Image Quality Assessment of Pixellated Systems

Transcription

1 Image Quality Assessment of Pixellated Systems Andreas Goedicke, Herfried Wieczorek, Henrik Botterweck, Wolfgang Eckenbach, Ling Shao, Member, IEEE, Micheal Petrillo, Member, IEEE, Jinghan Ye, and John Vesel, Member, IEEE Abstract-- The recent development of pixellated solid-state detectors for SPECT provides new opportunities for better image quality when the impact of energy resolution, spatial response and reconstruction technique is combined. In this investigation we assess the image quality of cadmium zinc telluride (CZT) based solid-state detector SPECT systems by use of Monte Carlo simulation, implemented in GEANT 4. Images are evaluated assuming different spatial and energy resolution, reconstruction algorithms and filtering. Phantoms for NEMA resolution and efficiency measurements, test phantoms for noise properties, and the Jaszczak phantom are used for this study. Monte Carlo results have found to be consistent with measured data and with analytical results. Jaszczak phantom simulations show that the better spatial resolution of pixellated detectors enables better lesion detectability compared to NaI(Tl) detectors. A hypothetical value of E = 3% gives better contrast than the typical E = 9.5% of an Anger camera. On the other hand, there is hardly any difference when energy resolution is improved from E = 5% to E = 3% for a CZT detector. Reconstruction shows a strong influence on SPECT image quality. Appropriate filters in the back-projection, and statistical reconstruction give better small lesion detectability and higher contrast-to-noise ratio, though peak-to-valley ratios and spatial resolution are reduced. P I. INTRODUCTION ixellated detectors, based on semiconductor detector materials like cadmium zinc telluride (CZT), will provide new opportunities to enhance the image quality of SPECT systems [1]-[6]. Their performance, expressed in NEMA defined parameters like system planar efficiency or spatial resolution, depends strongly on collimator design, detector energy resolution, and image reconstruction method. Those NEMA parameters, as well as signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR), are described by an analytical model that we developed recently [7]. Additionally, attenuation and scatter in the object, penetration of collimator septa, and reconstruction effects have to be considered. For Manuscript received October 4, Andreas Goedicke, Herfried Wieczorek, Henrik Botterweck and Wolfgang Eckenbach are with Philips Research Laboratories, Aachen, Germany (telephone: , Ling Shao, Micheal Petrillo, and Jinghan Ye are with Philips Medical Systems, Milpitas, CA 95035, USA (telephone: , ling.shao@philips.com). John Vesel is with Philips Medical Systems, Cleveland, OH 44143, USA (telephone: , john.vesel@philips.com). that reason we have set up a Monte Carlo (MC) simulation, calibrated it against analytical model and measurement results, assessed image quality, mainly on the Jaszczak phantom, and evaluated the impact of reconstruction methods. II. METHODS 1) Simulation environment GEANT4 is a software toolkit developed for simulating the passage of particles through material objects. It was developed at CERN and intended for detector design in high-energy physics, but due to its extension to low energy particle physics it is widely used in many nuclear physics related areas like nuclear medicine and bio-physics. The problem with SPECT calculations is the low collimator efficiency that requires a large number of primary events to be simulated. We used various optimization techniques to speed up simulation: Pre-allocated dynamically allocated memory, no parameterized volumes, energy cut for low-energy particles, restricted angles of incidence, and scatter and penetration background parameterization for large phantoms. 2) Simulation setup Simulation has been done for Tc-99m, kev radiation. A 150 mm diameter, 10 mm high water disk inside an acrylic disk holder was employed for system planar efficiency, and a line source of 150 mm length, slightly tilted with respect to one detector axis, was used to determine system planar spatial resolution. Phantom-collimator distance was 100 mm. A Jaszczak Phantom with cold rods of mm diameter inside an acrylic cylinder, half-filled with 740 MBq (20 mci) activity of Tc-99m in water, was implemented to visualize SPECT image quality. Projection images were taken under 128 different SPECT angles with 200 mm rotational radius at the collimator surface and 15 min total imaging time (Fig. 1). Two types of detectors, a NaI camera with VXGP collimator and CZT detectors with parallel-hole (PH) collimator, 5 mm CZT thickness, 2.46 mm pixel pitch, 0.35 mm septa thickness, and 40 mm or 45 mm septa length were used. For each detector we employed different energy resolution and different reconstruction methods to separate the influence of the essential parameters: spatial resolution, energy resolution, and reconstruction method, including different filters.

2 III. SIMULATION RESULTS A. NEMA system parameters Table 1 shows a comparison of MC simulated values with measurement results for Anger cameras with VXGP collimator, and with analytical values for a CZT detector with parallel-hole collimator of 45 mm septa length. TABLE 1 NEMA PARAMETER SIMULATION RESULTS Efficiency (cps / MBq) Spatial resolution (mm) NaI-VXGP, MC (measured) ( 80) ( 7.04) CZT-PH, MC (analytical) (74) (6.80) We see that Monte Carlo results compare well with the best values measured on real collimators and with analytical model calculations. The slight difference in numbers for the CZT detector is explained by penetration in the collimator septa which had not been included in the analytical model. Effects like limited absorption and fluorescence loss in the NaI or CZT detector, and an assumed electrical efficiency of 89% for the CZT detector, had been included in the model. B. Jaszczak Phantom Simulation of the cold rod section in the Jaszczak phantom was done to evaluate the impact of detector and reconstruction parameters on the image quality of a strongly scattering phantom. 1) Impact of spatial resolution Fig. 2 shows Jaszczak images obtained for a NaI-VXGP camera and two different CZT-PH detectors. Energy resolution E = 5% was assumed for all detectors. For reconstruction we used filtered back-projection (FBP) with Ram-Lak filter. We took the average of 40 transaxial slices to account for the high noise level in the images, caused by the relatively low dose that we used in the simulation. Contrast and detectability of small lesions, especially in the fifth and sixth sector, is increased for the CZT-PH detector with 45 mm septa length (center image) in comparison to a NaI-VXGP camera (left image) though the difference in spatial resolution is only 6.85 mm vs. 7.0 mm. The dependence may be amplified by depth-of-interaction effects in the thicker NaI layer that were not accounted for in the analytical model. When lower spatial resolution is simulated, 7.45 mm for a CZT-PH detector with 40 mm septa length (right image), image quality is worse than for the Anger camera. The impact of spatial resolution on contrast and detectability of small structures agrees well with previously published results [8]. Fig. 1: Geometry of Jaszczak Phantom and VXGP collimator Fig. 2: Jaszczak phantom images of a NaI-VXGP camera and two CZT-PH detectors with 45 and 40 mm septa height, respectively (from left to right). An energy resolution of 5% was assumed for all detectors. Reconstruction was done by filtered back-projection with Ram-Lak filter; 40 slices were averaged.

3 2) Impact of energy resolution Fig. 3 shows the influence of detector energy resolution on image quality. A theoretical improvement in energy resolution from E = 9.5% to E = 3% is accompanied by a slight increase in lesion detectability in the upper three sectors and of contrast on the whole phantom area (left and center image). There is also some difference seen when the two images are compared with the one simulated with E = 5% (Fig. 2, left image). We find, however, no statistically significant increase in the peak-to-valley ratio for the first sector. This may be due to the relatively low detector dose that we used in the simulation. The slightly better spatial resolution of the CZT detector, compared to a NaI camera, is expressed again in the better lesion detectability in the upper three sectors (right image). A comparison between E = 5% (Fig. 2, center image) and E = 3% (Fig. 3, right image) for the CZT detector, however, shows hardly any difference. 3) Impact of reconstruction method Reconstruction has a major influence on image quality and lesion detectability. We tried FBP with Ram-Lak, Shepp- Logan, Cosine, Hamming or Hann filter, and statistical methods, Maximum Likelihood Expectation Maximization algorithm (MLEM) and Ordered Subset Expectation Maximization (OSEM). Fig. 4 shows three examples. Fine details are preserved but noise is high for FBP with Ram-Lak filter while application of high frequency filters smoothes all structures but increases lesion detectability (left and center image). For the innermost rods in the first sector we found a 2-3 higher contrast-tonoise ratio (CNR) accompanied by only 20% reduction in the peak-to-valley ratio. Detectability of small lesions is better increased by application of OSEM (right image). Spatial resolution is high, but care has to be taken since the peak-tovalley ratio is reduced by 40%, compared to FBP with Ram- Lak filter. Fig. 3: Jaszczak phantom images of NaI-VXGP cameras with E = 9.5% and E = 3%, and a CZT-PH detector, 45 mm septa length, with E = 3% (from left to right). Reconstruction was done by filtered back-projection with Ram-Lak filter; average of 40 slices. Fig. 4: Jaszczak phantom images of a CZT-PH detector, 45 mm septa length, E = 5%, reconstruction by filtered back-projection with Ram-Lak filter, Hann filter, and by 5-pass OSEM (from left to right); average of 40 slices.

4 C. Image quality test phantoms We used planar hot disk phantoms [9] to investigate noise propagation in the image reconstruction and compared the resulting image quality with SNR values derived from theory. Planar disks with 40 mm diameter were simulated, imaged by a rotating slat detector of 2.46 mm pixel pitch at 256 different 7 angles, with a total number of decays per voxel. The image was reconstructed from the resulting line integrals by FBP, MLEM, OSEM, and Algebraic Reconstruction Technique (ART). This reconstruction of planar images from line integrals is equivalent to the reconstruction of transaxial slices in SPECT, except for attenuation in three-dimensional phantoms [7]. 1) Impact of reconstruction on image noise Fig. 5 shows the impact of different filters, used in FBP, on noise in the reconstructed image. The Ram-Lak filter is known to preserve fine details as well as noise, while high frequency filters like Cosine, Hamming and Hann filters are seen to smooth shot noise in the image considerably. A similar noise reduction is seen when reconstruction is done by 16-pass MLEM (Fig. 6). The same result is achieved much faster by use of 1-pass OSEM with 16 subsets. Repeated application of OSEM, however, results in an increasingly noisy image. ART, a non-statistical iterative method, causes even higher noise. TABLE 2 SNR VALUES OF HOT DISK PHANTOM SNR lin SNR near SNR th ξ A(ξ ) Q Ram-Lak ξ 0.82 Shepp- Logan Cosine πξ ξ sin c 0.5 ξ N πξ ξ cos 0.16 ξ N 2πξ Hamming ξ cos ξ N 2πξ Hann ξ cos ξ N OSEM OSEM OSEM ART 9.4 Fig. 5: Hot disk phantoms, reconstruction by FBP with Ram-Lak, Cosine, Hamming and Hann filter (from upper left to lower right). Fig. 6: Reconstruction by 16-pass MLEM, 2-pass and and 8-pass OSEM with 16 subsets, and ART (from upper left to lower right). Table 2 shows SNR values determined from a 52-pixel region of interest in the center of the hot disk, calculated for different reconstruction methods. SNR lin refers to Fig. 5 and 6, using linear interpolation in FBP, while SNR near uses nearest neighbor values and SNR th gives theory values [10]. The FBP filter function in Fourier space is defined by the variable ξ, normalized by Nyquist frequency ξ N, and commonly known filter functions A (ξ ). We define the noise reduction factor Q: Q = π ξ A( ξ) dξ. (1) The squared signal-to-noise ratio is given by p SNR 2 = AvTE (2) 2rQ where Av is the activity per voxel, T the imaging time, E the detector efficiency and 2 r / p the disk diameter divided by the pixel pitch [7]. We see from Table 2 that this noise theory describes FBP with nearest neighbor interpolation very well, and for statistical reconstruction we get comparable SNR.

5 2) Impact of reconstruction on spatial resolution From the different spacing of grid lines in Fig. 5 and 6 we see that spatial resolution is reduced by filtering but higher for statistical reconstruction. We used line phantom simulations for further evaluation. Spatial resolution, calculated as the FWHM of the line spread function (LSF), is calculated from all lines in the image perpendicular to the line source (Fig. 7). Fig. 7: Line phantom image, reconstructed by FBP with Ram-Lak filter, and line spread function calculated from the image data. Fig. 8 shows the correlation of SNR values from Table 2 with the FWHM of LSF. Comparing different filters in FBP, the higher SNR achieved by filtering is accompanied by slightly lower spatial resolution. Repeated application of OSEM gives lower SNR values, shown for 1-pass up to 5-pass iteration, while spatial resolution is increased, even beyond that of a parallel-hole detector applied without reconstruction. S / N ratio OSEM parallel hole 5x 3x 2x FBP Shepp-Logan Ram-Lak 1x Cosine Hann Hamming 0 6,5 7,0 7,5 8,0 8,5 9,0 9,5 spatial resolution / mm Fig. 8: Signal-to-noise ratio, measured on disk phantoms, vs. spatial resolution, measured on line phantoms, for different reconstruction methods. The value for a parallel-hole detector is given for comparison We have seen, however, that convergence properties of the OSEM reconstruction depend strongly on the object simulated so that we cannot generalize the correlation shown in Fig. 8. IV. CONCLUSIONS We have set up a Monte Carlo simulator for the evaluation of new SPECT detector concepts. As a first test, NEMA system parameters were simulated and shown to be well comparable to analytical results and best values measured on SPECT systems. Different test phantoms for signal-to-noise ratio and for spatial resolution, and the Jaszczak phantom were simulated to evaluate the impact of detection concept, energy resolution, and reconstruction method on image quality. Contrast and the detectability of small lesions are enhanced by a small improvement in spatial resolution. There seems to be a further advantage of pixellated systems in comparison to Anger cameras that is not yet explained. An improvement in energy resolution from 9.5% to a theoretical value of 3% for Anger cameras causes an increase in lesion detectability, but for pixellated detectors we hardly see any difference for energy resolution values of 3% and 5%. Results are, however, statistically not significant. High frequency filters in the back-projection as well as the application of statistical reconstruction increase the signal-tonoise ratio in reconstructed images, which is well understood. Contrast and lesion detectability are enhanced, but the peak-tovalley ratio in the Jaszczak phantom images is reduced and spatial resolution is diminished. A further evaluation will use forced detection in the Monte Carlo code to enable simulation of a much higher phantom or patient dose. This is necessary to get statistically significant values of image quality parameters and evaluate the in depth impact of reconstruction methods on lesion detectability. V. REFERENCES [1] J.F. Butler, C.L. Lingren, S.J. Friesenhahn, W.L. Ashburn, R.L. Conwell, F.L. Augustine, et al, CdZnTe solid-state gamma camera, IEEE Trans. Nucl. Sci., vol. 45, pp , June [2] R. Amrami, G. Shani, Y. Hefetz, I. Blevis, A. Pansky, A Comparison Between The Performance Of a Pixellated CdZnTe Based Gamma Camera and Anger NaI(Tl) Scintillator Gamma Camera, Proceedings of the 22 nd Annual EMBS International Conference, July 23-28, 2000, Chicago, IL pp [3] Y. Eisen, I. Mardor, A. Shor, Z. Baum, D. Bar, G. Feldman, et al., NUCAM3 A Gamma Camera Based on Segmented Monolithic CdZnTe Detectors, IEEE TNS, vol. 49, No 4, pp , [4] D.J. Wagenaar, S. Chowdhury, J.C. Engdahl, D.D. Burckhardt, Planar image quality comparison between a CdZnTe prototype and a standard NaI(Tl) gamma camera, Nuclear Instruments and Methods in Physics Research, vol. A 505, pp , [5] B. Mueller, M. O Connor, I. Blevis, D. Rhodes, R. Smith, D. Collins, et al, Evaluation of a small CZT detector for scintimammography, JNM, Vol 44, No 4, pp , April 2003 [6] M. Petrillo, J. Ye, J. Vesel, L. Shao, H. Wieczorek, and A. Goedicke, "Imaging performance of tiled solid-state detectors", presented at IEEE MIC 2004, Rome, Oct [7] H. Wieczorek, A. Goedicke, L. Shao, M. Petrillo, J. Ye, and J. Vesel, "Analytical model for pixellated SPECT detector concepts", presented at IEEE MIC 2004, Rome, Oct [8] G. Muehllehner, "Effect of resolution improvement on required count density in ECT imaging: a computer simulation", Phys. Med. Biol., vol. 30, pp , 1985 [9] T. F. Budinger, S. E. Derenzo, G. T. Gullberg, W. L. Greenberg, and R. H. Huesmann, "Emission computer assisted tomography with singlephoton and positron annihilation photon emitters",, J. Comput. Assist. Tomogr., vol. 1, pp , 1977 [10] M. A. Lodge, D. M. Binnie, M. A. Flower, and S. Webb, "Experimental evaluation of a prototype rotating slat collimator for planar gamma camera imaging", Phys. Med. Biol., vol. 40, pp , 1995.