From: AAAI Technical Report SS-94-05. Compilation copyright © 1994, AAAI (www.aaai.org). All rights reserved. uniformlyscattered throughoutk-space. Thus,unlike other techniquesof constrained imaging, FRMR/ WiliTedprior knowledge to control the image acquisitionprocessas wenas the imagereconstruction procedure. Using a pattern recognition technique to constrain the acquisition and reconstruction of MRimages YneCaoand DavidN. Levin Departmentof Radiology,University of Chicago Chicago, IL 60637 In an earlier repon2, FRMRIwasapplied to simulatedand experimentalphantomimageswith small matrices(16 x 32). In the following,wedescribethe appficationof the methodto 64 x 64 imagesof the humanbrain. This required the use of a numerical optimizationtechnique(simulatedannealing)to find the best subset of phase-encoded signals to be measured. Section2 is a brief discussionof the theoreticalbasis of FRMRI,whichwasdescribedin detail in ref. 2. The application of FRMRIto humanbrain imagesis discussedin Section3. Theimplicationsof these results are describedin Section4. ABSTRACT Atraining set of MRimagesof normaland abnormalheadswasusedto derive a completeset of orthonormalbasis functionswhichconvergedto headlike imagesmorerapidly than Fourierbasis functions. Thenewimagerepresentationwasused to reconstruct MRimagesof other headsfroma relatively small numberof phase-encodedsignal measurements.The trainingimages alsodetermined exactlyw.h/.ch.t .t .t .t~encodedsignals should be measuredto nnmmme nnage reconstructionerror. Thesesignals werenon-uniformly scattered throughoutk.space. Experiments showedthat headimagesreconstructedwith the newmethodhad less truncationartifacts than conventionalFourierimages, reconstructedfromthe samenumber of signals. 1. 2. METHODOLOGY The FRMRIalgorithmcan be divided into twophases. Thefirst phaseconsists of preprocessing or training computationswhichare performedjust once for a givenclass of images(e.g. the class of Tlweightedaxial imagesthroughthe lateral ventricles of the brain). Thesecalculations determine:(1) a formula for reconstructingthe training imagesfroma reducedset of PEsignals; (2) the indices of the phase-encodings whichproducethe mostaccurate suchreconstruction. Oncethe training phaseis completed for a particular class ofimages, "unknown" objects (objects notinthe training set) can be imagedby followingthe steps. First, the scanneracquiresa reducedset of PEsignals at the locationsin k-space specified bythe training computations.Usingformulasderivedin the training phase,these measurements are linearly transformedto producean imageof the unknown object. INTRODUCTION Conventionalmagneticresonanceimages (MPA) are reconstructedby Fourier uansformation of limited numberof phase-eneeded if’E) and frequeneyencoded(FE)signals in the center ofk space. There alwaysa tradeoff betweenimagingtime andS’l~rhl resolution. If onewishesto reducedamacquisition time, fewerPEsignals can be measured,andthe resulting imageswill havemoretruncationartifacts and worsespatial resolution. Manyconstrainedimage reconstructionalgorithmsattemptedto improvethis tradeoff by usingprior knowledge. Mostof these I did not attempt to optimizethe image methods acquisition;the measured PEsignals werestill uniformlyspacedin the center of k-space. Unmeasm’ed high-ksignals wereestimatedfromthe measuredlow-k data by using expficit mathematicalmodels,which reflecteda particularauthor’sintuition aboutwhatthe imagesshouldlook like. 2.1 Training images Thealgorithmutilizes prior knowledge in the formof a set of "trainingimages",comprised of a large collection of high resolutionimagesof objects resemblingthe"unknown" objects tobescanned. For example,a large set of normalandabnormalbrain scans wouldbe usedto train the algorithmto acquireand reconstructbrain images.Thetraining imagesare acquiredby conventionalMRscanning,i.e., by measuringPEsignals, whichsamplerows in the kspacerepresentationof the image.Thesampledpoints forma Cartesiangrid in the centerof k-spaceandhas uniformdistribution in eachdirection. Thespacing betweenpoints is determinedby the fields of view (FOV)Df and Dpin the FEand the PEdirections, respectively. Thenumberof sampledpoints in the FE direction(Nf)arid in the PEdirection(Np)is determined by the desiredresolutionalongeachaxis. Let gn(a) representthe complex data at the point with FEindexa andPEindex n. Thevalues ofgn(a) for each fixed valueof a (i.e. eachcolumn in k-space)canbe regarded Werecently developeda newapproachto 2, called Feature-Recognizing constrainedMRimaging MRIfb’R MRD. In this scheme,a set of training imagesreflected our prior knowledge about the objects to be imaged.Anautomaticpattern recognition techniquewasusedto examinethe space of all possible imagesandto t’mdthat subspaee,whichbest described the training set. Since the imagesof subsequently scannedobjects werepresumed to lie in this subspace, they couldbe reconstructedfroma subset of the usual array of PEsignals. Furthermore,the training images automaticallydeterminedwhichsignals should be measured in orderto reduceimagereconstructionerror. In mostcases, the optimal measurements werenon215 as a complexcolumnvector in an Np-dimensional imagespace. Since the Ixaining imagesare similar to oneanother,the vectorscorresponding to different training imagesare usuallyclusteredin the imagespace. 2.3 Let<gn(a)> betherawdA~array ofthe average training image, anddefine fn(a) tobethe diffe~nce benueen therawdata ofa specific image and this average training image: fn(a) = gn(a) - <&nfa)> TheFoudcx Iransformation offn(a) isa complex function l(x,y), which represents thedifference between the averagetraining imageandthe imagecorresponding to gnfa ): ~’z+tq’u l(x,y)ffi . t ~’.~".fs(a)e (I) Dt.~-~p .=~...s;.= .17"..up Theaim of the FRMRIalgorithmis to recover the functionl(x,y) froma subset of the N/) PEsignals. relxesentedby gnfa) orfnfa). Thedesiredimage isthen produced by adding theaverage waining image tol(x,y). 2.2 FR basis functions Theimagespaceis spanned bymany complete sets oforthononnal unit vectorswhichate "rotated" withrespect totheaxes corresponding totheFomier imagecomponentsfn(a). Theimage’s components (vn(a)) suc h an alt ernative rdinate coo tem sysate related to its Fourier components by an Npx Np unitary Iransformation,Unn’: f.(a)= ~’~Um,(a)vs(a) (2) a’=I...N# Theadoptionof such a "rotated" coordinatesystemis equivalentto representing the imagewith a completeset of orthonormalfunctions Van(X,y): l(x,Y) = ..~..s¢! a=~...sc v.(a)Vas(x.Y) O) whe~ V,n(x,y) Image reconstruction If wewantto reconstruct animagefroma subset of PEsignals with indices [L) (L < Np),it is necessaryto truncate the FRexpansionofeach column of k-spacedata sOthat it has L or fewerterms.In othe~ words,Eq(2) andEqO) mustbe truncatedafter the firs M(a)terms whereM(a)< L. Then, the L measured components Offn(a)Canbeused tO CalC))latOthe values ofthefirst M(a)FRcomponents, vn(a in ).Asshown reference 2,these ~ components canbeusedto estimate the Np-Lunmeasured PEsignals: fh(a)= [Q(aXR(a)+R(a))’lR(a)+]hlfk(a) k e{C} whereRkm(a)is an L x M(a)matrix with elements Rkm(a)= Ukm(a)for m = l...M(a) and k e {L) whereQign(a)is the (Np-L)x M(a)matrix: Qhm(a) Uhra(a)form= l..a~lfa) andh ~.tL). Finally, the imagecanbereconslructedby conventionalFourier transformation(Eq (1)) of the L measuredPEsignals andthe Np-LestimatedPEsignals in Eq(6). 2.4 Optimal PE signals to be measured In principle, the imagecan be reconstructed from any subset of L PEsignals. However,some choices ofthese measured signals mayleadto amplification of envrs dueto measurement noise and series truncation. Theoptimal PEsignals to be measured are those which lead to least error in the FR reconsgucfion of the training images.If e is the root meansquarednoise in the PEsignal measurements, th, averageerror in the FRreconstructionof the training imagesis: < ~r> =e2,,~...t¢ NiLe2 + Tr[(Q(a)÷Q(a)XR 41 (a)*R(a)) t I. e#, ~ Y Un’ n(a)dq’~’ (4) + ~.~...s, Tr[W(a)T(a)W(a)+] Sinceall of these alternativebasissets arc complete,eachonecan be usedto representan arbitrary imageexactly as well as the Fourier expansionin Eq (1). However, if these representations arcmmcaled at anypoint, there is onespecial basis set whichprovides the mostaccuraterepresentationof the training images. Asdemonstrated in reference2, the optimalbasis set can be derivedfromTnn’(a),the covariancematrixof the training images: (5) TM.(a) =l ~. f~(ad)fs(aJS* CO whe~ W~(a) = -[Q(a)(R( a)÷R( a)) 4 w~.(a) =&h. for h 4 (Ll, t e tLl, h’ ~(L} (8) Ourtaskis to find the indices {L) of the PEsignal measurements whichminimize<E~,>.Notice that the matrices Qand R in Eq (7) dependon the mmcation points M(a)andthe indices (L). order to avoid to examinethe hugesearch spacedefinedby the simultaneousvariation of the M(a)andthe indices (L) wehaveused the followingitemtive procedureto look for optimalvaluesof M(a)and(L}. First, weesfimat, the values of M(a)whichled to an acceptableamount averagemmcation error in the representationof the wainingimagesby M(a)FRbasis functions. Next. simulatedannealingwasusedto identify the optimal indices[L), i.e. the optimalk-spacelocationsfor the measuredPEsignals used to calculate those M(a) J j.,~..a wherefn(aj)is the Fourier representationof the/h training image.Theunitary transformationto the optimalcoordinatesystem(i.e. Unn’(a)in (2)) is matrix whichtransformsTnnfa) into diagonalform with descendingeigenvalues.In other words,the optimalimagebasis functionsare the principal components of the columnvectors of the training 2. images 216 After the optimalset of measuredPEsignals wasfoundin this way,weexamined the eigenvaluesof the matrix R(a)+R(a),whichmust be inverted to performthe FRreconstruction.If a smalleigenvalue wasfound, M(a)waslowereduntil all eigenvalueswere acceptablylarge, therebyavoidingerror amplification duringimage reconstruction. components. Finally M(a ) wasadjusted,as needed, minimize ill-conditioningof the calculationof the M(a) FRcomponentsfrom the optimally.located L PE measurements. Asstated above,the M(a)wereinitially chosen to lead to an acceptablelevel of truncationerror. As shown in reference2, the meansquarederror in the truncatedFRrepresentationof the training imagesis givenby the sumof the last (Np.M)eigenvaluesof Tnn’(a). Themeansquaredmagnitudeof the training imagesis givenby the sumof all of these eigenvalues. Therefore,the relative size of the meansquared truncationerror is givenby the ratio of these two quantities: 2.5 Imaging unknownobjects After thetraining procedureis completed, "unknown" imagescan be quicklyreconstructed. First, the operator choosesthe numberL of PEsignals to be acquired;for example,this mightbe determinedby the time available for scarming.Next, the scannermeasures the optimalPEsignals with indices (L} determined duringthe training phase.Imagereconstructionis performedby "back-projecting"these measuredFourier components onto the M(a)-ffnnensionalreconstruction subspace,spannedby the In’st M(a)FRbasis functions. <Era(a)2>-i,.M(a~l..A’p gi(a)/,-.~..N, k(a) where),i(a) is the I th eigenvalueof Tnn’(a).TheM(a) wereinitially chosento lead to an acceptablevalueof this relativetruncationerror. Givenan initial choicefor M(a),the nextstep wasto find the indices [LJ of measuredPEsignals whichminimized the error in the FRreconstructionof the training images;namely,Eq(7). This involved lookingat manypossible choicesof {L); for example, for Np- 64andL ffi 32there are 64!/32!(64-32)I 1018choicesfor these indices. Weusedsimulated 3 since it is an efficient meansof finding annealing global or nearly global minimain such hugesearch spaces.In our case, the state of the systemto be annealedwasdefinedby the indices(L.h andthe system’senergywasgivenby the teconslructionerror in Eq(7). After the initial state of the systemwaschosen randomly,a proposedperturbation("move")was stochasticallygenerated;i.e. someindices in (L) were deleted, and an equal numberof indices were~ to the set. TheenergychangeASassociated with the proposedmovewascalculated. If dE < 0, the move wasacceptedanddefinedthe next state of the system; i.e. the systemalwayswent"downlfill". If AE> 0, the con’esponding Boltzmann factor e"AE/T wascalculated, where T wastheuser-controlled "tempemtme" ofthe system. Themovewasaccepted withhigh(low) probability iftheBoltzmann factor wasclose toI (0); otherwise, themovewasrejected. Inother words, the systemwaslikely to makelarge "uphill" movesat high T, but usually madeonly small uphill movesat low T. After a large series of moveswasmade,the temperature parameterT waslowered,anotherseries of moveswas generated,T wasloweredagain, etc. If the initial temperatureis high enough,enoughmovesare madeto approachthermodynamic equilibriumat each temperature,andT is loweredslowlyenough,this procedureis highly likely to movethe systemto a 3. In other words,the global or nearly global minimum systemwas"frozen"or annealedin the "ground"state (or nearlygroundstate). 3. 3.1 APPLICATION TO IMAGES OF THE HUMAN HEAD Training images Weapplied the FRalgorithmto 64 x 64 imagesof the humanhead. First, Tl-weightedaxial imagesof 75 normaladult volunteersand patients withoutlarge lesions werecollectedby a 1.5 T scanner. A spin echopulse sequencewasused with parameters: TR/rE= 500/16 msec, slice thickness = 6 mm,FOV= 24 cm,and2 excitations. If necessary,the imageswere manuallyIranslated a fewpixels in the FEand PE directionsso that the headwasretrospectivelycentered in the FOVno matter howthe patient waspositionedin the scanner.Next, the k-spacedata werenormalized to unit magnitudeat k=0in order to compensate for variations in receiver gain fromscanto scan. Finally, the averagetraining imagewassubtractedfromeach image. 3.2 FR basis functions Thetraining imageswereusedto calculate and diagonalizethe covariancematrixTnn’(a)(E,q (5)). This proceduredeterminedthe FRbasis functionsVan(X,y) given by Eq (4). Thesecompriseda completeset orthonormal functionswhichwere"tailored" to convergerapidly to head-likeimages.Thefirst FR functionsdepict features whichare common in the training images,suchas the scalp, diploicfat, brain surface, andventricles. Higherorderbasis functions describefeaturesrarely foundin the trainingset; for example,they are distributedin regionswhichare filled with air in mostimages. 3.3 Optimal PE signals to be measured We"Irained"the algorithmto reconstruct imagesfrom8, 16, 24, 32, or 40 PEsignals. For each 217 slightly less aocmatethan the FRimagesreconstructed from the samenumberof signals with TR= 500 ms. This was becauseimagesacquired with TR= 300 ms andI000msare expectedto be "farther" fromthe cluster of training images(TR= 500 ms). Nevertheless,the FRreconstructions for TR= 300 ms and1000msstill havesignificantly less truncation artifact than the comparable FTreconstructions.This suggeststhat the FRalgorithmderivedfroma single Tl-weighted training set can be usedto reconstruct imagesacquiredwith a widevariety of Tl-weighted pulse sequences.TheFRalgorithmwasalso used to reconmuct a training subject, after the data were modifiedto containa simulated"lesion" unlike any lesion in the training set. It is apparentthat the FR algorithmreconstructedthe lesion accuratelywhenit waslocatedinside the head. This is becausethe IruncatedFRseries is fairly completeover the head regionwheremosttraining imageshavea variety of structural detail.~. of these choicesof L, the first step wasto estimate M(a),the numberof FRbasis functions to be usedin the reconstructionof the ath columnin k-space. As describedin Section2, this wasdoneby requiringthat the averagemnr, afion ezror <Erel(a)2> be less than a certain upperlimit: e.g. 0.01 for L=32.This criterion usually led to valuesof M(a)whichwerelowerfor central columnsin k-spacethan for peripheralcolumns. For example,for L = 32, M(a)was10, 27, and 30 for a = 32, 16 and 1, respectively. Next, simulatedannealing wasusedto determineoptimalor nearly-optimalindices (L} of the PEsignals to be measured.Theenergyto be minimizedwasgiven by Eq (7). Weused a value of (the noise in each PEsignal), whichwouldproduce signal-to-noiseratio of 7 in a conventionalFTimage. These calculations took 1-15 CPUhours on a Convex C-3supercompuer. This training procedureis practical since it only has to be doneoncefor eachclass of images. 3.4 Image reconstruction 4. Theaboveproceduredetermined the indices(/.,) of PEsignals whichled to a reasonablysmall average error in the FRreconstructionof the training images. TheFRreconsu’uctionformula,Eq (6), uses those signals to reconstructthe ath columnof a k-spaceimage as a sumof M(a)FRbasis functions. Notice that this imagereconstructionformulainvolvesthe inverse of a matrix R(a)+R(a).For a few k-space columnsthis matrix hadsmall eigenvalues(e.g. conditionnumber greater than200)whichcausedits inverseto be illconditioned.Therefore,wedecreasedM(a)for those columns(i.e. usedfewerFRbasis functions)so that the matrixinversewasbetter definedandthere wasless error ampfificationin the reconstructionprocess. DISCUSSION ConventionalMRIhas spatially-uniform resolutionbecauseit utilizes Fourierbasis functions (sines andcosines)whichoscillate uniformlyacrossth~ FOV.Furthermore,the measuredPEsignals are uniformlydistributed in the center of k-space.This strategy for imageacquisitionandreconstructionis appropriate in situations in whichthere is no prior informationabout the object being scanned. However, the method suffers fromthe fact that sharpedgesare poorly representedby Fourier functions. Suchedges generatetruncationartifacts whichcan obscureadjacent structures. In contrast, FRMRIutilizes prior informationin the formof training imagesto generate "customized"set of basis functionswhichconvergeto similar imagesmorerapidly than the generic Fourier functions. Thesebasis functions maybe highly localizeddistributions andcan containsharpedges. Th FRimagereconstructionprocessis characterizedby Spatial resolution whichvaries across the FOV, Regionscontainingmuchstructm’al detail in the training imagesare morehighly resolved and"come into focus" earlier in the FRseries expansion.TheFt expansionof an imageis best calculatedfroma set of PEsignals whichsamplek-space in a non-uniform fashion, also dictated by the training images.Thus,th FRstrategy is to use prior informationto optimizetin scanningprocessas well as the imagereconstruction Oncethis training procedurewascompleted, "unknown" objects were imagedby measuringthe optimalL PEsignals and substituting themin Eq(6). The FRimagesof a typical "unknown" headwere reconstructedfrom8, 16, 24, 32 and40 PEsignals. Themeasuredsignals wereacquiredwith the samepulse sequenceusedto collect the wainingdata. Theresults are comparedto conventionalFTreconstructionsfrom the samenumberof PEsignals, distributed uniformly in the center of k-space. TheFRimageshaveless truncation artifact than the FTimages. Common featuresin the training images(the scalp, diploic marrow,brain surface, ventricles, andinterventricular septum)"comeinto focus" even for highly mmcatedFR reconstructions(e.g. L = 8, 16, 24). This happens becausethese featuresare "built into" the first FRbasis functionsandare addedinto the FRseries, complete with highlyresolvededges, as soonas they are "detected"in the measured signals. 1. Z.P. Liang, F.E. Boada,R.T. Constable, E.M. Haacke,P.C. Lauterbur, and M.R.Smith. Rev. Mag.Res. in Med., VoL4, pp. 67-185, 1992. 2. Y. Cao and D~. Levin. Mag. Res. Med., Vol. 30, pp. 305-317,1993. 3. S. Kirkpatrick, C.D.Gelatt, and M.P. Vecchi. Science, Vol. 220, pp. 671-680,1983. TheFRalgorithmwasused to reconstruct imagesof unknown headsscannedwith different TRs: such as TR=300and 1000ms. The FRreconstructions from PEsignals with TR= 300 ms and 1000mswere 218