Вы находитесь на странице: 1из 5

Published by Elsevier Science on behalf of IFAC ELSEVIER

IFAC
PUBLICATIONS
www.elsevier.com/locale/ifac

AN IMPROVED OPTIMAL IMAGE SAMPLING SCHEDULE FOR MULTIPLE


ROIS IN DYNAMIC SPECT

Lingfeng Wen····, Stefan Eber..·····, Dagan Feng..·····, Jing Raj'

* Dept. ofBiomedical Engineering, Tsinghua University, Beijing. 100084, China


** Biomedical and Multimedia Information Technology (BMIT) Group.
School ofInformation Technologies. the University ofSydney. NSW 2006. Australia
*** Dept. of PET & Nuclear Medicine, Royal Prince Alfred Hospital. Missenden Road.
Camperdown. NSW 2050, Australia
**** Centerfor Multimedia Digital Signal Processing.
Dept. ofElectronic& Information Engineering.
Hong Kong Po~vtechnic University. Hung Hom. Kowloon. Hong Kong

Abstract: Optimal image sampling schedule (OISS) has been applied successfully to
dynamic images for positron emission tomography (PET) with improved parameter
estimation precision and significantly decreased image storage and processing
requirements. OISS is usually based on one average curve for tissue, and the performance
of OISS is still unclear when multiple regions of interest are considered in clinical
diagnosis. An improved OISS is investigated with a modified objective function through a
simulated study of the tracer 5- '23 1-iodo-A-85380. The study shows that the improved
OISS is at least as efficient in providing the accuracy and precision of parameter
estimates as conventional and traditional OISS.

Keywords: Sampling Intervals, Monte Carlo Simulation, Parameter Estimation,


Physiological Model, Quantitative analysis

I. INTRODUCTION time, with the conclusion that there was a monotonic


improvement in the index of parameter accuracy with
In the traditional dynamic system identification increasing sampling frequency, especially in the early
experiments, instantaneous measurements were stage. An optimal image sampling schedule (OISS)
usually obtained from sampling at discrete-time was proposed by Li. et at. (1996) with a modified
points according to empirical sampling schedules. cost function based on integrated activity
Since DiStefano (1979) introduced the advantage of concentration for PET modelling. The methodology
optimal sampling schedule (OSS), several algorithms was then successfully applied to investigate 01SS of
have been investigated to find OSS based on the input function, output function and whole-body
Fisher information matrix criterion for instantaneous image acquisition for PET (Feng. et at. 1997;
discrete-time sampling (Cobelli. et at. 1985; HoShon. et at 1996; Li. et at. 2001; Li. et at. 2000)
Distefano 1981, 1982; Mori and Distefano 1979). with the advantage of improved parameter estimation
precision and significantly decreased image storage
OSS for positron emission tomography (PET) was and processing requirements. Furthermore, OISS has
investigated for temporal sampling of the glucose been found applicable for dynamic SPECT of TI-201
model using 18-fluoro-deoxy-D-glucose (FDG) by (Lau. et al. 1998).
Hawkins. et al. (1986). Delforge. et al. (1989)
presented the general framework and various criteria OISS is usually derived from one tissue time activity
to estimate receptor-ligand reaction model curve (TTAC) that is simi lar to the average of the
parameters with dynamic PET data. These methods TTACs. The performance of OISS is still unclear for
were based on D-optimal criterion and a cost multiple regions of interests. Therefore an improved
function, assuming that sample points represent OISS was investigated for the nicotinic receptor
instantaneous activity concentration at the sample tracer 5- 123 I-iodo-A-85380 in this work, to consider

139
15()J· .....---,..- ~-'-- .•- .-~.~-
interval is adjusted in turn in the direction which
increases the determinant sum of the information
matrices with the TTAC resampled with the adjusted
interval. When the interval k is adjusted, the adjacent
interval k+ I is adjusted by the same amount in the
oppositc direction to maintain the same total
collection time. The interval will be merged with the
next interval if the determinant for the merged
interval keeps increasing and the duration of it falls
below the lower bound for intervals (10 s in this.
PTAC
study). If the determinant decreases during the
00 20 «l 60 80 100 120 140 160 180 200
Tme(mn) adjustment, a simple local search is carried out for a
maximum with adjustment of interval k and k+ I.
Fig.1 the plasma tissue activity curve (PTAC) and
TTACs for cerebellum, frontal cortex and 2.2 Evaluation
thalamus
Dynamic projection data were generated from high-
the overall effccts of multiple regions of interests count Monte Carlo simulations of a mathematical
with a modified objective function. Monte Carlo human brain phantom and using the kinetic
simulations were perfonned to generate dynamic parameters listed in the Table I, which were derived
projection sets from a brain phantom with different from the experiments of neuronal nicotinic
Poisson noise levels in line with the kinetics of the acetylcholine receptor (nAChR) tracer 5_[,23I]_iodo_
selected tracer. Kinetic parameters of interest were A-85380 (Kassiou. et al. 200 I). The projections were
estimated and compared to the known values used in scaled and Poisson noise was added to generate
the simulation to evaluate the different sampling projections sets without noise and with eight levels of
schedules. noise based on experimental count levels, with
maximum pixel count for a similar 5 min frame,
2. METHODS ranging from 15 to 50 counts. Images were then
reconstructed using Ordered Subsets Expectation
2.1 Optimum Image Sampling Schedule Maximization (OS EM) with moderate number of
subsets and iterations equivalent to 40 traditional
OISS aims to minimize the covariance matrix of the iterations and attenuation correction, but without any
estimated parameters by rearranging the sample post-reconstruction filtering (Wen. et al. 2002). The
intervals. According to D-optimality criterion and TTACs were derived from the reconstructed data,
Cramer-Rao theorem. minimization of the objective which was resampled according to the respective
function (/J,(p) is approximately equivalent to optimum sampling schedules.
maximizing the detenninant of the Fisher
infonnation matrix M, which is a measure for Table 1 The reference parameter sets for the regions
parameter infonnation in the data (Li. et al. 2001). of interests

CI>,(pl = Min (COVAR ( p) '" Max (IMI) (I) Cortex


Cerebellum 0.275 0.063 0.029 0.035
Here p is an unbiased estimate of the parameter Frontal cortex 0.277 0.059 0.038 0037
Thalamus 0.284 0.061 0.143 0.041
vector p. OISS depends on the particular shape of
TTAC. Usually there are mUltiple regions of interest The TTAC were fitted with a four-parameter, three-
in clinical diagnosis. which may have distinctly compartment model (Figure 2) using non-linear least
different TTACs, e.g., cerebellum, thalamus and squares fitting. Bias and reliability of the kinetic
frontal cortex for the tracer 5-[1231]-iodo-A-85380 parameters (K I -k4 ) for thalamus, cerebellum and
(Kassiou. et at. 200 I) as shown in Figure I. Thus a frontal cortex were estimated. Bias was calculated by
modified objective function (/Je(P) was proposed to comparing the estimated values with the known
consider the effect of the information matrix for all parameters in Table 1. The coefficient of variation
the regions of interest with the adjustment of (CV) was calculated from the standard deviation
sampling intervals as follows, derived from the fitting, and divided by the known
parameters in Table I.
<l>,(p) = Max (~]M.I) (2)
,.1

Therefore OISS was investigated through iterative


adjustment of sampling schedule in order to
maximize the objective function. starting from the
conventional sampling schedule (CSS) which has a
relatively large number of frames and short frame
durations initially and increasingly longcr frame Fig 2. Definition of rate constants for 3-compartment
durations during the later parts of the study. Each model used in the curve filling

140
cerebellum. OlSS3 generally displayed the highest
3. RESULTS bias for the frontal cortex. For the thalamus, the
OlSS techniques performed as well as CSS, except
3. J Optimum Sampling Schedule for k4 in the thalamus, where OISS techniques
showed higher bias.
OISS I, OlSS2, OlSS3 were based on the individual
TTAC respectively for cerebellum. frontal cortex and CV for the OlSS techniques (see Figure 4) was again
thalamus using the traditional objective function similar to CSS, with again OlSS3 demonstrating
l/>,(p). OlSS4 was searched for with the modified overall slightly higher CVs. A similar trend for bias
objective function l/>;>(p), which considered all the and reliability was also found for the estimated
regions of interest for nAChR study. Table 2 shows parameters at the other noise levels. Considering the
the results of OISS as well as the conventional overall effects on the bias and reliability of estimated
sampling schedule (CSS). parameters (Kt -k4 ) as compared with CSS, the OISS
based on the modified objective function (OISS4)
Table 2. Results of sampling schedule provides improved parameter estimates over those
based on traditional function.
Protocol Sampling schedule (min)
CSS* 15 x I min, 9x 5 min, l2x 10 min
OISSI 7.9,20.3,85.5,66.3 3.3 Effect ofnoise
OISS2 7.0, 15.6,84.9,72.5
OlSS3 6.8,15.7,87.3,70.2 The effect of noise for the sampling schedule was
OlSS4 7.3, 21.8, 81.3, 69.7 investigated between the conventional and optimal
*CSS: Conventional Sampling Schedule, 15 x I min sampling schedule, i.e. CSS and OISS4, for the
represents 15 frames, I min each etc parameter K/, which is one of primary interests for
the nAChR study.

3.2 Effect ofoptimum sampling schedule OISS achieves better estimation of parameters than
CSS for cerebellum and thalamus especially at the
Figure 3 and 4 show the comparisons of the higher noise levels with smaller bias, while
percentage bias and CV for the estimated parameters maintaining the same reliability (CV). While the bias
respectively as a function of the different sampling of OISS is not always less than that of CSS (see
schedules at the noise level of maximum pixel count Figure 5), it provides more consistent results and
of 30 in the selected frame. The selected sampling displays less sensitivity to the noise level than CSS.
schedule had little effect on the bias for the

---e- cerebe'lm --a- fronal cortex ---b- thalamus

0 100

~-10

~ 20
m-
G>

'"
to
..
~
to
m
G>

'"
to
50

0
/
r
'E -30 'E
~
G>
~
G> G> -50
Q. -40 Q.

-50 -100
CSS 015S1 015S2 015S3 015S4 CSS 015S1 OlSS2 OlSS3 OlSS4
(a) (b)

150 30

j\
20
~ 100
..
.l!!
5
~
.l!!
.. 10

m m 0
G> G>

'"
to
'E 0
/ to
'E
'" -10 \,

/~
I
G> G>
~ ~ -20
G> G>
Q. -50 Q.
-30 \
'-
-100 -40
CSS 015S1 015S2 015S3 015S4 CSS 015S1 OlSS2 OlSS3 OlSS4
(C) (d)

Fig 3. Percentage bias of estimated K , (a), kc (b), k j (c) and k~ (d) as a function of varying sampling schedules

141
~ cerebelun -a-- fronal corte x ---i:T- thatamus

8 25
Q
1\
/ \

j~~
6 20

~ ~ 15 ! \
>
u
> /
u

8sLs----:o---S~S---1--0:--IS~S:--2--0---IS~S---3::----0-::--:'1SS4 tss OISS2 OISS3 OISS4


(a) (b)

50r----__- - - ~ - - ~ - - _ _ , 30r---~---~--~---,

40

~30
>
u 20

10 5

O'-------~---~--~_::_::_-_____:__::'
CSS 0lSS1 0lSS2 0lSS3 01SS4 8S:.S;;-----=O:::IS:-:S:::1;-------:0I~S;:::S2;;---0I=S-:;:S:::-3
--::OI=SS4
(c) (d)

Fig 4. CV of estimated K, (a), k} (b), k J (c) and k 4 (d) as a function of varying sampling schedules

4. CONCLUSION pixel count for a similar 5 min frame in the


generation of dynamic projection
This study shows that improved OISS based on TTACs
from multiple regions promises to provide improved dynamic images compression (Ho. et at. 1997), fast
parameter estimates compared with those obtained parameter estimation (Feng. et at. 1999) and content-
from OISS of a single region. Bias and CV of based retrieval (Cai, et al. 2000).
parameter estimates are at least as good as those from
CSS, thus allowing advantage to be taken from the
reduced computational burden of the small number of ACKNOWLEDGEMENT
frames of OISS. The smaller number of longer frames
of the OISS technique implemented also reduces the This work was supported by the Australian Research
sensitivity of the parameter estimate bias to noise. Council and the University Grant Council of Hongkong.

Furthermore, OISS for 1-123 provides the possibility to


further study dynamic SPECT in wider areas such as REFERENCES

S
Cai. Y.D., D. Feng and R. Fulton (2000). Content-
::!.L..!=_gs::....:ISS::...:.4--'--J1 based retrieval of dynamic PET functional images.
IEEE T 1nl Technol. Biomed., 4,152-158.
;;i
~ , Cobelli, c., A. Ruggeri, J.J. Distefano, et al. (1985).
'"'l1 I Optimal-Design of Multioutput Sampling
CD I
gj,.42 " Schedules Software and Applications to
III
C I Endocrine Metabolic and Pharrnacokinetic Models.
ID i
2-441 IEEE TrailS. Biomed. Eng., 32,249-256.
ID I
Q. I Del forge, J.. A. Syrota and B.M. Mazoyer (1989)
-46[ Experimental-Design Optimization - Theory and
Application to Estimation of Receptor Model
-480'-------=50~-45~-~40--3~5--30~ 25 20
Parameters Using Dynamic Positron Emission
NoISe level
Tomography. Phys. Med. BioI., 34, 419-435.
DiStefano, J.J. (1979). Matching the model and the
FigS Percentage bias of estimated K, for frontal cortex
experiment to the goal: data limitations,
as a function of increasing noise level. 0 denotes
complexity and optimal experiment design for
noise free, and other numbers represent maximum
dynamic systems with biochemical signals. J.
Cvbern. Illf Sci., 2, 6-20.

142
Distefano, 1.1. (1981). Optimized Blood-Sampling Kassiou, M., S. EberL S.R. Meiklc, et aL (200 I). In
Protocols and Sequential Design of Kinetic- vivo imaging of nicotinic receptor upregulation
Experiments. Am. J. Physiol., 240, R259-R265. following chronic (- )-nicotine treatment in baboon
Distefano. 1.1. (1982). Algorithms, Software and using SPECT. Nucl. Med BioI., 28,165-175.
Sequential Optimal Sampling Schedule Designs Lau, CH., S. EberL D.G. Feng, et a!. (1998).
for Pharmacokinetic and Physiologic Experiments. Optimized acquisition time and image sampling
Math. Comput. Simul., 24, 531-534. for dynamic SPECT of TI-20 I. IEEE Trans. Med
Feng, D., D. Ho, K.K. Lau, et aL (1999). GLLS for Imaging, 17,334-343.
optimally sampled continuous dynamic system Li, X.1., D. Feng and K.P. Wong (2001). A general
modeling: theory and algorithm. Comput. Meth. algorithm for optimal sampling schedule design in
Programs Biomed., 59, 31-43. nuclear medicine imaging. Comput. Meth.
Feng, D.G., X.1. Li and W.C Siu (1997). Optimal Programs Biomed, 65, 45-59.
sampling schedule design for positron emission Li, X.1., D.G. Feng and K.W. Chen (1996). Optimal
tomography data acquisition. Control Eng. image sampling schedule: A new effective way to
Practice,5,1759-1766. reduce dynamic image storage space and
Hawkins, R.A., M.E. Phelps and S.C Huang (1986). functional image processing time. IEEE Trans.
Effects of Temporal Sampling, Glucose Metabolic Med. Imaging, 15, 710-719.
Rates, and Disruptions of the Blood-Brain-Barrier Li, X.1., D.G. Feng and K.W. Chen (2000). Optimal
on the Fdg Model with and without a Vascular image sampling schedule for both image-derived
Compartment - Studies in Human-Brain Tumors input and output functions in PET cardiac studies.
with Pet. J. Cereb. Blood Flow Metab., 6, 170-183. IEEE Trans. Med. Imaging, 19,233-242.
Ho, D.. D. Feng and K. Chen (1997). Dynamic image Mori, F. and 1.1. Distefano (1979). Optimal
data compression in spatial and temporal domains: Nonuniform Sampling Interval and Test-Input
theory and algorithm. IEEE TRANS INF Design for Identification of Physiological Systems
TECHNOL BIOMED, 1, 219-228. from Very Limited Data. IEEE Trans. Autom.
HoShon, K., D.G. Feng, R.A. Hawkins, et al. (1996). Control, 24, 893-900.
Optimized sampling and parameter estimation for Wen, L.F., S. Eberl, K.P. Wong, et a!. (2002). Effect of
quantification in whole body PET. IEEE Trans. reconstruction and filtering on dynamic SPECT
Biomed. Eng., 43,1021-1028. kinetic parameter bias and reliability. J. Nucl. Med,
43, 831.

J·B

Вам также может понравиться