Iterative image reconstruction with random correction

for PET studies

Jyh-Cheng Chen', Ren-Shyan Liu', Kao-Yin Tua, Henry Horng-Shing LU*C, Tai-Been Chenc, and
Kuo-Liang Choub

alflstjtute Øf Radiological Sciences, National Yang-Ming University, Taipei, 1 12, Taiwan

bNational PET/Cyclotron Center, Taipei Veterans General Hospital, Taipei, 1 12, Taiwan
Clnstitute of Statistics, National Chiao Tung University, Hsinchu, 30050, Taiwan

ABSTRACT

A maximum likelihood-expectation maximization (ML-EM) reconstruction algorithm has been developed that allows
random coincidence correction for the phantom we used and the reconstructed images are better than those obtained by
convolution backprojection (CBP) for positron emission tomography (PET) studies in terms of spatial resolution, image
artifacts and noise. With our algorithm reconstruct the true coincidence events and random coincidence events were
reconstructed separately. We also calculated the random ratio from the measured projection data (singles) using line and
cylindrical phantoms, respectively. From cylindrical phantom experiments, the random event ratio was 41.8% to 49.1% in
each ring. These results are close to the ratios obtained from geometric calculation, which range from 45.0%to 49.5%. The
random ratios and the patterns of random events provide insightful information for random correction. This information is
particularly valuable when the delay window correction is not available as in the case of our PET system.

Keywords: Random coincidence, true coincidence, positron emission tomography, iterative reconstruction

1. INTRODUCTION

'Nuclear medicine imaging diagnosis not only can detect functional and metabolic abnormality but also can do quantitative
studies, thus it becomes an important clinical diagnostic modality. Recently there is a trend using PET for metabolic and
physiological studies, for example glucose metabolic rate in the brain. In doing so, we need to generate accurate parametric
images. In order to get such accurate images for quantitative analysis, we need to use better image reconstruction algorithms.
Many commercial PET systems use the CBP technique for image reconstruction due to its fast computational speed.
However, this method has several drawbacks. First, it is derived from the transmission tomography, not the emission

*
Correspondence: Email: hslu@stat.nctu.edu.tw; Tel: 886-3-573 1870; Fax: 886-3-5728745

In Medical Imaging 2000: Image Processing, Kenneth M. Hanson, Editor,

1218 Proceedings of SPIE Vol. 3979 (2000) • 1605-74221001$15.00

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 06/15/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 tomography. It is therefore incapable of handling accidental coincidence (AC) or random coincidence (RC) events and
attenuation that occur frequently in practice. Second, it does not consider the random variations that are inherent of the
photon counts of coincidence events. Hence, the reconstructed images are typical noisy and inaccurate. Therefore, it is
necessary to consider a realistic model for PET with RC events and attenuation. Based on this model, reconstruction
methods with regularization are important to reduce the noise and edge artifact. Due to fast progress of computer technology,
computers with very fast computing speeds are developed. Thus, using iterative reconstruction algorithms for such as ML-
EM', in clinical setting have been reported recently. This paper describes our work on RC correction using ML-EM and
shows some comparisons with CBP reconstructed images.

2. MATERIALS AND METHODS

The principle of PET is based on annihilation coincidence detection and tomographic images are formed through
reconstruction step2. The PET scanner is Scanditronix PC4096-15WB whole-body scanner3. A total of 4096 crystals are
arranged in eight rings of 512 crystals each. Two dual photomultiplier tubes (PMTs) are mounted on each assembly of 16
crystals. Transaxial field of view (FOV) is 55 cm and axial FOV is 9.75 cm. We used Ga68 that were produced by Ge68 —
Ga68 generator with a half-life of 68 mm4. In order to understand the detection, performance characteristics and data format,

we designed the following experiments; then used those data for image reconstruction.
(1) A line phantom with an inner diameter of 0. 1 cm and 15 cm of length were fixed on a patient couch. PET was
acquired for 20 mm to obtain original sinogram data.
(2) In order to compare the reconstructed images between ML-EM and CBP, we used two test phantoms. One has 5
line sources in the FOV, and the other has 1 9 line sources.
(3) The positions of 2 line sources are fixed at two positions that are in the opposite side about the central FOV.
Acquisition time was 10 mm for each sinogram. Three experiments are performed in order to account for RC events. The
first experiment located a thin line source in the left side. Because there was only one thin line source in air, random or
scatter coincidence (SC) events were reduced to minimal. The detected coincidence events were mainly true coincidence
(TC) events. Then, the next experiment put in another thin line source in the right side. Since there were two thin line
sources in air, true, random, and few scatter coincidence events were recorded simultaneously. The third experiment
removed the thin line source in the left. Again, only one thin line source existed and mainly truce coincidence events
were recorded.
(4) Replaced line phantom with a cylindrical phantom, and 1 mCi of Ga 68 was distributed uniformly in this water
phantom. Acquisition time was 10 mm.

2.1. The ML-EM Algorithm

According to the configuration of V.G.H.-Taipei PET system, the target image is partitioned into B=128* 128 boxes (pixels)
and there are D256* 193 detector tubes recorded in a sinogram of a slice. The intensity at each box, denoted by

1219

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 06/15/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 'r '(q) q = ''T 'g" suosoidai oq '!suu! U!4!M oq &iipuodsauoo •xoq qj ioqmnu Jo ouopiornoo
S2UOAO ppioooi UT qoo .iopaop iatd si pooup
*
q u '(p) p = i z' '•' a • s jous opoui u) oiJdwoou!

(oids si puinss o q

(p) piui UOSiOJ ?) '((P) (1)

'is' (p) =
j=q
'q)d '(p (q) (z)

OJH PF'-'! cpupu0dpU, P °I plnq3Sp y (P) ! °'P OJOA13 '"! Jo UOp!OUiOO

SUAO E ioooop oqni 'p 'q)d (p s oq uosu ijiqqoid woij xoq q o ioooop qm p im opnjouiat uoinuo
uoonoo ddoq) pim 'i!PA '!PA 'ddqg pu (cUIUJ1W)J icj osothnd jouicjdd oq F1 'uiio ouo uo umss
oq oidmoooods opoui s
U (p'q) Plu' (p'q)d)uoss.zoj ((q) . (i:)

OqJ. JA-flJAJ W4!JO1E ST AUO s sMojjoJ qj pooqjo-oj uoiounjut ojdwooui gp oods s!

a g * * *
'r)i (= }
t=p j=q
u (p)lul ? p) [( 'r p) {( + :Rn3:sUoz,
— . (t7)

UJ 'SJ.UOO 0142 pOOq!jO)jij-O UOUOUflJ U! oq oAoq jdwoo I2Ep oods s!

i 'k') (=
1=qI=p
u} 'q) (p 'q)j]u'r(p [(q) — 'q)d ?(P {(q) + uvjsuo: . (c)

qj ds- spou o ondwoo oq uopuo uouodxo qjo oojdwoo ods pOOqLj!j-O I s!

ci g q)aT (q)10?(p
= 3j} ') u ' {( = u} (p) ? (p) 'q)d]u 1'(P [(q) — 'q)d {(q)(p + JUVJ8U03 C

(9)

(p) = p,o?(p'q)d' (q) (L)

qj ds-j sinbi T4 UO!Z!a[iXtU uouodx ui ds- siqjuo q pAiqo 1q U!uoI

= 'o q = ''T . . . ' .g (8)
(q)

L ozz

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 06/15/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 &) D p(b, )2old(b) i
ai (b)
={n(d) d=1 2*old (j) ..Z (b)
—p(b,d)}=O. (9)

Hence, the updating formula is

2(b) = 2old(b) D n (d)

p(b,.) d=1
(b' d
/Z old (d)
(10)

where
D
p(b,.)=d=1>p(b,d). (11)

In summary, the MLE-EM algorithm for PET is as follows.

'Algorithm 1 . The EM Algorithm for PET:

Step 1 . Initial: ,old (b) b = 1,2, . . . , B is obtained by the method of moments estimate (MME), the convolution

backprojection (CBP) method, or other methods.

Step 2. Project: 2old (d) = p(b, d)201d (b) for d=1,. . .,D.

Step 3. Backproject: for b=1, 2, ..., B,

2°(b)
,%old(b) D
p(b,.)
n (d)
d=1
(b d
old
2* (d)

Step 4. Stop: If 1 (2 ) — 1 (2l ) < tol , for a tolerance level,

then stop. Else, 2old (b) anew (b) , and go to Step 2.

2.2. The ML-EM Algorithm for Detecting Random Coincidence Events

A new model for these three sinograms by considering decay correction is proposed as follows. Suppose

n(d) i.n.i.d. Poisson (c12 (d)), i = 1,2,3, d = D, (12)

where n (d), i = 1, 2, 3, are mutually independent,

B
2, (d) = p(b, d)21 (b), I = 1,2, (13)
b=1

2 (d) = p(b, d)(23 (b) + 2Rc (b)), Rc(b) O, (14)

1221

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 06/15/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 and
..Z3(b) = )t1(b)+22(b). (15)

For example, the half-life of Ga68 radionuclide is 68 minutes and the exponentially decay rate is k = !_ . Therefore, the

decay factors for three sinograms are

Cl = e' (1 — et )I(kt1),i — 1,2,3. (16)
The log-likelihood function in the above incomplete data space is

l(A "2 ,k) = {n (d)1n[c1; (d)] - c2 (d)} + constant . (17)

d=lj=1

The MLE-EM algorithm for this new model can be derived as follows. Suppose the complete data space is

n(b,d) i.n.i.d. Poisson (c1p(b,d)2, (b)), I = 1,2 , (18)

n3(b,d) i.n.i.d. Poisson (c3p(b,d)(23 (b) + ARC (b))) , (19)

for b = 1,2, , B, d = 1,2, , D and n1 (b, d), i = 1, 2, 3, are mutually independent. Then the log-likelihood function
in complete data space becomes

l(A,A2,k) = (b,d)1n[cp(b,d)2(b)]
d=lb=1 j=1

+ n3 (b, d) ln[c3p(b, d)(23 (b) + ARC (b))]

—
p(b,d)[c1A1 (b) + c222 (b)]
—
c3p(b, d)[23 (b) + ARC (b)] }

+ constant . (20)

The E-step computes the conditional expectation:

Q= E{l R "2 ,ARC )n , A*old , i = 1,2,3)}

DB2 * p(b, d)A01!dl (b)

= n1(d) 1n[cp(b,d)A1(b)]
A*old(d)
old old
+ n (d) p(b, d)(A3 (b)± ARC ( ln[c3p(b, d)(A3 (b) + ARC (b))]
—
p(b, d)[c1A (b) + c2A2 (b)J
— (21)
c3p(b, d)(A3 (b) + ARC (b))} + constant,

where

1222

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 06/15/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 B
2*old (d) = : p(b, d)2"
b=1
(b) j = 1,2 , (22)

B
2*old (d) = : p(b, d)(2 (b) + )t (b)) ,
b=1
(23)

and

21d (b) = o1d (b) + 21d (b) . (24)

The M-step needs to maximize the conditional expectation in the E-step. This can be achieved by solving
=0 =O,forallj=1,2,b=1,2,•••,B. Since
821(b) 'a2RC(b)
D
aQ =
_____ {n(d) p(b,d)2"(b)
A,*old (d) 2 (b)
1

82 (b) d=1

+ n (d) p(b, d)(2r (b) + 2 (b)) 1

,;o1d (d) 23(b) + RC (b)

—c1p(b,d) —c3P(b,d)} = 0,] = 1,2, (25)

and

3Q =
_______
D * p(b, d)(2r (b) + 2 (b))
{fl3 (d) ________________________________
1

a2RC (b) d=1 A,*old (d) 23(b) + RC (b)

— c3p(b, d)} =0.
(26)
Let

2old(b) D
A1 (b) = n (d) p(b, d) = 1,2, (27)
p(b,.) d=1 A,*o/d (d)

and

=23 (b) 2 (b)

+ *
p(b, d)
A3 (b) (d) (28)
p(b,.) d=1 ,old(d)
Then the above equations (25) and (26) become

A1(b) + A3(b)
=c +c3, 1 (29)
21(b) 2(b)+22(b)+2Rc(b)
4(b) + 4(b)
=c2+c3, (30)
22(b) 21(b)+ 22(b)+ 2Rc(b)
and

1223

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 06/15/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 A3 (b)
=c .
(31)
,%l(b)+A2(b)+2Rc(b)
In order to maximize the conditional expectation Q and maintain the nonnegativeness of the new estimates, the following
generalized EM (GEM) algorithm is considered.
S
Algorithm 2. The GEM Algorithm for detecting RC events:
Step 1 . Initial:

2old (b) for b = 1,2, . . . , B by the MME, CBP, or other methods.

A(b) =0.

Step 2. Project: 2*old (d) for d = 1,2,• . •,D by (13), (14), (15).

Step 3. Backproject:

A1 (b) by (27), (28).

2new(b)< 4(b),
Cl

A2(b)
2new(b)<
C2

2new (b) < 2 (b) + )t4ew (b)

,%$,(b) < max{(----),0}.

C3 C1 C2

Step 4. Stop:

If l(2Jv, A7C', j =1,2)

—
1%o1d , , j = 1,2) <tol for a tolerance level, then stop. Else,

gold (b) Anew (b), A (b) — A (b), j = 1,2, and go to step2.

2.3. Random Coincidence Event Rate

The single rays were used to calculate the random coincidence event rates by the formula: RAC= 2cRR where-c is the

coincidence resolving time of the detectors and associated electronics, and R or R is the single-channel6, i.e., the

noncoincidence count rate in detector i or j. These RC rates were calculated using the raw data from the sinograms and were
compared to the results obtained by geometric method. The geometric method uses the following formula:

1224

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 06/15/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 Rs 2E e g exp(—pD I 2)
RT _ E * e2 g exp(—pD)

RAC =2*r*R
where E is the source emission rate (positron/sec), c is the intrinsic efficiency of a detector pair (assumed to be equal), i.e.,
fraction of incident photons detected, and g is the geometric efficiency of an individual detector, i.e., fraction of emitted
photons intercepted by the detector. D is the object thickness and i is the linear attenuation coefficient ofthe object.

3. RESULTS AND DISCUSSION

In 5 line-source experiments, we aligned the phantom horizontally and 1 7 degrees from the horizontal direction,
respectively, and acquired data for 30 mm in each phantom orientation. We then calculated the transition matrix based on
the geometry and did the EM reconstruction7 to get tomographic images. A representative slice image is shown in Fig. IA.
The same data reconstructed by the CBP is shown in Fig. lB. Figures 1C&D show the corresponding images when the
phantom was tilted 17 degrees. From the images it is clear that we can easily see 5-points images reconstructed by the ML-
EM method. However, the images reconstructed by the CBP show some artifact and noise.

Figure 2 shows the reconstructed images of the 1 9 line sources. The interval between each line is 1 cm and the phantom
was imaged horizontally and 5 1 degrees from the horizontal direction. Again, the images reconstructed by the ML-EM
method show less blurring than the ones obtained by the CBP method.

Assume for a single line source, there are only TC events, however, in the case of two line sources there will be RC in
addition to TC. From our experiments, we found RC events were roughly uniformly distributed in the whole FOV using
iterative ML-EM reconstruction after 10 iterations, which was chosen based on our previous studies89. Examples of three
experiments were shown in Fig. 3. The reconstructed images can be displayed as four parts, including the left source, the

right source, two sources and the RC events which show (b), '2 (b), 2 (b) and 2RC (b) respectively. They are

demonstrated in Fig. 4.

Using single-channel rate from original sinogram data, we found RC count I TC count ranges from 0. 13% to 0. 19% for 2
line sources. From measurements, BGO detector is 0.3 cm wide and 1.2 cm long, therefore g = O.3x1.2/(4x3.14x50.5x50.5).
The intrinsic efficiency of BGO is 95%. Thus, RC count /TC count ranges from 0.08% to 0.27%. For uniform source in
the cylindrical phantom, RC count /TC count ranges from 41.8% to 49.1% using raw sinogram data compared with 45.0%
to 49.5% using geometric method, see tables 1-3. We see for both uniform source or 2 line sources, RC/TC using
experimental data and RC event formula is close enough to the RC/TC obtained by geometric method, which validates the
formula of RC event rate is correct.

1225

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 06/15/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 4. CONCLUSIONS
The ML-EM does outperform the CBP as compared in this study. Furthermore, the random coincident events can be
estimated by experimental studies as suggested in this study. We also explore the relationship between true and random
coincident events. The ML-EM reconstructed images are superior to those obtained with CBP. Furthermore, the random
coincident events can be estimated by experimental studies as suggested in this study. The patterns of random coincidence
events provide insightful information for random coincidence event correction, which the hardware correction by the delay
window can not provide. This information is particular valuable when the delay window correction is not available as in the
case of Taipei Veterans General Hospital PET system. We also explore the relationship between true and random coincident
events. The results obtained by geometric method were close to the results calculated from experimental data, which
verified the random coincidence rate formula.

REFERENCES
1. L A. Shepp and Y. Vardi, "Maximum likelihood reconstruction for emission tomography," IEEE Trans. Med. Imaging 1,
pp. 113-122, 1982.

2. Spinks TJ, Jones T, Gilardi MC, Heather JD. Physical performance of the latest generation of commercial positron
scanner," IEEE trans. Nuci. Sci. 35, pp. 721-725, 1988.
3. SCANDITRONIX PC-4096-1 5WB Whole body PET scanner Service Menu.
4. G. B. Saha, Fundamentals ofNuclear Pharmacy, 1998.
5. Y. Vardi, L. A. Shepp, and L. Kaufman, ''A statistical model for positron emission tomography," Journal of the
American StatisticalAssociation. 80, pp. 8-20, 1985.
6. J. A. Sorenson and M. E. Phelps, Physics in Nuclear Medicine, 2ed. Saunders, Philadelphia, 1987.

7. K. Lange and R. Carson, "EM reconstruction algorithms for emission and transmission tomography," .1 Comput. Assist.
Tomogr. 8, pp. 306-3 16, 1984.
8. H. H.-S. Lu, C.-M. Chen, and 1.-H. Yang, "Cross-reference weighted least square estimates for positron emission
tomography," IEEE Trans. Med. Imaging. 17, pp. 1-8, 1998.
9. H. H.-S. Lu and W.-J. Tseng, "On accelerated cross-reference maximum likelihood estimates for positron emission
tomography," IEEE Nuclear Science Symposium 2, pp. 1484-1488, 1997.

1226

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 06/15/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 (A) (B)
ML-EM reconstruction CBP reconstruction

Figure 1. Image of 5 line sources reconstructed (A) with ML-EM algorithm and (B) with
CBP reconstruction method.

ML-EM reconstruction CBP reconstruction

Figure 2. Image of 19 line sources reconstructed (A) with ML-EM algorithm and (B)
With CBP reconstruction method.

1227

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 06/15/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 880211 dO s9 em 9 880320 calOb s9 em 9 880403 calOb s9 em 9
Activity ratio 1:1 Activity ratio 3:2 Activity ratio 3:1

Figure 3. Random coincidence distribution in different activity ratios.

C D
Figure 4. ML-EM reconstruction of line source, (A),(B): single source reconstructed image,
(C): two line sources reconstruction, (D): random coincidence event image.

1228

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 06/15/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx

 Table 1. The calculated ratio ofrandom vs. true coincidence events by the formula
based on single rays for uniform sources.

UrFaMsaJE
EllGR1Rll RD4CXNf 1RIE GN
_________
RANXIRW/o)
Ring 1 400988 8.68E+05 46. 19614
Rh2 378268 8.62&05 43.89625
Riiig3 368J1 a82E+05 41.75566

Ring4 423399 a62E-05 49.10676

Rfrig 5 399049 8.9E-05 44.85%O

Ring6 384667 a87E+05 43.39004
Rüg7 381958 &- 43.01125

Rhig 8 392686 8.84E+05 44.40572

Table 2. The calculated ratio of random vs. true coincidence events by the formula
based on single rays for line sources.

LINE SOURCE
LINE SOURCE RANDOM COUNT TRUE COUNT RANDOM/TRUE(%)
4.4uCi 18.823 14179.177 0.132751005
8.l83uCi 36.525 26806.475 0.136254394
3.8uCi 16.6471 12341.3529 0.134888777

— 7.88uCi 26.0559 15373.9441 0.169480908

13.35uCi 41.0018 21550.9982 0.19025476
5.47uCi 11.6482 6538.3518 0.178151931
5.luCi 22.53 17502.47 0.128724689
9.76uCi 48.0711 32459.9289 0.14809367
4.67uCi 20.2428 15301.7572 0.132290689

Table 3. The calculated ratio of random vs. true coincidence events by the formula
based on geometric calculation for line and uniform sources, respectively.

LINE SOURCE tANDOM COUNT TRUE COUNT RANDOM/TRUE (%)

4.4uCi 21891.8 2.43E+07 0.08994166
8.l83uCi 75742.5 4.53E+07 0.16725737
3.8uCi 16333.6 2.IOE+07 0.077670679
7.88uCi 70237.2 4.36E+07 0.161064942
13.35uCi 201594 7.39E+07 0.272870504
5.47uCi 128.843 116733 0.110374102
5. luCi 29420.8 2.82E+07 0.104244056
9.7596uCi 107749 5.40E+07 0.199490854
4.6681uCi 24563.3 2.58E+07 0.095248676
UNIFORM
lmCi 62110 1.38E+05 44.96716695
1.lmCi 75153.1 1.52E+05 49.46398131

1229

Downloaded From: http://proceedings.spiedigitallibrary.org/ on 06/15/2016 Terms of Use: http://spiedigitallibrary.org/ss/TermsOfUse.aspx