
Figure2.4.
(a)(b)
Figure2.4:(a)WaveletDecomposition(b)DecomposedImagedirection
ThishasfurtherbeenappliedinWaTMRI[29]inadditiontousingtotal
variation,parentchildrelationshipsofwaveletsareconsidered.Theauthors
groupallparentchildcoefficients.Theresultantoptimizationequation(11)
is
𝑚𝑖𝑛!{ 𝐹 𝑥 = !!𝐴𝑥− 𝑏 !
! +𝛼 𝑥 !" + 𝛽 𝜙! ! + (𝜙!)! !!∈! }(11)
Theadditionofthe“parent-child”group(g)termenforcestheabove
optimizationtoconsidercoefficientsinparentchildpairsandtherebyaims
todiscardfromtheresult,weakorunderperformingcoefficientgroups.
Theequation(11)issolvedusingFastIterativeShrinkageThresholding
Algorithm(FISTA)[26],[27].
28
2.1.3 Drawbacks
AlthoughtheWaTMRIapproachconsiderscorrelationamongthe
waveletcoefficients,itisrudimentary.Italsosuffersfromthedisadvantages
ofconventionalwavelettransform:sensitivetoshifting[31],poor
directionality[32]anddearthofphaseinformation[33].
2.1.4 ProblemStatement
Problemtoberesolvedisimagereconstructionwithminimalerroras
wellastospeeduptheruntimeofthealgorithm.Thedirectionalityof
waveletistakenintoconsideration.Sincethetargetistogeneralizethe
algorithmtocatertogeneralcaseofimagereconstructionthisisachallenge
initselfsincetherearetraditionalimageprocessingalgorithmsthatworks
quietefficientlyaswell.ConsiderationofNon-MRIdatatypeprovidesa
generalizationofdifferentdatadistributedthatcouldbeusedtosimulate
otherMRIdatasetthatarehardtoobtain.
29
Chapter3
3.1 ProposedSolutions
Inthispaper,anewmodelisproposedwithwavelettreesparsitytaken
intoconsiderationalongwithentirewavelet“group-coefficients”.Thetree
structureshowninFigure3.1,ismodeledasanoverlappinggroup
regularization,withgroupssharingGrandparent(GP)->Parent(P)->Child(C)
coefficients.Theoriginalproblemof20%samplingratecanbedecomposedto
two-subproblemsandsolvedveryefficientlybyusingtheFISTA[26],[27].
Equation(12)presentsourapproachintermsoftheoptimizationterms.
min!{ F x = !!Ax − b !
! + α x !" + β( ϕx ! + (ϕ!)! !)}!∈! , s. t. t > τ(12)
whereτisprescribedco-efficienttreethresholdforeachlevelofhierarchy.
Subsett ∈TiswhereTisthenumberofhierarchicaltreegroups.
Theequation(12)hasthreeterms,whichspecifythethreecomponentsofthe
algorithm.Thefirsttermensuresreconstructionerroriswithinrange.The
secondtermsenforcessmoothnessontheresult.Thefinaltermenforcessparse
selectiononlyfromthegroupsoftreecoefficientgreaterthanprescribed
thresholdvalue.
30
(a)(b)Figure3.1:WaveletQuad-TreeStructure(a)AcardiacMRimage(b)Corresponding
WaveletTreeGroupofHierarchyGP-P-CStructure
Weproposedthreenewalgorithms,whichwouldexploitthehierarchical
approach.
• WeightedThresholdWaveletTreeStructure(WTWTS)
• DirectionalWaveletTreeStructure(DWTS)
• WeightedDirectionalWaveletTreeStructure(WDWTS)
3.1.1 WeightedThresholdWaveletTreeStructure(WTWTS)
Thehierarchy-basedapproachiscombinedwiththeweighted
thresholdingtoprovideabaselineforthisresearch.Theothertwoalgorithms
arebasedonthisconceptandbuildonitwithdifferenttechniquesusedinthe
shrinkageofembeddedzerotreealgorithm( z!)asshowninequation(13)
whichisconsideredinpseudocodeinTable3.1.
z! = shrinktree((Tϕ!)!,!!, τ)(13)
wherethez!canbefurthersimplifiedto
z! = max ((Tϕ!)! −!!, 0)
31
andwhereϕ! ≥ τandϕ! < τforz!=0suchthatthasthefollowingthreshold
range
Child (C) = 25% < τ ≤ 50%ofCco-efficient
Parent (P) = 25% < τ ≤ 75%ofPco-efficient
GrandParent(GP) = 75%ofGPco-efficient
Table3.1:PseudoCodeofWTWTS
Input:𝑟!=𝑥!!,𝑡!=1,α,β,λ, 𝜏
forn=1toMaxIteration=50do
1. Calculate𝑧! =𝑠ℎ𝑟𝑖𝑛𝑘𝑡𝑟𝑒𝑒 ((𝑇𝜙!)! ,!!, 𝜏)
2. 𝑥! = 𝑟! − 𝑎𝑟𝑔𝑚𝑖𝑛!!{!!𝐴𝑟! − 𝑏 !
! + !!∥ 𝑧! − (𝑇𝜙!)!! ∥}
3. 𝑥! = 𝑔! 𝑥 + 𝛼 ∥ 𝑥!" − 𝑥! ∥!!
4. 𝑥! = 𝑔! 𝑥 + 𝛽 ∥ 𝜙𝑥 − 𝑥! ∥!!
5. 𝑥! = (!!!!!)!
6. 𝑡!!! = [!! !!! !! !
!
7. 𝑟! + 1 = 𝑥! + !!!!
!!!!(𝑥! − 𝑥!!!)
wherehyper-parametersandparametersareexplainedbelow:
α=0.001
β=0.035=positiveparameter
λ=0.2xβ=0.007=multiplier(usedinallconvexmodels)
n=iterationnumber
τ=Weightedthresholdconditionmentionedbelow
Child (C) = 25% < τ ≤ 50%ofCco-efficient
Parent (P) = 25% < τ ≤ 75%ofPco-efficient
GrandParent(GP) = 75%ofGPco-efficient
(Tϕ!)!=Non-overlappinghierarchicaltreeinsparsedomain
32
r!=x!!eachsignalcoefficientvaluefrombasediteration
g! x =α x !"
g! x =β ϕ! !
t!=1,α,β,λ, τallthevaluewhichisinitializedforthefirstiterationandlater
onbasedonthethresholdvaluetheyareiterativelyupdatedfornsamples.
Thesteps5,6,7intheTable3.1arereferredfromWatMRI[29].Resultsshow
theefficacyofourresentedapproach.
3.1.2 DirectionalWaveletTreeStructure(DWTS)
Thesecondapproachworthhighlightinginthispaperistheconceptof
waveletdirectionalityasexpressedinDirectionalFilterBanksTransform
(DFB)[33]asshowninFigure3.2andFigure3.3.TheDirectionalfilter
banksarerealizedusingiteratedquincunxfilterbanks.Thedirectional
filterbankconceptisextendedtomulti-resolutionwaveletsubband
decompositionasbyaddingthefeatureofdirectionalityinanappropriate
mannerimprovementmayoccurinthenonlinearapproximationof
waveletsasshowninFigure3.4.
Figure3.2:Directionalfilterbankfrequencypartitioningusingeightdirections
(a)(b)
Figure3.3:(a)Verticaldirectionalfilterbanks(b)Horizontaldirectionalfilterbanks
33
Figure3.4:ExampleofDirectionalfilterbankonrepresentation
Thesecondmethodutilizesthepriordiscusseddirectionalwavelet
transforms.Directionalityincreasestheefficiencyofthewavelettransforms.
Hencewavelettransformscanbereplacedwithdirectionalwavelets.This
enablesbetterconsolidationofthe‘wavelettrees’andtherebyenablesbetter
sparseoptimizedimagereconstruction.Figure3.5showstheeightdirectional
filterbankonanimage.
Figure3.5:EightDirectionalfilterbanksrepresentusingclassicBarbaraImage
34
ThealgorithmissimilartotheproposedalgorithmWTWTSusingHybrid
WaveletsandDirectionalFilterBanks[33],asbothusethresholdingtoreduce
thepotentialnumberof“candidategroupofTrees”forsparseprocessing.The
reductionofcandidategroups,bythresholdinggroupswithlowscores,enables
sparseselectionfromrichlypopulatedgroups.Shrinkagealgorithmusesthe
eightDirectionalfilterbankwaveletasshowninTable3.2.
Table3.2:PseudoCodeofDWTS
Input:𝑟!=𝑥!!,𝑡!=1,α,β,λ, 𝜏
forn=1toMaxIteration=50do
1. Calculate𝑧! =𝑠ℎ𝑟𝑖𝑛𝑘𝑡𝑟𝑒𝑒 ((𝑇𝜙!)! ,!!, 𝜏)
2. 𝑥! = 𝑟! − 𝑎𝑟𝑔𝑚𝑖𝑛!!{!!𝐴𝑟! − 𝑏 !
! + !!∥ 𝑧! − (𝑇𝜙!)!! ∥}
3. 𝑥! = 𝑔! 𝑥 + 𝛼 ∥ 𝑥!" − 𝑥! ∥!!
4. 𝑥! = 𝑔! 𝑥 + 𝛽 ∥ 𝜙𝑥 − 𝑥! ∥!!
5. 𝑥! = (!!!!!)!
6. 𝑡!!! = [!! !!! !! !
!
7. 𝑟! + 1 = 𝑥! + !!!!
!!!!(𝑥! − 𝑥!!!)
whereDirectionalWaveletisconsidered.
Hyper-parametersandparametersareexplainedbelow:
α=0.001
β=0.035=positiveparameter
λ=0.2xβ=0.007=multiplier(usedinallconvexmodels)
n=iterationnumber
τ=Weightedthresholdconditionmentionedbelow
Child (C) = 25% < τ ≤ 50%ofCco-efficient
35
Parent (P) = 25% < τ ≤ 75%ofPco-efficient
GrandParent(GP) = 75%ofGPco-efficient
(Tϕ!)!=Non-overlappinghierarchicaltreeinsparsedomain
r!=x!!eachsignalcoefficientvaluefrombasediteration
g! x =α x !"
g! x =β ϕ! !
t!=1,α,β,λ, τallthevaluewhichisinitializedforthefirstiterationandlater
onbasedonthethresholdvaluetheyareiterativelyupdatedfornsamples.
Thesteps5,6,7intheTable3.2arereferredfromWatMRI[29].Resultsshow
theefficacyofourresentedapproach.
3.1.3 WeightedDirectionalWaveletTreeStructure(WDWTS)
Inthefinalapproach,inadditiontothedirectionalwavelet,weenforce
thresholdconstraintonthedirectionality.Inadditiontowaveletthresholding,
newestalgorithmalsoincorporatesdirectionalwaveletthresholding(DWTS).
Asseeninthelasttermofequation(12),boththresholdingofwavelet
coefficientgroup’sdirectionalityandenergyareconsidered.Theshrinkage
algorithminequation(13)ismodifiedwithweighteddirectionalcondition.
Table3.3showsthepseudocodeofWDWTSalongwiththeconditionbelow:
36
Table3.3:PseudoCodeofWDWTS
Input:𝑟!=𝑥!!,𝑡!=1,α,β,λ, 𝜏
forn=1toMaxIteration=50do
1. Calculate𝑧! =𝑠ℎ𝑟𝑖𝑛𝑘𝑡𝑟𝑒𝑒 ((𝑇𝜙!)! ,!!, 𝜏)
2. 𝑥! = 𝑟! − 𝑎𝑟𝑔𝑚𝑖𝑛!!{!!𝐴𝑟! − 𝑏 !
! + !!∥ 𝑧! − (𝑇𝜙!)!! ∥}
3. 𝑥! = 𝑔! 𝑥 + 𝛼 ∥ 𝑥!" − 𝑥! ∥!!
4. 𝑥! = 𝑔! 𝑥 + 𝛽 ∥ 𝜙𝑥 − 𝑥! ∥!!
5. 𝑥! = (!!!!!)!
6. 𝑡!!! = [!! !!! !! !
!
7. 𝑟! + 1 = 𝑥! + !!!!
!!!!(𝑥! − 𝑥!!!)
wherehyper-parametersandparametersareexplainedbelow:
α=0.001
β=0.035=positiveparameter
𝜆=0.2xβ=0.007=multiplier(usedinallconvexmodels)
n=iterationnumber
τ=Weightedthresholdconditionmentionedbelow
s. t. t 2,3,6,7 Inner Direction = 75% of co− efficient1,4,5,8 Outer Direction = 25% of co− efficient P,C
P=Parent
C=Child
(Tϕ!)!=Non-overlappinghierarchicaltreeinsparsedomain
r!=x!!eachsignalcoefficientvaluefrombasediteration
g! x =α x !"
g! x =β ϕ! !
𝑡!=1,α,β,λ, 𝜏allthevaluewhichisinitializedforthefirstiterationandlater
onbasedonthethresholdvaluetheyareiterativelyupdatedfornsamples.
37
3.2 ComplexityAnalysis
Supposewerepresentxasanimagewithnpixels,overallcomplexity
wouldbeO(n’+nlogn)wheren’isnumberofnon-zerohierarchicaltree
element.SothatTisnon-overlappinghierarchicaltreewouldhaveO(n’).
BelowisacomplexityofeachstepinthreealgorithmnamelyWTWTS,DWTS,
WDWTS.
Step1:O(n’+nlogn)
Step1showsthecomplexityforWTWTS,DWTSandWDTWTS.Notethat
weconsidered8directionsforDWTS,WDWTSitwouldbe
O(n’+8nlogn)≈O(n’+nlogn).
Step2:O(nlogn)
Step3:O(n)
Step4:O(nlogn)
Step5:O(n)
Step6:O(n)
Step7:O(n)
Resultsinthenextsectioncomparethethreeproposedtechniquesandthe
currentbenchmarktechnique.
38
Chapter4
4.1 Experimentsandresults
Numerousvarietiesofexperimentshavebeenconductedfordifferent
parametersandhyperparametersselection.Thissectionalsoshowsfinalresult
oftheproposedalgorithmsperformancecomparedwiththeWatMRI[29].Inthe
MRimagingproblem,AisthepartialFouriertransformwithmrowsandn
columns.Thisstudyconsidered20%sampling.MRscanningtimewillbe
reducediflessmeasurementissampled.Allmeasurementscontain0.01Gaussian
WhiteNoise.SignalToNoiseRatio(SNR)andMeanSquareError(MSE)are
consideredforevaluation.
4.1.1 TestEnvironmentSpecification
Detaileddescriptionofthetestenvironmentthatwasusedforthethesisis
mentionedbelow:
• HardwareSpecification:
o System:MacOSv10.13.1
o Processor:2.4GHzIntelCorei5
o Memory:8GB1333MHzDDR3(DoubleDataRateSynchronous
DynamicRandom-AccessMemory)
o Graphics:IntelHDGraphics3000512MB
• ProgrammingLanguage:MATLABR2017a(Aproprietaryprogramming
languagedevelopedbyMathWorks)
4.1.2 DataSet
BothMRIandNon-MRIdatatypesareconsideredtovalidatethis
39
approach.500setsofeachdatatypeswithdifferentsizefrom16x16,32x32,
64x64,128x128,and256x256resolutionsareconsidered.
MRIData:
MRIdatawasobtainedfromthereferencepaper[29].Brain,heart,
shoulder,chestimagesareconsideredasshownintheFigure4.1.
Figure4.1:MRIData
ThesamplingmaskthatisappliedforspatialfilterisasshowninFigure4.2.
Figure4.2:SamplingMask
ThetwobasictypesofMRIimagesareT1-weightedandT2-weightedimages,
oftenreferredtoasT1andT2images.
T1images–highlightfattissuewithinthebodyasshowninFigure4.3(a)
T2images–highlightfattissueandwatercontentwithinthebodyasshownin
Figure4.3(b)
40
(a)(b)
Figure4.3:(a)T1imageand(b)T2image
Experimentsalsoconsidered50imagesofT1,T2,whichrepresentdifferent
detailsinthebraindataoftoidentifyifthereisanyvariationbasedonthe
directionofthedata.Thedatasetwasobtainedfromgithublinkasshownbelow:
https://github.com/muschellij2/Neurohacking_data/tree/master/BRAINIX/DIC
OM
NonMRIData:
Non-MRIdatatypeswerepickedupfromGoogleImagebasedonfollowing
criteriasuchas
• PortraitImages–Thisdatatyperepresentsimagesthathaveboth
combinationsoflowandhighpasssignalsthatmeantheonlyfewofthe
datacanbesparselyrepresented.FilesarenamedasLena,Obama,
Einstein,andCameramanthatisJPGfilesasshowninFigure4.4.
Figure4.4:PortraitData
41
• CrowdImages–Thisdatatyperepresentsimagesthathavemajorlyhigh
passsignalthatmeanstherearelotsofvariationsinthedata.Filesare
namedasCrowd1,Crowd2,Crowd3andCrowd-riceimagesthatareJPG
filesasshowninFigure4.5.
Figure4.5:CrowdData
• LandscapeImages–Thisdatatyperepresentsimagesthathavemajority
ofthelowpasssignalsthatindicatesthatsomeofthedataareredundant
whichcouldbesparselyrepresented.FilesarenamedasLandscape1,
Landscape2,andLandscape3imagesasshowninFigure4.6.
Figure4.6:LandscapeData
4.1.3 Hyper-Parameters
Eachsimulationonthedataiscarriedoutwith50iterationswithapproximate
20%sampling.WaveletTreeStructureisconfiguredtochoosetheGrandparent,
42
parentandchildrelationshipthataregeneralizedforCompressedSensingin
MRI.Differentthresholdingsetisalsoconsideredashyper-parametersthatare
configured.
4.1.4 Parameters
WaveletTreeGroupingiscustommadebasedonwhichofthewaveletco-
efficientcontributestowardsaccurateimagereconstruction.WaveletCo-
efficientchangesarebasedonthethresholdingsetsthatneedtobeiteratively
updatedtoprovideco-efficientstoimprovethereconstructionoffinalimages.
4.1.5 PreliminaryResults
4.1.5.1 SelectionofThresholdset,ImagesizeandthresholdbasedonWaveletDirections
Numberofzero-tree(zerovalueco-efficientatalllevel)variedagainst
rangeofthreshold,therewasnoparticularonevalue.Ithasbeendifferentfor
eachdataofMRIandnon-MRIdata.
Figures4.7to4.10showsvariationsofthenumberofzerotrees
occurrence(eliminatedduringshrinkage)ateachhierarchylevelwithdifferent
setsofdatadisplaysaspecificrangeofthresholdτwhichnormallyisbetween
0-100.
43
Figure4.7:VariationofnumberofzerotreesbasedonGP-P-CHierarchyforonebrain
image
Figure4.8:VariationofnumberofzerotreebasedonGP-P-CHierarchyforT1images
44
Figure4.9:VariationofnumberofzerotreesbasedonGP-P-ChierarchyforT2image
Figure4.10:VariationofnumberofzerotreesbasedonGP-P-Chierarchylevelforentire
500brainimages
45
Thevariationsofnumberofzerotreesoccurrenceinthedatawereattributed
tovariationintheimage.Tocomparetheentropydistributionoftheimage
basedonthelevelwouldgiveabetterunderstandofdistributionpatternin
eachlevelirrespectiveofthetypeofimagesasshownintheFigure4.11.
Figure4.11:3DGraphofrangeofthresholdsetsagainstEntropyofdataatParent-Child
Level
TheParentandChildleveldistributionledtousingtheconceptsofweighted
thresholdingsincethereisnoonethresholdvaluethatsatisfiesallthedata
type.Thereshouldbedifferentrangeofthresholdsettocatertodifferent
distributioninlevel.Henceweightedtreestructureofdifferentcombinationof
ratioofthresholdisbeingconsidered.(GP,P,C)thresholdinpercentagesare:
• 75,75,50–Set1
• 75,75,25–Set2
• 75,50,50–Set3
• 75,50,25–Set4
46
• 75,25,25–Set5
Figures4.12to4.15showscomparisonofthe5setsofthresholdvaluesagainst5
setsofdifferentresolutionsize(16,32,64,128,256).
Figure4.12:ComparisonofdifferentthresholdingsetsagainstSNRfordifferentresolution
sizesforBrainImages(RedlineindicatereferenceSNRofWatMRIalgorithm)
Figure4.13:ComparisonofdifferentthresholdingsetsagainstSNRfordifferentresolution
sizesforHeartImages(RedlineindicatereferenceSNRofWatMRIalgorithm)
Figure4.14:ComparisonofdifferentthresholdingsetsagainstSNRfordifferentresolution
sizesofShoulderImages(RedlineindicatereferenceSNRofWatMRIalgorithm)
47
Figure4.15:ComparisonofdifferentthresholdingsetsagainstSNRfordifferentresolution
sizesforChestImages(RedlineindicatereferenceSNRofWatMRIalgorithm)
Itwasfoundthat256x256wasthemoststablecomparedtoothers.Inregards
tothresholding,itwasfoundthatSet2and3gavethebestandmostreliable
resultswithhigherSNRvaluecomparedtoWatMRIasshownintheFigure
4.16.
Figure4.16:SNRcomparisonbetweenWatMRI(Greencolor)andWTWTSwithSet2(Red
color)and3(Bluecolor)approaches
48
Figure4.17showsthebehaviorofnon-MRIdataagainstthethresholdingset2
and3whencomparedtothereferencealgorithm.
Figure4.17:ComparisontheSNRforPortrait,CrowdandLandscapeimagesforthresholdingset2and3(RedlineindicatesreferenceSNRofWatMRIalgorithm)
Figures4.18to4.21showsthevariationofhybridwaveletcoefficientsinthe8
directionsforeachMRIdata.Direction2,3,6,and7thatisconsideredasinner
directionhashighercoefficientsthandirection1,4,5and8thatisconsideredas
outerdirection.ThisistheconditionforthealgorithmWDWTStochoosethe
weighteddirectionalthresholdingintheParentandChildlevelwhileallvaluesof
GrandParentlevelwillbeconsideredsinceallthecoefficientscontributetowards
theimagereconstruction.
Figure4.18:Histogramof8DirectionsatParentandChildlevelforBrainImage
49
Figure4.19:Histogramof8DirectionsatParentandChildlevelforHeartImage
Figure4.20:Histogramof8DirectionsatParentandChildlevelforShoulderImage
50
Figure4.21:Histogramof8DirectionsatParentandChildlevelforChestImage
Aboveresultsshowsthatexperimentswerecarriedouttoselectthresholdsets,
imagesizeandthresholdfordirectionalwaveletonMRIdata.Theseexperiments
werecarriedoutnon-MRIdataobtaininggoodresultthatcouldbegeneralized
foranytypeofimages.Figures4.22to4.24showthePortrait,Crowdand
Landscapeimagesresultsbasedontheselectionmadeforthresholdset,image
sizeandthresholdonwaveletdirection.
51
Figure4.22:ComparisonofPortraitimageagainstoriginaldatausingWTWTSand
WDWTSalgorithmswithchosenthresholdset
Figure4.23:ComparisonofCrowdimageagainstoriginaldatausingWTWTSandWDWTSalgorithmswithchosenthresholdset
52
Figure4.24:ComparisonofLandscapeimageagainstoriginaldatausingWTWTSandWDWTSalgorithmswithchosenthresholdset
4.1.6 ResultsandEvaluation
Thissectiondisplaystheevaluationofproposedalgorithmsincorporating
thefindingofhyper-parameterssuchasimagesize,thresholdsets,thresholdfo
weighteddirectionetc.
Figures4.25to4.31belowshowtheresultcomparisonofthethreenew
algorithmsdevelopedasthepartofthisthesisagainstreferencepaperWatMRI.
EachdatatypehasitsSignaltoNoiseRatio(SNR)valuesandMeanSquare
Error(MSE)
53
Figure4.25:Brainimagereconstructionusingthe3algorithmscomparedagainstWatMRI
Figure4.26:Heartimagereconstructionusingthe3algorithmscomparedagainstWatMRI
54
Figure4.27:Shoulderimagereconstructionusingthe3algorithmscomparedagainst
WatMRI
Figure4.28:Chestimagereconstructionusingthe3algorithmscomparedagainstWatMRI
55
Figure4.29:Portraitimagereconstructionusingthe3algorithmscomparedagainst
WatMRI
Figure4.30:Crowdimagereconstructionusingthe3algorithmscomparedagainst
WatMRI
56
Figure4.31:Landscapeimagereconstructionusingthe3algorithmscomparedagainst
WatMRI
Table4.1showstheSNRandMSEcomparisonofMRIdataforallthethree
algorithmsagainstWatMRI.
Table4.1:SNRandMSEcomparisononMRIdata
Method
MRIDataTypes
Brain Heart Shoulder Chest
SNR MSE SNR MSE SNR MSE SNR MSE
WatMRI 19.73 28.74 20.80 16.32 23.61 6.36 20.39 63.68
WTWTS 19.98 27.16 20.87 16.07 23.98 5.85 20.41 63.35
DWTS 20.54 23.86 21.47 13.99 23.99 5.80 20.44 62.97
WDWTS 20.83 22.33 21.69 13.32 25.07 4.54 21.06 54.60
57
Table4.2showtheCPUtimecomparisononMRIdatasetsforallthethree
algorithmsagainstWatMRI.
Table4.2:CPUTimecomparisononMRIdata
Method
MRIDataTypes
Brain Heart Shoulder Chest
CPUTime(s) CPUTime(s) CPUTime(s) CPUTime(s)
WatMRI 1.50±0.03 1.53±0.05 1.60±0.04 1.59±0.03
WTWTS 1.38±0.09 1.35±0.1 1.43±0.11 1.41±0.09
DWTS 1.40±0.08 1.43±0.09 1.50±0.07 1.47±0.05
WDWTS 1.42±0.08 1.66±0.05 1.70±0.6 1.51±0.11
Table4.3showtheSNRandMSEcomparisononNon-MRIdatasetsforallthe
threealgorithmsagainstWatMRI.
Table4.3:SNRandMSEcomparisononNon-MRIdata
Method
NonMRIDataTypes
Portrait Crowd Landscape
SNR MSE SNR MSE SNR MSE
WatMRI 23.40 80.10 14.26 420.3 19.81 104.05
WTWTS 24.19 66.90 14.46 401.56 20.11 97.09
DWTS 24.37 64.15 14.46 401.07 20.15 96.14
WDWTS 24.67 59.88 14.70 379.26 20.20 94.98
58
Table4.4showtheCPUtimecomparisononNon-MRIdatasetsforallthethree
algorithmsagainstWatMRI.
Table4.4:CPUTimecomparisononNon-MRIdata
Method
Non-MRIDataTypes
Portrait Crowd Landscape
CPUTime(s) CPUTime(s) CPUTime(s)
WatMRI 1.50±0.01 1.60±0.03 1.55±0.04
WTWTS 1.45±0.09 1.50±0.05 1.42±0.08
DWTS 1.49±0.07 1.53±0.06 1.47±0.04
WDWTS 1.55±0.12 1.61±0.05 1.49±0.08
4.1.7 ComparisonbetweenWTWTS,DWTS,WDWTS
1. AllthreealgorithmsnamelyWTWTS,DWTS,WDWTShaveshownbetter
qualityofreconstructedimagerepresentedbySNRvalueandreductionin
MSEvaluethanthereferencepaper(WatMRI[29])result.
2. WTWTSproduced≈4%imagequalitywith≈8%reductioninalgorithm
runthetimeofthealgorithm.(≈symbolrepresentsapproximation)
3. DWTSproduced≈10%qualitywithreductionof≈4%inruntime.
4. DWTSalsoconsumedmorememoryspaceforstoringdirectional
informationinthetreestructure
5. WDWTSproduced≈20%betterimagequalitywith≈6%reductioninrun
timeofthealgorithm.
6. TradeoffofWDWTSisthememoryspaceconsumptionforweighted
directionaltreestructure.
59
Chapter5
FutureWork
Thisresearchcanbetakenforwardtoimprovetheresults.Ideasareasfollows:
• UseDeepNeuralNetworktotrainmachinetolearnthealgorithm
withoutanyassistancefromradiologistsandtrainthereconstruction
errorbackintothenetworktominimizetheerror.Theoverviewisas
shownFigure5.1.
Figure5.1:OverviewofDeepNeuralNetworkforCSinMRI
TrainReconstructionError:Flowchartofthetrainingreconstructionerroris
showninFigure5.2.
60
Figure5.2:ReconstructionErrorflowchart
• UpdatealgorithmforMulti-slice3DMRI–MultiplesliceofDataforone
directionalimagecanbeconsideredoreven3DMRIdatacanbe
consideredasshowninFigure5.3.Thethreecrosssectionaldirectional
ofthebrainimageisrepresentedinFigure5.4.
Figure5.3:Multi-SliceofBrainMRI[38]
ReconstructionErrors
TrainNeuralNetwork
MinimizeReconstruction
Errors
61
Figure5.4:3DBrainMRIdisplayingeachdimension[40]
DynamicMRIwithTensorwillbeusedtoworkon3DMRIdataset.Eachofthe
dimensionsisconsideredtobe2Ddatasetalongwithtime.Figure5.5shows
theoverviewof3DMRItoTensor(3Directionalaxisorequivalentofvolume
datastructure)[37,38].
Figure5.5:DynamicMRIrepresentedinTensor[46]
62
TheformulaforTensorissimilartoCSwithSVDbutthedatais3D[45].Sothe
needistoconsiderHigherOrderSingleValueDecomposition(HOSVD)as
showninequationbelow
𝑀 = 𝑈𝑆𝑉! = 𝑆 ×!𝑈!×!𝑉! = 𝑆 ×!𝑈!×!𝑈!(14)
whereU&V=unitarymatrices,whileS=pseudo-diagonalmatrixwhichis
representinequation(15)
𝑆 = 𝐴!×!𝑈!!×!𝑈!!×!…… .×!𝑈!! (15)
whereA0=nthordertensorunfoldedmatrixasshownbelowinequation(16):
𝑖!!! − 1 𝐼!!!𝐼!!!…… . . 𝐼!𝐼!𝐼!…… 𝐼!!! +
𝑖!!! − 1 𝐼!!!𝐼!!!…… . . 𝐼!𝐼!…… 𝐼!!! +⋯+
𝑖! − 1 𝐼!𝐼!…… . . 𝐼!!! + 𝑖! − 1 𝐼!𝐼!… . . 𝐼!!! +⋯+ 𝑖!!!(16)
Figure5.6representsthe3Dtensorand4DtensorusedforSVDcalculation.
Thisiscommonlyusedink-tFOCUSS[42,43]methodinmedicalcommunity.
Figure5.6:VisualizationofHigherOrderSingleValueDecomposition(HOSVD)(a)3D
Tensor(b)4DTensor[45]
63
Chapter6
Conclusion
TheclearimpetusofchoosingCompressiveSensing(CS)approachisto
utilizesignificantlesssamplingwhilegeneratingahighqualityreconstruction.
RecentmethodscanreconstructMRimageswithgoodqualityfrom
approximate20%sampling.Theseresultshavealsoshownsuchtrendsaswell.
Theproposedmethodsshowimprovementoverthecurrentbenchmark
algorithminthisfield.Althoughsomealgorithmshavebeenproposedto
improvestandardCSrecoverybyutilizingthetreestructureofwavelet
coefficients[10][11][12][13],noneofthemexploitthetreesparsitytoany
significantinstantandonlyafewofthemconductexperimentsonMRimagesto
validatethepracticalbenefitforacceleratedMRI[51][52].
ThisresearchhasfuturefocustobecombinedwithDeepNeuralNetwork
andMulti-slicingof3DMRItothatwouldincorporatemulti-dimensionaldata
intoCompressedSensingforMRI.Thissurelywouldcontributetotheexisting
research.
64
Bibliography[1] W.C.Shiel,“MRIscan”,https://www.medicinenet.com/mri_scan
[2] FigureisadaptedfromSingaporeSportsandOrthopedicClinic.Link:
https://www.orthopaedics.com.sg/treatments/other-services/magnetic-
resonance-imaging-mri/
[3] M.A.Qureshi,M.Deriche,“Anewwaveletbasedefficientimage
compressionalgorithmusingcompressivesensing”,articleinMutlimedia
ToolsandApplication,2015.
[4] H.Nyquist,"Certaintopicsintelegraphtransmissiontheory".Trans.
AIEE.47:617–644.(April1928)
[5] C.E.Shannon,"Communicationinthepresenceofnoise".Proceedingsof
theInstituteofRadioEngineers.37(1):10–21.(January1949)
[6] C.E.Shannon,"AMathematicalTheoryofCommunication".BellSystem
TechnicalJournal.27(4):623–666.(October1948).
[7] E.T.Whittaker,"OntheFunctionswhicharerepresentedbytheexpansions
oftheInterpolationTheory".Proc.RoyalSoc.Edinburgh.35:181–
194.(1915).
[8] R.G.Baraniuk(2007)Compressivesensing[lecturenotes].Signal
ProcessingMagazine,IEEE24(4):118–121
[9] R.G.Baraniuk,E.Candes,M.Elad,Y.Ma(2010)“Applicationsofsparse
representationandcompressivesensing“.ProceedingsoftheIEEE98(6):
906–909
[10] E.Candes,T.Tao,“Near-optimalsignalrecoveryfromrandomprojections:
Universalencodingstrategies?”IEEETransactionsonInformationTheory
52(12):5406–5425.(2006)
65
[11] E.Candes,M.B.Wakin,“Anintroductiontocompressivesampling”.Signal
ProcessingMagazine,IEEE25(2):21–30.(2008)
[12] E.Candes,J.Romberg,T.Tao,“Robustuncertaintyprinciples:Exactsignal
reconstructionfromhighlyincompletefrequencyinformation”.IEEE
TransactionsonInformationTheory52(2):489–509.(2006)
[13] E.Candes,J.Romberg,T.Tao,“Stablesignalrecoveryfromincompleteand
inaccuratemeasurements.”Communicationsonpureandapplied
mathematics59(8):1207–1223.(2006)
[14] D.L.DonohoD.L.Compressedsensing.IEEETransactionsonInformation
Theory52(4):1289–1306.(2006)
[15] E.Candes,J.RombergandT.Tao,“Robustuncertaintyprincipals:Exact
signalreconstructionfromhighlyincompletefrequencyinformation,”IEEE
Trans.Inf.Theory,vol.52,pp.489-509,2005.
[16] E.Candes,T.Tao,“Decodingbylinearprogramming.”IEEETransactionson
InformationTheory51(12):4203–4215.(2005)
[17] E.Candes,J.Romberg,“Sparsityandincoherenceincompressivesampling.
Inverseproblems”.23(3):969.(2007)
[18] G.H.GolubandD.P.O’Leary,Somehistoryoftheconjugategradient
methodsandLanczosalgorithms:1948-1976,SIAMRev.,31(1989),pp.
50–100.
[19] M.Lustig,D.Donoho,J.M.SantosandJ.M.Pauly,“Compressedsensing
MRI,”IEEESignalProcess.Mag.,vol.25,no.72,pp.72-82,2008.
[20] M.Lustig,D.Donoho,andJ.M.Pauly,“SparseMRI:Theapplicationof
compressedsensingforMRimaging”,MagneticResonanceinMedicine,vol.
58,no.6,pp.1182-1195,2007.
[21] T.CaiandL.Wang,“OrthogonalMatchingPursuitforSparseSignal
RecoveryWithNoise”,IEEETransactionsonInformationTheory,Volume:
57,Issue:7,July2011
[22] S.MallatandZ.Zhang.MatchingPursuitwithTimeFrequencyDictionaries.
IEEETrans.Sign.Proc.(1993).
66
[23] J.Yang,Y.ZhangandW.Yin,“Afastalternatingdirectionmethodfortvl1-l2
signalreconstructionfrompartialFourierdata,”IEEEJ.Sel.TopicsSignal
Process.vol.4,no.2,pp.288-297,2010.
[24] S.Ma,W.Yin,Y.Zhang,andA.Chakraborty,“Anefficientalgorithmfor
compressedMRimagingusingtotalvariationandwavelets,”Proceedingof
IEEEConferenceonComputerVisionandPatternRecognition,pp.1-8,
2008.
[25] J.Huang,S.ZhangandD.Metaxas,“EfficientMRimagereconstructionfor
compressedMRimaging,”MedicalImageAnalysis,vol.15,no.5,pp.670-
679,2011.
[26] A.BeckandM.Teboulle,“Afastiterativeshrinkage-thresholdingalgorithm
forlinearinverseproblems,”SIAMJ.onImagingsciences,vol.2,no.1,pp.
183-202,2009.
[27] A.BeckandM.Teboulle,“Fastgradient-basedalgorithmsforconstrained
totalvariationimagedenoisinganddeblurringproblems,”IEEETrans.on
ImageProcess.,vol.18,no.113,pp.2419-2434,2009.
[28] C.ChenandJ.Huang,“ThebenefitoftreesparsityinacceleratedMRI”,
MICCAIworkshop2012.
[29] C.ChenandJ.Huang“CompressivesensingMRIwithwavelettree
sparsity,”proceedingofthe26thAnnualConferenceonNeuralInformation
ProcessingSystems,pp.1115-1123,2012.
[30] J.Huang,T.ZhangandD.Metaxas,“Learningwithstructuredsparsity,”J.of
MachineLearningResearch,vol.12,pp.3371-3412,2011.
[31] R.G.Baraniuk,V.Cevher,M.F.DuarteandC.Hegde,“Model-Based
compressivesensing,”IEEETrans.onInf.Theory,vol.56,no.4,pp.1982-
2001,2010.
[32] F.Back,R.Jenatton,andG.Obozinski,“Structuredsparsitythroughconvex
optimization,”TechnicalReport,HAL00621245-v2,toappearonstatistical
science,2012.
[33] R.EslamiandH.Radha,“NewImageTransformsUsingHybridWavelets
67
andDirectionalFilterBanks:AnalysisandDesign”,ICIP,2005.
[34] Chen,J.Huang,“Exploitingthewaveletstructureincompressedsensing
MRI”,MagneticResonanceImaging32(2014)1377–1389.
[35] N.Kingsbury,“Imageprocessingwithcomplexwavelets,”Philosophical
TransactionsoftheRoyalSocietyofLondonA,no.357,pp.2543–2560,
1999.
[36] F.Fernandes,“Directional,shift-insensitive,complexwavelettransforms
withcontrollableredundancy,”Ph.D.thesis,RiceUniversity,2002.
[37] Y.Kim,M.Altback,T.Trouard,andA.Biligin,“Compressedsensingusing
dualtreecomplexwavelettransform”ProceedingoftheInternational
SocietyforMagneticResonanceinMedicine,p.2814,2009.
[38] VRunge,JRichter,TBeck,MPiccirelli,AValavanis,“SimultaneousMulti-
Slice(SliceAccelerated)InstituteforDiagnosticandInterventional
RadiologyClinicsforNeuroradiologyandNuclearMedicineUniversity
HospitalZurich.
[39] J.Zamani,A.N.Moghaddam,H.S.Rad,“MRIReconstructionThrough
CompressedSensingUsingPrincipleComponentAnalysis(PCA)”,ICEEMay
2012.
[40] ImagefromHitachiproductgeneratedaNeurologicalMRIusingisoFSE.
[41] DiagramfromMathiasBlasche,ChristophForman,SiemensMAGNETOM
Flash,2016.
[42] H.Jung,J.C.Ye,E.Y.Kim,(2007),“Improvedk-tBLASTandk-tSENSEusing
FOCUSS”.Physicsinmedicineandbiology52:3201–3226.
[43] H.Jung,K.Sung,K.SNayak,E.Y.Kim,J.C.Ye,(2009),“k-tFOCUSS:Ageneral
compressedsensingframeworkforhighresolutiondynamicMRI”.
MagneticResonanceinMedicine61:103–116.
[44] M.Hong,Y.Yu,H.Wang,F.Liu,andS.Crozier,"CompressedsensingMRI
withsingularvaluedecomposition-basedsparsitybasis,"Physicsin
MedicineandBiology,vol.56,p.6311,2011.
[45] Y.Yeyang,J.Jin,L.Feng,S.Crozier,MultidimensionalCompressedSensing
68
MRIUsingTensorDecomposition-BasedSparsifyingTransform”,PLosone
Volume9,Issue6,e98441.(June2017)
[46] Y.Xiaomei,L.Yuewan,C.Siji,Z.Xiujuan,Y.Qin,L.Kai,“DynamicMRI
reconstructionfromhighlyundersampled(k,t)-spacedatausingweighted
Schattenp-normregularizeroftensor”,MagneticResonanceImaging37
260–272.(2017).
[47] https://www.radiologymasterclass.co.uk/tutorials/mri/t1_and_t2_images
[48] http://multiple-sclerosis-research.blogspot.com/2015/01/the-only-thing-
that-is-black-and-white.html
[49] https://github.com/muschellij2/Neurohacking_data/tree/master/BRAINI
X/DICOM
[50] M.H.Conde,“FundamentalsofCompressiveSensing”,Chapter3from
CompressiveSensingforthePhotonicMixerDevice,DOI:10.1007/978-3-
658-18057-7_3
[51] ThisthesiswaspresentedatGeorgiaCouncilofGraduateSchools(GCGS
April2018)asapartof3MinuteThesis
[52] S.S.JanneyandS.Chakravarty,“NewAlgorithmforCS-MRI:WTWTS,
DWTS,WDWTS”,ResearchPosterpaperacceptedby40thAnnual
InternationalIEEEEMBSConferenceJuly2018
[53] M.F.Jiang,J.Jin,F.Liu,“Sparsity-constrainedSENSEreconstruction:an
efficientimplementationusingafastcompositesplittingalgorithm”
MagneticResonanceImaging.2013;31(7):1218–1227.doi:
10.1016/j.mri.2012.12.003.
[54] J.Huang,S.Zhang,H.Li,D.Metaxas,“Compositesplittingalgorithmsfor
convexoptimization”,ComputerVisionandImageUnderstanding(2011).
doi:10.1016/j.cviu.2011.06.011.
[55] P.L.Combettes,J.-C.Pesquet,Aproximaldecompositionmethodforsolving
convexvariationalinverseproblems,InverseProblems24(2008)1–27.
[56] P.L.Combettes,Iterativeconstructionoftheresolventofasumofmaximal
monotoneoperators,J.ConvexAnal.16(2009)727–748.
69
[57] J.Yang,Y.Zhang,W.Yin,AfastalternatingdirectionmethodforTVL1-L2
signalreconstructionfrompartialfourierdata,IEEEJ.Sel.Top.Signal
Process.4(2)(2010)288–297.
[58] J.Liu,P.Musialski,P.Wonka,J.Ye,Tensorcompletionforestimating
missingvaluesinvisualdata,in:ProceedingsofICCV,2009.
[59] S.Wright,R.Nowak,M.Figueiredo,Sparsereconstructionbyseparable
approximation,IEEETrans.SignalProcess.57(7)(2009)2479–2493.
[60] J.Eckstein,B.F.Svaiter,Generalprojectivesplittingmethodsforsumsof
maximalmonotoneoperators,SIAMJ.ControlOptim.48(2009)787–811.
[61] D.Gabay,B.Mercier,Adualalgorithmforthesolutionofnonlinear
variationalproblemsviafinite-elementapproximations,Comput.Math.
Appl.2(1976)17–40.
[62] R.Glowinski,P.L.Tallec,AugmentedLagrangianandoperator-splitting
methodsinnonlinearmechanics,SIAMStud.Appl.Math.9(1989).
[63] P.Tseng,Applicationsofasplittingalgorithmtodecompositioninconvex
programmingandvariationalinequalities,SIAMJ.ControlOptim.29
(1991)119–138.
[64] B.S.He,L.-Z.Liao,D.Han,H.Yang,Anewinexactalternatingdirection
methodformonotonevariationalinequalities,Math.Program.92(2002)
103–118.
[65] S.Vasanawala,M.Alley,R.Barth,B.Hargreaves,J.Pauly,andM.Lustig.
FasterpediatricMRIviacompressedsensing.InProc.AnnualMeetingSoc.
PediatricRadiology(SPR),Carlsbad,CA,Apr.2009.
[66] M.YuanandY.Lin.Modelselectionandestimationinregressionwith
groupedvariables.J.Roy.Stat.Soc.Ser.B,68(1):49{67,2006.
[67] M.Davenport.Randomobservationsonrandomobservations:Sparse
signalacquisitionandprocessing.PhDthesis,RiceUniversity,Aug.2010.
[68] Y.Chi,L.Scharf,A.Pezeshki,andR.Calderbank.Sensitivitytobasis
mismatchincompressedsensing.Preprint,May2010.
70
[69] A.Cohen,W.Dahmen,andR.DeVore.Instanceoptimaldecodingby
thresholdingincompressedsensing.InInt.Conf.HarmonicAnalysisand
PartialDifferentialEquations,Madrid,Spain,Jun.2008.
[70] A.Cohen,W.Dahmen,andR.DeVore.Compressedsensingandbestk-term
approximation.J.Amer.Math.Soc.,22(1):211{231,2009.
[71] S.Mun,J.E.Fowler,“Blockcompressedsensingofimagesusingdirectional
transforms.”,IEEEInternationalConferenceonImageProcessing(ICIP),
IEEE,pp.3021–3024.(2009)
[72] E.Candes,J.Romberg,“Sparsityandincoherenceincompressivesampling.
Inverseproblems”,23(3):969.(2007)
[73] N.Ahmed,T.Natarajan,K.R.Rao,“Discretecosinetransform.”,IEEE
TransactionsonComputers100(1):90–93.(1974)
[74] J.Trzasko,A.Manduca,“AUnifiedTensorRegressionFrameworkfor
CalibrationlessDynamic,Multi-ChannelMRIReconstruction”;2013;Salt
LakeCity,USA.(2013).
[75] C.Prieto,M.Usman,J.M.Wild,S.Kozerke,P.G.Batchelor,“Groupsparse
reconstructionusingintensity-basedclustering.”MagneticResonancein
Medicine69:1169–1179.(2012)
[76] WangS-J,SunM-F,ChenY-H,PangE-P,ZhouC-G(2012)STPCA:sparse
tensorprincipalcomponentanalysisforfeatureextraction.Pattern
Recognition(ICPR),201221stInternationalConferenceon.TsukubaJapan.
2278–2281.
[77] P.Comon,“Tensordecompositions.”,MathematicsinSignalProcessingV:
1–24.
[78] D.Liang,EVR.DiBella,R.RChen,L.Ying“K-TISD:Compressedsensingwith
iterativesupportdetectionfordynamicMRI.BiomedicalImaging:From
NanotoMacro,2011IEEEInternationalSymposiumon:IEEE.1264–1267.
(2011).
[79] M.Akcakaya,T.A.Basha,B.Goddu,L.A.Goepfert,K.V.Kissinger,“Low-
dimensional-structureself-learningandthresholding:Regularization
71
beyondcompressedsensingforMRIReconstruction.”,MagneticResonance
inMedicine66:756–767.(2011)
[80] T.G.Kolda,B.W.Bader,“Tensordecompositionsandapplications.”,SIAM
review51:455–500.(2009)
[81] F.L.Hitchcock,“Multipleinvariantsandgeneralizedrankofap-waymatrix
ortensor.”JournalofMathematicsandPhysics7:39–79.(1927)
[82] WangS-J,YangJ,SunM-F,PengX-J,SunM-M,etal.(2012)Sparsetensor
discriminantcolorspaceforfaceverification.NeuralNetworksand
LearningSystems,IEEETransactionson23:876–888.
[83] W.F.Sacco,H.A.Filho,P.Claudio,“Cost-basedoptimizationofnuclear
reactorcoredesign:Apreliminarymodel”,2007
[84] A.K.Cline,I.S.Dhillon,“ComputationoftheSingularValueDecomposition”,
bookchapter45inHandbookofLinearAlgebra,TheUniversityofTexasat
Austin.
72
Appendix
ListofAlgorithms
Aspartofthisthesis,someofthewell-knownalgorithmsareusedandsomealso
experimentedwith.Herearethelistofalgorithmsnamesandtheirpseudocode.
Allthesepseudocodeisreferencefrom[54].
1) IterativeShrinkage-ThresholdingAlgorithm(ISTA)[26]
Algorithm1.ISTA
Input:𝜌 = 1 𝐿! , 𝑥!
repeat
fork=1toKdo
𝑥! = 𝑝𝑟𝑜𝑥!(𝑔)(𝑥!!! − 𝜌∇𝑓(𝑥!!!))
endfor
untilStopcriterions
2) FastIterativeShrinkage-ThresholdingAlgorithm(FISTA)[26]
Algorithm2.FISTA
Input:𝜌 = 1 𝐿! ,𝑟! = 𝑥!,𝑡! = 1
repeat
fork=1toKdo
𝑥! = 𝑝𝑟𝑜𝑥!(𝑔)(𝑟! − 𝜌∇𝑓(𝑟!))
𝑡!!! = !! !!!(!!)!
!
𝑟!!! = 𝑥! + !!!!!!!!
(𝑥! − 𝑥!!!)
73
endfor
untilStopcriterions
3) CompositeSplittingDenoising(CSD)[55,56]
Algorithm3.CSD
Input:𝜌 = 1 𝐿,𝛼,𝛽,{𝑧!!}!!!,….,! = 𝑥!
repeat
forj=1toJdo
fori=1tomdo
𝑥! = 𝑎𝑟𝑔𝑚𝑖𝑛! !!!"
||𝑥 − 𝑧!!!!||! + 𝑔!(𝐵!𝑥)
endfor
𝑥! = !!
𝑥!!!!!
fori=1tomdo
𝑧!! = 𝑧!
!!! + 𝑥! − 𝑥!
endfor
endfor
untilStopcriterions
4) CompositeSplittingAlgorithm(CSA)[55,56]
Algorithm4.CSA
Input:𝜌 = 1 𝐿,𝑥!
repeat
fork=1toKdo
fori=1tomdo
𝑦!! = 𝑝𝑟𝑜𝑥! 𝑔! (𝐵! (𝑥!!! − !!∇𝑓! 𝑥!!! ))
endfor
74
𝑥! = !!
𝐵!!!!!!! 𝑦!!
endfor
untilStopcriterions
5) Fastcompositesplittingalgorithms(FCSA)[26]
Algorithm5.FCSA
Input:𝜌 = 1 𝐿,𝑟! = 𝑥!,𝑡! = 1
repeat
fork=1toKdo
fori=1tomdo
𝑦!! = 𝑝𝑟𝑜𝑥! 𝑔! (𝐵! (𝑟! − !!∇𝑓! 𝑟! ))
endfor
𝑥! = !!
𝐵!!!!!!! 𝑦!!
𝑡!!! = !! !!!(!!)!
!
𝑟!!! = 𝑥! + !!!!!!!!
(𝑥! − 𝑥!!!)
endfor
untilStopcriterions
6) FastcompositesplittingalgorithmsusedforCompressedSensing(FCSA-
MRI)[20]
Algorithm6.FCSA-MRI
Input:𝜌 = 1 𝐿 ,𝛼,𝛽,𝑟! = 𝑥!,𝑡! = 1
repeat
75
fork=1toKdo
𝑥! = 𝑟! − 𝜌∇𝑓 𝑟!
𝑥! = 𝑝𝑟𝑜𝑥!(2𝛼||𝑥||!")(𝑥!)
𝑥! = 𝑝𝑟𝑜𝑥!(2𝛽||Φ𝑥||!)(𝑥!)
𝑥! = (𝑥! + 𝑥!) 2
𝑡!!! = !! !!!(!!)!
!
𝑟!!! = 𝑥! + !!!!!!!!
(𝑥! − 𝑥!!!)
endfor
untilStopcriterions
7) FastcompositesplittingalgorithmswithLowRankTensorCompletion
(FCSA-LRTC)[58]
Algorithm7.FCSA-LRTC
Input:𝜌 = 1 𝐿! ,𝛼,𝛽,𝑟! = 𝑥!,𝑡! = 1
repeat
fork=1toKdo
fori=1tomdo
𝑦!! = 𝑝𝑟𝑜𝑥! 𝛼!||𝑥!||∗ (𝑟!! − 𝜌𝐴∗(𝐴(𝑟!! − 𝑏)))
endfor
𝑥! = !!
𝐵!!!!!! 𝑦!!
𝑡!!! = !! !!!(!!)!
!
𝑟!!! = 𝑥! + !!!!!!!!
(𝑥! − 𝑥!!!)
endfor
untilStopcriterions
76
8) SingularValueDecomposition(SVD)[43,84]
Algorithm8.BiorthogonalizationSVD
Input:𝑚,𝑛,𝐴where𝐴is𝑚 ×𝑛.
Output:Σ,U,𝑉sothatΣisdiagonal,Uand𝑉haveorthonormalcolumns,U
is𝑚 ×𝑛,𝑉is𝑛 ×𝑛,and𝐴=U Σ𝑉! .
1. U ← 𝐴.(Thisstepcanbeomittedif𝐴istobeoverwrittenwithU)
2. 𝑉 = 𝐼! × !
3. Set𝑁! = ( 𝑢!,!!!!!!
!!!! ),𝑠 = 0,and𝑓𝑖𝑟𝑠𝑡 = 𝒕𝒓𝒖𝒆
4. Repeatuntil𝑠! ! ≤ 𝜀!𝑁!and𝑓𝑖𝑟𝑠𝑡 = 𝒇𝒂𝒍𝒔𝒆.
a. Set𝑠 = 0and𝑓𝑖𝑟𝑠𝑡 = 𝒇𝒂𝒍𝒔𝒆.
b. for𝑖 = 1,… . .𝑛 − 1
i. for𝑗 = 𝑖 + 1,… . ., 𝑛
• 𝑠 ← 𝑠 + ( 𝑢!,!𝑢!,!!!!! )!
• Determine𝑑!,𝑑!, 𝑐 = cos (𝜃),and𝑠 = sin (𝜑)such
that:
𝑐 −𝑠𝑠 𝑐
𝑢!,!!!
!!!
𝑢!,!𝑢!.!
!
!!!
𝑢!,!𝑢!.!
!
!!!
𝑢!,!!!
!!!
𝑐 𝑠−𝑠 𝑐 = 𝑑! 0
0 𝑑!
• U ← 𝑈𝑅!,! 𝑐, 𝑠 where𝑅!,! 𝑐, 𝑠 istheGivens
rotationmatrixthatactsoncolumns𝑖and𝑗during
rightmultiplication.
• 𝑉 ← 𝑉𝑅!,! 𝑐, 𝑠
5. for𝑖 = 1,… . .𝑛
a. 𝜎! = 𝑢!,!!!!!!
b. 𝑈 ← 𝑈Σ!!
77
9) PrincipalComponentAnalysis(PCA)[39,83]
Algorithm9.PCA
Generateaninitialsolution𝑂𝑙𝑑_𝐶𝑜𝑛𝑓𝑖𝑔
for𝑛 = 0to⋕ofiterations
Generateastochasticperturbationofthesolution
𝑖𝑓Fitness(𝑁𝑒𝑤_𝐶𝑜𝑛𝑓𝑖𝑔)>Fitness(𝑂𝑙𝑑_𝐶𝑜𝑛𝑓𝑖𝑔)
𝑂𝑙𝑑_𝐶𝑜𝑛𝑓𝑖𝑔:=𝑁𝑒𝑤_𝐶𝑜𝑛𝑓𝑖𝑔
Exploration()
𝑒𝑙𝑠𝑒
Scattering()
𝑒𝑛𝑑 𝑖𝑓
𝑒𝑛𝑑for
Exploration()
for𝑛 = 0to⋕ofiterations
Generateasmallstochasticperturbationofthesolution
𝑖𝑓Fitness(𝑁𝑒𝑤_𝐶𝑜𝑛𝑓𝑖𝑔)>Fitness(𝑂𝑙𝑑_𝐶𝑜𝑛𝑓𝑖𝑔)
𝑂𝑙𝑑_𝐶𝑜𝑛𝑓𝑖𝑔:=𝑁𝑒𝑤_𝐶𝑜𝑛𝑓𝑖𝑔
𝑒𝑛𝑑 𝑖𝑓
𝑒𝑛𝑑for
return
Scattering()
𝑝!"#$$%&'() = 1− 𝐹𝑖𝑡𝑛𝑒𝑠𝑠(𝑁𝑒𝑤_𝐶𝑜𝑛𝑓𝑖𝑔)
𝐵𝑒𝑠𝑡 𝐹𝑖𝑡𝑛𝑒𝑠𝑠
𝑖𝑓𝑝!"#$$%&'() > 𝑟𝑎𝑛𝑑𝑜𝑚 0, 1
𝑂𝑙𝑑_𝐶𝑜𝑛𝑓𝑖𝑔: = 𝑟𝑎𝑛𝑑𝑜𝑚 𝑠𝑜𝑙𝑢𝑡𝑖𝑜𝑛