IMAGE RECONSTRUCTION FROM TRUNCATED PROJECTIONS : A LINEAR PREDICTION APPROACH N. Srinivasa* * V. Krishnan' K.R. Ramakrishnan Dept. of Elect. Engg., Indian Institute x K. Rajgopal * of Science, Bangalore, INDIA. "School of Automation, Indian Institute of Science, Bangalore, INDIA. 'Dept. of Elec. Engg., University of Lowell,Lowell, Abstract Unambiguous reconstruction is not possible with truncated projections. In order to r e d u c e t haem b i g u i titynhree c o n s t r u c t eidm a g es ,e v e r a l a simple 'completion' a u t h o rhsa vseu g g e s t etdh a t involving extrapolation of t threu n c a t pe dr o j e c t ion sufficient. is In this paper method a based Prediction approach proposed is for on a Linear o b t a i n i n gt h em i s s i n gp a r ti nt h ep r o j e c t i o n .R e c o n s t ruction then is carried out using the Convolution 111. Simulationresults Backprojection(CBP)method s h o w i n g t h e e f f e c t i v e n e s s of this method in extracting significantly more information from truncated projections are presented. Introduction Image reconstruction from projection relies onthe fact t h a t h eo b j e c ft u n c t i o ni sr e p r e s e n t e d c o m p l e t e l y by a set of lineintegrals so t h a t a knowledge of the projections is sufficient to unambiguously d e t e r m i ntehset r u c t u r e of t hoeb j e c t . In medical imaging, projections physically correspond the to l i n ea at tre n u a t i o n of X-rays, time-of-flight the of aunl t r a s o n ipcu l soetrhteo t arla d iaoc t i v i t y isotope. in aenl e m e n t av lo l u m e of a distributed O t h ea rp p l i c a t i oanr e awsh e ri m e a g ea sr e c o n s t r u c tfer dop m r o j e c t i oaN nr sue c l eMa ra g n e t i c Resonance, electron microscopy, radio astronomy, etc.,. In practice,therearemanysituationswherein is t r u n c a t e (da l s roe f e r r e d the available projection to as 'limited field of v i e wa' n d' r e s t r i c t e dr e g i o n F oirn s t a n c et ,hper o j e c t i o nasrnee c e scan') [ 2 ] . s s a r i l tyr u n c a t e d if the object being examined has d i m e n s i o nl as r g et hr at nhf ei e l d of view of t h e as in t h e case of G a m m ac a m e r a s imagingsystem, nuclear in medicine. In 'clinical situations where t h ed i s e a s e da r e ai sa l r e a d yl o c a l i z e df r o m a previous a followupscan of t h eR e g i o n Of Interest studyand a r e s t r i c t e rde g i o snc a inds o n e (ROI) is required, to minimize the radiation dosage. Here order in as t h e projection data intentionally is truncated to radiation. In external region not issubjected electronmicroscopy,it is knownthatinnegatively of certain biological objects, stained preparations the overall distribution of stain does not preserve t h sey m m e t r i eostfh oe b j e c te, s p e c i a l l yt h oe u t e r p a r t s of theprojections,thusresultingintruncated cases, reconstruction has projections. In atlhl e s e Mass., USA. to b e a t t e m p t e d w i t h t h e a v a i l a b l e t r u n c a t e d p r o j e c t ions. Algorithms based Radon the on inversion as the CBP when directly used with formula such truncated projections does not properly reconstruct of view. t hoeb j e cf ut n c t i oonu t s i dtehfei e l d a d d i t i o na ,r t e f a c et sx t e nidn t ohfei e lovdfi e w [3,4,5]. Thus indirect methods are being devised so t h a t a f a i r reer c o n s t r u c t i ocnaboneb t a i n e d f rom truncated projections. In as Algebraic ReconstructI t e r a t i v em e t h o d ss u c h (ART) and Simultaneous Iterative ion Technique Reconstruction Technique (SIRT) have been investigated for reconstruction from truncated projections [6]. major A dis-advantage of t haipsp r o aicsh of c o m p u t e r t hrai ett q u i r ecso n s i d e r a b laem o u n t time. Several authors have shown that reasonable a c c u r aitm e a gw e siltehs saer t e f a cct os usl dt i l l CBP method of reconstruction b eo b t a i n e df r o mt h e [3,4,51. u s'icnogm p l e tter du 'n c apt er od j e c t i o n s The completion of projection in [3] is a c h i e v e d byestimatingthemissingdatabyfittingpolynomial f u n c t i o ntso a few end points on either side of thetruncatedprojection. . In [ 5 ] , t h ec r o s ss e c t i o n a l outline of theobjectfunctionisobtainedbyoptical is approxm e t h o d sa n dt h em i s s i n gp r o j e c t i o nd a t a i m a t e d as the p r o d u c t of the path length and a meanvalue of t h eo b j e c tf u n c t i o nw h i c hi se s t i m a t e d from the available data. This paper presents a method for c o m p l e t ittnhrgue n c a tperdo j e c t i o n by enabling using Linear Prediction theory, there to algorithms based Radon on inversion formula e x t r a csti g n i f i c a n t l m y o rien f o r m a t i o fnr o m incomplete projections. Projection completion by Linear Prediction Linearpredictormodelshavebeendeveloped for many signals for different applications such as prediction or forecasting, control, data compression and spectral e s t i m a t i o nO. n e o i the most widely a truncated investigated model for extrapolating a s p e c t r adle n s i tfyu n c t i o n signal or interpolating is the Autoregressive (AR) model. Fitting an AR model to t ha ve a i l a bol end ei m e n s i o ndaalitsa equivalent to the Maximum Entropy Method [71. M a x i m uemn t r o p yr o c e s s i ne gn s u rtehsfaetw e s t possibleassumptionsaremadeabouttheunmeasured for modelling the d a tTathm .ihiseso t i v a t i o n as an o u t p u t of a linearpredictor. p r o j e c t i o nd a t a 34.3.1 ICASSP 86, TOKYO CH2243-4/86/0000-1733 $1.00 0 1986 IEEE 1733 SimulationResults L e t g(x,y) r e p r e s e n t a t w od i m e n s i o n a ol b j e c t a c i r c l e of radius function which zero is outside T hRe a d otnr a n s f o r m [ R g(]8 , tit)sh e b (Fig. 1). :RB (Fig.1). lineintegral of t h e f u n c t i o n a l o n g t h e l i n e w h e r ed si st h e l e m e n t a dl i s t a n c eo nt h el i n eA B representedbytheequation x cos O+ y sin0 = t Theprojection [Rg] (0 ,t)is a set of l i n ei n t e g r a l s to t h e evaluated along parallel lines perpendicular 0 withthex-axis as vectorwhichmakesanangle showninFig. 1. . Consider a d a t m a atrix P w h e r ea ne l e m e n t p(i,j) represents a line integral [Rg] (i,j) and each a c o m p l e tper o j e c t i o n at aann g l e r o rwe p r e s e n t s as missing A i. truncatedprojectionmanifestsitself c o l u mtinhdnseam t aa t r iFxi. g . 2 d e p i ctths e of v i e ws i t u a t i o nw h e r teh R e O(Ir < a limited field 0 ' s 0 < 179". inFig. I ) is projected from all directions E a c h of t haev a i l a b lter u n c a t epdr o j e c t i o n , a row as a A nR process i tnhde a tm a a t r i xim ,s o d e l l e d N. p(i,j) expressed is approximately as order of a linear combination of the 'past' values. A . . where p(l,l) . IS an approximation of p(i,j) and aN(k) of o r d e r N. T he r r oor r a r teh A e Rc o e f f i c i e n t s r e s i d u able t w etehaec R t u a.vl.a l u. e . of p(i,j) and the predicted value p(1,~) IS glven by A . e(j) = p(i,j) - p ( l , ~ )= N 1 a#) p(i,j-k) k=O whereaN(0) = 1. The AR parameters are obtained minimising by to t hme e aotnor t as lq u a r eedr r owr i trhe s p e c t e a c h of tphaer a m e t e V r sa. r i oaulsg o r i t h m a rse [81. Each available t o o b t a itnhAecRo e f f i c i e n t s projection then is completed by e x t r a p o l a t i nt gh e truncated projection forwards and backwards using of t hdee r i v eldi n e aprr e d i c t oTr .hfei n i teex t e n t the object which may be known apriori or can be to determine determined experimentally used is t h ne u m b e r of points t o be x t r a p o l a t e od eni t h e r side. If d a t a points in t he ex t r a p o l a t esde g m e n t set t o zero.Thecompleted a r ne e g a t i v et,h e ya r e projection thus consists of the original truncated p r o j e cetaxihnotedrna p o l asteegdm e nTthse. completion of t h e d a t a m a t r i x i s t h u s a c h i e v e d w i t h o u t as t h eo b j e c ft u n c t i o n makinganyassumptionssuch or c o n s t a nftotrhaer eoau t s i dtehRe O I izse r o r ae a s u r e m e n t s ( a < r < b) nrioetrq u i r ae nseyx t m o u t s i d e t h e ROI. of 9 superimposed T h e test phantomcomposed (64x64) used in t h ec o m p u t esr i m u ellipses of size 3. 180 equispaced lation studies shown is Fig. in a t 91 values of t (corresprojections,eachsampled ponding t o a parallelprojection)distributeduniformly of view (a=45) encompassing o vtehfruef li le l d thewholephantomwerecomputed.Thereconstruct i t ht h i fsu l l tion of t h pe h a n t o mw a cs a r r i e do uw d a tuas i ntghCeBm P e t h o(dF i g . 4). The Rama[l] and a Hamming c h a n d r a n - Lakshminarayanfilter window was used for filtering and smoothing the projection data., This operation was carried out tfihrne q u e ndcoym auisniFn Fag lTg o r i t h m s . A weightedbackprojectionwasthenusedtoobtain t hree c o n s t r u c t eidm a g e . In o r d e r t o s i m u l a tteh e m e a s u r e m e n t s of truncated projections, the outer t o rays p a r t s of t hper o j e c t i odna tcao r r e s p o n d i n g of i n t e r e s t w h i c hd on o ti n t e r s e c tt h ec i r c u l a rr e g i o n t o zero. Three cases of t h e ROI ( r > aw) e rsee t w i t h a = 35, 30 and 25 are considered. Fig. 5a, 6a an7sdha o wt hrse c o n s t r u c t i oonb t a i n ebdy using the CBP method directly with the truncated cases a = 3 5 a, = 3 0a n da = 2 5 p r o j e c t i o n sf o rt h et h r e e r e s p e c t i v e l yT. heer r o rpsr e s e natreev i d e n t lcyo n of t ht er u n c a t epdr o j e c t i o n s siderable. Completion as mentionedintheprevious w e r et h e nc a r r i e do u t section assuming elliptical an boundary. A study of t h e p a r t i a l a u t o c o r r e l a t i o n f u n c t i o n a n d t h e r e s i d u a l for various projections showed that an AR model of order 5 was sufficient t o m o d et lhter u n c a t e d [91 d a t a in t hper e s e ncta s eT.hBe u ragl g o r i t h m whichisfound to b ee f f i c i e n tf o rs h o r td a t al e n g t h s was used t o o b t a ti hnAecRo e f f i c i e ndt si r e c t l y from the data values. Fig. 5b, 6b and 7b shows t h ee f f e c t i v e n e s s of theproposedmethod of complecases tion of t h et r u n c a t e dp r o j e c t i o n sf o rt h et h r e e r e s p e c t i v e l y .F o rt h es a k e of comparison,thereconstruction obtained by using the positive constraint additive ART method with truncated projection f otrhfei r st twcoa s eas= 3a5nad= 3a0rseh o w n in Fig. 8 a n d 9. Conclusion Theproposedmethod of c o m p l e t i n g t h e t r u n c a t e d a significantlybetterreconstprojectionsresultsin t o the reconstruction obtained ruction compar,ed by using the incompleted ones. Qualitatively the r e c o n s t r u c t e di m a g e sa r ev e r yg o o dw i t h i nt h ef i e l d of view. Also, features which are visual ly missing of viewcanbeseen in t h er e g i o no u t s i d et h ef i e l d w h etnhter u n c a t epdr o j e c t i o nasrceo m p l e t e da,l thoughquantitativeerrorscanbeobserved. A detailederroranalysisisbeingcarriedout.Anadvantage of thismethod is t h a ti td o e sn o tr e q u i r ea d d i t i o n a l of view. The time m e a s u r e m e n tosu t s i dtehfei e l d t a k e nf o rt h ec o m p l e t i o n of t h et r u n c a t e dp r o j e c t i o n s is being extended to insignificant. is The method i n c l u dtehoe t h etrw o cases - t hhe o l l o w a n tdh e limited angle situations. 34. 3. 2 1734 ICASSP 86, TOKYO Predictionessentiallyreliesonthe fact t h a t is a b l et oe x t r a c tt h ec o r r e l a t i o n t h el i n e a rp r e d i c t o r p r e s e n t in t h ed a t aT. h es i m p l ec o m p l e t i o nm e t h o d of theprojectionindivipresentedheremodelseach a one dimensional predictor model. dually using fact t hatalhptler o j e c t i o n s This overlooks the as t h e ya r ea l ld e r i v e d arenotactuallyindependent from t hsea mcer o ssse c t i o n . In o r d eteorx p l o i t this fact, further investigations are being carried out using two dimensional linear predictor models and block adaptive methods. References [I] G . N . R a m a c h a n d raA ann.dV . L a k s h m i n a r a y a n , 'Three dimensional reconstruction from radiographs : Application of Convolution andelectronmicrographs instead of Fourier Transform', Proc. Nat. Acad. Sci., vo1.68, pp2236-2240,1971. [2]A.K.Louis andF.Natterer,'Mathematicalproblems of Computerized Tomography', Proc. IEEE, vo1.71, ~ ~ 3 7 9 - 3 8 91983. , [3] R.M.Lewitt and R.H.T.Bates, 'Image reconstruction from projections', Optik, pp19-33;Part 111 : pp189-204,1978. vol. 50, P a r t I: J.C.Gore [4] and %Leeman, 'The reconstruction ofobjectsfromincompleteprojections',Phys.Med. Bioi., Vo1.25, pp129-136,1980. [ 5 ] W . W a g n'eRre, c o n s t r u c t i of nr rosems t r i c t e d - Newmeans to r e d u c et h ep a t i e n t r e g i o ns c a nd a t a dose',IEEETrans.Nucl. Sci., Vol. NS-26,pp2866-2869, April1979. [6] B.E.Oppenheim, 'More accurate algorithms for iterative 3-dimensional reconstruction', IEEE Trans. Nucl.Sci,,NS-21,pp72-77,1974. [7] A . V a n D e n B'oAsl,t e r n a tiinvtee r p r e t a t i o n maximum entropy spectral analysis', IEEE Trans. Inform.Theory, Vol.1T-17, pp493-494,1971. of [SI V.Krishnan, 'Adaptive Estimation Algorithms', of Science,Bangalore, LectureNotes,IndianInstitute India,October1984. Fig. 3 [9] J.P.Burg, 'nAeawn a l y sties c h n i q uf oet irm e at theNATOAdvancedStudy s e r i e sd a t a ' ,p r e s e n t e d Inst.SignalProcessingwithemphasisonUnderwater Acoustics,Enschede,Netherlands,1968. Fig. 4 34. 3. 3 ICASSP 86, TOKYO 1735 Fig. 5a Fig. Fig. 6a Fig. 6b Fig. 7a Fig, Fig. 7b Fig. 8 34. 1 5b 9 3. 4 ICASSP 86, TOKYO