DirectContrastSynthesisforMagneticResonanceFingerprintingPatrickVirtue1,2,JonathanI.Tamir1,MariyaDoneva3,StellaX.Yu1,2,andMichaelLustig1

1ElectricalEngineeringandComputerSciences,UniversityofCalifornia,Berkeley;2InternationalComputerScienceInstitute,Berkeley;3PhilipsResearchEurope,Hamburg

Target Audience. Researchers and developers working to improve acquisitionandreconstructionforfastandquantitativeMRI.Introduction.Magneticresonancefingerprinting(MRF)cangeneratequantitativemapsoftissueandsystemparameters(PD,T1,T2,B0,B1)fromasingleacquisition[1].MRFalsohasthepotentialtoreplacestandardradiologicalsequencesbyusingtheparametermapstoindirectlysynthesizecontrastimages,suchasT1-andT2-weightedimages,Fig.1,dotted-bluelines.TheconceptofMRIsynthesisdatesbackto1985[2]andtechniquessuchasQRAPMASTER[3]haverecentlybeenshowntoproduce clinically viable images [4]. MR fingerprinting data contains richparametric information and can be used for contrast synthesis. However, MRIsynthesis techniques fromparameters are significantly limited by biases, due toeffectsthataredifficulttosimulate,suchastimevaryingsignals,partialvoluming,flow, diffusion, magnetization transfer, and others. We propose training neuralnetworks to directly synthesize contrast-weighted images from MRF data,bypassinginsufficientparametermodeling,Fig.1,solid-redline.Methods.Fingerprintingacquisitionandtrainingdata.Wescanned13volunteerswitha1.5TPhilips Ingenia scannerusing13receivechannels.Weacquired fourconsecutiveaxialheadsequences:T1-weightedspinecho,TE=15ms,TR=450ms;T2-weightedturbospinecho,TE=110,TR=2212;fingerprintingbalancedfastfieldecho (bFFE) sequence with 500 repetitions, constant TE=3.3 and TR=20, andsmoothly varying flip angles between 0-60 degrees. The spiral MRF acquisitionwasreconstructedtoimagespaceaftergriddingandcoilcombinationwithPhilipsCLEAR.Weusedthedatafrom11volunteersfortrainingandvalidationandusedthedatafromtwovolunteersonlyforthefinaltestresults.Indirect Contrast Synthesis. We simulated MRF signals with the extended phasegraph (EPG) algorithm [5][6], and used cosine similarity to match nearestneighborasin[1].ThenearestneighbormapswereconvertedtoT1-weightedandT2-weighted contrast images by simulating spin echo sequences: !" 1 −

%&'()'*'+ %&

'*', ,where!"istheprotondensity.

Direct Contrast Synthesis. The neural network for direct contrast synthesis wastrained on three million 3x3 patches from the in vivo MRF data using anarchitecturesimilartothetwo-channelreal/imaginarynetworkfrom[7],Fig.2.Results. Direct contrast synthesis consistently produced higher quality resultsthan the indirect contrast method, where T1-weighted and T2-weighted bothcontain significant artifacts, especially in thevasculature and cerebrospinal fluid(CSF),Fig.3.Discussion & Conclusion.We show that our deep learning model for directcontrastsynthesiscanbypassincompletesimulationmodelsandtheirassociatedartifacts. We look forward to expanding our experiments to include additionaltrainingdataaswellasadditionalcontrastimages,suchasFLAIR.

Figure 1. Contrast synthesis from MRF: indirect(blue-dottedlines)versusdirect(solid-redline).

Figure 2. Neural network architecture for directcontrastsynthesis.3x3spatialpatchesareflattenedand passed through three convolutional layers andthen three fully connected layers, resulting in acontrastvaluepredictionforthecenteroftheinputpatch.BetweeneachlayerisaReLUnon-linearfilter.The number of feature channels are shown aboveeach block, while the size of the temporaldimensionsareshownbelow.AnL2lossfunctionisusedtopenalizepredictedvaluesthatdonotmatchtheacquiredcontrastvalue.

Figure 3. Results from indirect contrast synthesis(a,d) and direct contrast synthesis (b,e). Note thatboth indirect synthesis methods presentinconsistent vessel contrast (white arrows), mostnoticeably in the superior sagittal sinus.

References.1.D.Ma,V.Gulani,N.Seiberlich,K.Liu,J.L.Sunshine,J.L.Duerk,M.A.Griswold,“MagneticResonanceFingerprinting,”Nature,vol.495,no.7440,pp.187–192,2013.2.S.A.Bobman,S.J.Riederer,J.N.Lee,S.A.Suddarth,H.Z.Wang,B.P.Drayer,J.R.MacFall,“CerebralMagneticResonanceImageSynthesis,”AJN,vol.6,no.2,pp.265–269,1985.3.J.B.M.Warntjes,O.D.Leinhard,J.West,P.Lundberg,“RapidMagneticResonanceQuantificationontheBrain:OptimizationforClinicalUsage,”MRM,vol.60,no.2,pp.320–329,2008.4.L.N.Tanenbaum,A.J.Tsiouris,A.N.Johnson,T.P.Naidich,M.C.DeLano,E.R.Melhem,P.Quarterman,S.X.Parameswaran,A.Shankaranarayanan,M.Goyen,A.S.Field,“SyntheticMRIforClinicalNeuroimaging:ResultsoftheMagneticResonanceImageCompilation(MAGiC)Prospective,Multicenter,MultireaderTrial,”AJN,2017.5.J.Hennig,“MultiechoImagingSequenceswithLowRefocusingFlipAngles,”JMR(1969),vol.78,no.3,pp.397–407,1988.6.M.Weigel,“ExtendedPhaseGraphs:Dephasing,RFPulses,andEchoes-PureandSimple,”JMRI,vol.41,no.2,pp.266–295,2015.7.P.Virtue,S.X.Yu,M.Lustig,“BetterthanReal:Complex-valuedNeuralNetworksforMRIFingerprinting,”ICIP,2017

FLAIR

MRFSignal

T1w T2w

ParameterMaps

SequenceSimulator

NeuralNetworkDirectContrast

Synthesis

EPGSimulator&NearestNeighbor

ISMRMAbstractFigure4: Networkarchitecture

816 32 512 256 1

TemporalConvolutions

Patchdata3x3x2=18

250125 62 1 1 1

FullyConnected

500

T2w

T1w

IndirectSynthesis DirectSynthesis TrueAcquisition

FLAIR

a b c

d e f

g h i