1、核医学影像中的数据处理核医学影像中的数据处理中国科学院高能物理研究所北京市射线成像技术与装备工程技术研究中心贠明凯Modern Nuclear Medical ImagingScanners Computers UsersOutline lData organizationlCorrection methodslRebinning lImage reconstructionlImage registration and fusionlDICOM and PACSOutline lData organizationlCorrection methodslRebinning lImage reco
2、nstructionlImage registration and fusionlDICOM and PACSData organizationlList modelHistgramlSinogramlLinogram5010015020025020406080100120SinogramPET0Sinogram rlProjections and SinogramSinogramPET0SinogramrlProjections and SinogramSinogramSPECT2D VS. 3DlSepta between crystal ringslLower sensitivitylL
3、ower randomlLower scatterl2D reconstructionlNo septalHigher sensitivitylHigher randomlHigher scatterl3D reconstruction or hybrid reconstructionOutline lData organizationlCorrection methodslRebinning lImage reconstructionlImage registration and fusionlDICOM and PACSScatter CoincidenceTrues Coincidenc
4、eRandom CoincidenceTrue Counts & NoiseNormalizationABCDAttenuationABCDScatterABCDNeed to correct the dataCorrection methodslrandoml“dead time”lnormalizationlscatterlattenuationldecaylArc correctionlDepth of interactionlMotion correctionlPartial volumelAxial of rotationlCamera head tiltRandomFinite t
5、ime window withEnergy window Coincidence timing windowActivity Random lTail fittinglsimplestlSmall changes in tail, great changes in estimatelEstimation from singles rateslMeasure the single count rate on each detector for a given time windowlSubtracting from the prompts between detector pairlSingle
6、s rate is much larger than that of coincidence eventslSingle rates change in the same way over timeDelayed coincidence channel estimationlOne channel is delayed before being sent to coincidence processinglSubtracted form prompt coincidenceslAdvantage lAccuratelSame dead time environment as prompt ch
7、annellDisadvantage lIncreased system dead timelDoubling of the statistical noise due to randomDead time correctionlDecaying source experiment is performedDead time correction (con)lLook up tablelUniform sourcelKnown quantitylShort livedlLinear extrapolation from count rate for a given level of activ
8、ityNormalization lCauses of sensitivity variationslSumming of adjacent data elementslDetector efficiency variationslGeometric and solid angle effectslRotational samplinglTime window alignmentlStructural alignmentlseptaSumming of adjacent data elementsGeometric and solid angle effectsRotational sampl
9、inglLOR at the edge are sampled less than LOR close to the centerCrystal interface factorsTime window alignment factorsNormalization methods (con)lDirect normalizationlSimplest approachlAdequate statistical qualitylVery uniform activity sourceslScatter in normalization should be substantially differ
10、ent from normal imagingNormalization methods (con)lComponent-based normalizationuivjaxvaxutrvjtruivjuiscatteruivjtbbbbuivjaxuvtruivjaxvaxutrvjtruivjuitrueuivjmggbbbbScatter correctionlLORs recorded outside object boundary can only be explained by scatterlThe scatter distribution is very broadlScatte
11、red coincidences fall within the photo-peak window mainly due to scattered once Scatter correctionlEnergy spectra distribution of scattered 511KeV photons according to the number of times each photon scattersScatter correctionlEmpirical scatter correctionslFitting the scatter tailslDirect measuremen
12、t techniquelEnergy window techniqueslDual energy window methodslMultiple energy window methodslConvolution and de-convolutionlSimulation-based scatter correctionlAnalytical simulationlMonte Carlo simulationFitting the scatter tailslSimplest approachlFit an analytical function to scatter tailslSecond
13、 order polynomial or 1D GaussianlCoincidences outside the object are entirely scatter eventslNot always well approximated, particularly in thoraxDirect measurement techniquelOnly applicable to PET with retractable septalStepslMake a measurement of the same object with and without septalScaling septa
14、 extended projections for different efficiencylSubtract from projections of polar angle 0lEstimate the oblique scatter by interpolation of the direct plane scatterDual energy window methodspwsclwscscCCRpwunsclwunscunscCCRunscsclwscpwpwunscRRCRCCDual energy window methodsMultiple energy window method
15、sScatter CorrectionAnalytical simulationScatter CorrectionSingle Scatter - Model based correctionCalculate the contribution for an arbitrary scatter point using the Klein-Nishina equationBeforeScattercorrectionAfterScattercorrectionAttenuation correctionlAttenuation in the body is equal to that of s
16、ource lying along the same LORZaidi H, Hasegawa B. J Nucl Med 2003; 44:291-315.SPECTPETAttenuation correction (con)lMeasured attenuation correctionlCoincidence transmission datalLong-lived positron emitterlNormally more than one rod source are usedlSinogram windowing is applied provide location of r
17、odlImpractical in 3DlSingles transmission datalShielded point transmission sourcelSeparate blank scan is neededlSignificant scatter and broad beamMeasured attenuation correctionlCoincidence measurement using rod sourcelTransmission measurement using point sourceCT scanlAdvantage lHigh statistical qu
18、alitylHigh spatial resolutionlSignificant reduction in scan timelDisadvantagelFaster CT, slower PETlSmaller FOV of CTlDifficulty in registrationl values do not scale linearlyAttenuation correction for PETTypes of transmission imagesCoincident photon Ge-68/Ga-68(511 keV)high noise15-30 min scan timel
19、ow biaslow contrastSingle photon Cs-137(662 keV)lower noise5-10 min scan timesome biaslower contrastX-ray(30-140kVp)no noise1 min scan timepotential for biashigh contrastOther attenuation correction methodslCalculated attenuationlRegular geometric outlinelConstant tissuelSegmented attenuationlSegmen
20、t transmission image according to tissue typelAssigning known attenuation coefficientslForward projectionattenuation correction05010015020025030000.20.40.60.811.21.4att05010015020025030000.10.20.30.40.50.60.70.8050100150200250300050100150200250300350Attenuation/Scatter correctionUniversity of Pennsy
21、lvania PET CenterNo AC or Scatter CorrAC and Scatter CorrPhilips AllegroArc correctionlDifferent sampling distance at different radial positionlEqual sampling distance is required in analytical methodlInterpolation methodlNearest interpolationlLinear interpolationlB-spline interpolation (negative va
22、lues!)DOIdepth of interactionDOIdepth of interaction（con） Dual LayerA Point Spread Function (PSF) describes the response of an imaging system to a point source or point object. A system that knows the response of a point source from everywhere in its field of view can use this information to recover
23、 the original shape and form of imaged objects. PSFs are used in precision imaging instruments, such as microscopy, ophthalmology, and astronomy (e.g. the Hubble telescope) to make geometric corrections to the final image.Point Spread Function (PSF) Motion correctionlCardiac motion and respirationMo
24、tion correction（con）lGated frameslList modelRespiratory motion is distributed throughout the whole bodylImpact is rarely on detection, but often affects quantitationStatic wholebodySingle respiratory phase(1 of 7, so noisier) 1 cc lesion on CTWhole-body respiratory gated PET/CT: PatientsPartial volu
25、me effectlCharacterslObject or structure being imaged only partially occupies the sensitive volume of scannerlSignal amplitude becomes diluted with signals from surrounding structureslThe degree of underestimation of radioactivity concentration will depend not only on its size but also on the relati
26、ve concentration in surrounding structureslCorrection methodslResolution recoverylUse of anatomical imaging dataA Point Spread Function (PSF) describes the response of an imaging system to a point source or point object. A system that knows the response of a point source from everywhere in its field
27、 of view can use this information to recover the original shape and form of imaged objects. PSFs are used in precision imaging instruments, such as microscopy, ophthalmology, and astronomy (e.g. the Hubble telescope) to make geometric corrections to the final image.Point Spread Function (PSF) Partia
28、l volume effectMAPassumptions:camera moves along circular orbitorbit is reproducible calibration method finds system geometryproblem 1: tilting detectorassumption: camera moves along circular orbitAORAxial of rotationlOffset of AORlRotation of AORlNutation of AORCamera head tiltlHeads need to be exa
29、ctly parallel to axis of rotationCorrect alignmentHead tiltpinhole calibrationDirk Bequ, Kathleen Vunckxcircular orbitcircular orbit + new modelextension 2: circular orbit + arbitrary small deviationsmeasurementmodelMichel Defrise, Chris Vanhoveextension 2: circular orbit + arbitrary small deviation
30、soldnewtranslationsrotations1mm-3mm1.5mm-1.5mm1.5mm-1mm1o-2o1.5o-1.5o3.5o-2.5o1mm1.21.41.61.82mmOutline lData organizationlCorrection methodslRebinning lImage reconstructionlImage registration and fusionlDICOM and PACSRebinninglConvert 3D data to 2DSSRB and MSRBlSSRB- Single-slice rebinninglDetectio
31、n: center slice lSimple lFast lResolution losslMSRB- Multi-slice rebinninglDistribute along all intermediate sliceslDe-blurring along z-axisFourier rebinningOutline lData organizationlCorrection methodslRebinning lImage reconstructionlImage registration and fusionlDICOM and PACSImage reconstructionl
32、AnalyticallFBPlBPFlFDKl3D RPllIterativelARTlMLEMlOSEMlOSLSlMAPlAnalytical algorithmslFor example, FBP (Filtered Back-projection)lTreat the unknown image as continuouslPoint-by-point reconstructionlRegular grid points are commonly chosenlTreat projection process as line integral theoretically解析重建解析重建
33、-FBPFBPback projection (BP) = summation of projectionsfiltered back projection (FBP)FDKlFeldkamp、Davis、KressFDKstzODstOSP(t,s,z)rQQQDFocus losusxy) (ttstzddOO DD)(11SS)(22SS2022222)(),()(21),(dpdpDDpphpRsDDzstfcossin)sincos(yxDyxDsDDtpcossinyxDDzsDDz3D RPRe-projectionSteps of 3D RPlExtract 2D sinogr
34、amslReconstruct each with 2D FBP and stack to form 3D imagelForward project to calculate missing LORslExtract 2D projection data of all oblique sliceslTake 2D Fourier transformlBack project data through 3D image matrixlRepeat for all angles and oblique slicesWhat is iterative reconstructionlDiscrete
35、 measurements, discrete imagelOptimizationAttractions of iterative methodslEither consistent or inconsistent is OKlComplex geometrylPhysical effects and detection processes can be modeledlNon-negativity lGreat reducing streaking artifactslBetter contrast recoverylClassification of iteration reconstr
36、uction methodslART (algebraic reconstruction techniques)lMART (multiplicative ART)lAART (additive ART)lSIRT (simultaneous iterative reconstruction)lSMART (simultaneously MART)lBI-ART (block iterative ART)lBI-SMART (block iterative SMART)lRBI-SMART (rescaled BI-SMART)Statistical algorithmslMAP: lMaxi
37、mize the conditional probability P(image|data)lMLEM:lMaximize the probability P(data|image)Statistical algorithmsGaussian assumptionP is projection column matrix, A is system matrix, F image column matrix, C is the covariance matrix of the dataAssumed all standard deviations are identical and equal
38、to 1, idealized parallel projection, perfect resolution and no attenuation or other degrading affectsStatistical algorithmsPoisson assumptionjikjkkifjipjdjipjipifjif)1( )1()() () , ()(),(),()(),(jjsjiksjkkifjipjdjipjipifjif)1( )1()() () , ()(),(),()(),(实测数据迭代重建迭代重建-MLEM&OSEM正投影比较更新重建MLEM, OSEM, .lik
39、elihooditerationSinogramr1 3 2 4Subset order012341040ordered subsets1 iteration of 40 subsets(2 proj per subset)System matrixlScan geometrylCollimator/detector responselAttenuationlScatter (object, collimator, scintillator)lDuty cycle (dwell time at each angle)lDetector efficiencylDead-time losseslP
40、ositron rangelNon-colinearitylCrystal penetrationConsiderations of system matrixlQuantitative accuracylSpatial accuracylComputation time lStorage spacelModel uncertaintieslArtifacts due to over simpleificationslSystem matrix trickslFactorizelSymmetrylSparsenesslApproximationlPartial Monte CarlolSyst
41、em matrix modelReconstruction image of uniform sourceFBP VS. OSEMlFBPanalyticallPros：lSingle pass lLinearlFastlCons：lStreak artifactlPoor resolutionlCorrection not built-inlOSEMiterationlPros：lBetter resolutionlBetter contrastlLower noiselCons：lExtensive time consuminglMemory consuminglRequired user
42、 trainingFBP VS. OSEMlPhantom test (left)lClinical results (right)Outline lData organizationlCorrection methodslRebinning lImage reconstructionlImage registration and fusionlDICOM and PACSImage RegistrationPETCTPET/CTVoxel based image registrationImage RegistrationImage Registrationl算法流程图相似性测量一般用到的函
43、数有相似性测量一般用到的函数有：相同模态图像：残差（sum of square difference）不同模态图像：互信息（mutual information）一般用来做配准的优化算法有：一般用来做配准的优化算法有：六参数或十二参数的优化一般使用 Powell 优化算法多参数优化一般使用LBFGS（ limited-memory BroydenFletcherGoldfarbShanno ）优化算法（由牛顿算法演变而来）Image FusionlAlpha Blending basedlAdaptive alpha blending Alpha blendingAdaptive Alpha
44、blending( , )( , )*( , )( , )*(1-( , )Ri jRi jalpha i jRi jalpha i jFAB( , )( , )*( , )( , )*(1-( , )Gi jGi jalpha i jGi jalpha i jFAB( , )( , )*( , )( , )*(1-( , )Bi jBi jalpha i jBi jalpha i jFABOutline lData organizationlCorrection methodslRebinning lImage reconstructionlImage registration and fu
45、sionlDICOM and PACSDICOM and PACSlDICOMlDigital image and Communication in MedicinelCreated by ACR (American College of Radiology) and NEMA (National Electrical Manufacturers Association) in 1985lTwo components:lCommunication protocols and image format standardsDICOM image VS. General imagelStructur
46、e:lDICOM contains header and image data sections;lOther image file such as BMP, JPG, TIFF, which contain also two sections;lSize:lThe header size of DICOM is variable;lThe header size of many general image is constantlContents:lDICOM contains additional patients data such as basic information, study
47、 information and so on;lGeneral image header describes basic image parameters, such as image size, compression typeDICOM logical layerslPatient information:lPatients name, patients ID, patients birth date, hospital information systemlStudy information:lStudy information such as dose, injection time
48、and additional examination information such as study name study date;lSeries information:lSeries ID, manufacturer and institution namelImage information:lThe size of pixel, image size, pixel value and how it is encodedPACSlPACSlThe Picture Archival and Communication SystemlA system for storage of im
49、ages and transferring images between computers in different facilities through networkslPACS is helpful to provide comparative studies among different image modalities PACS (con)Thanks Our Group is Our Group is growing up growing up TOF-PETTOF信息信息PET原始数据偶然符合校正发射图像衰变校正死时间校正弓形几何校正衰减校正散射校正归一化校正图像重建数据重组数据校正
