Interpreting results from a comparative study of lesion detectability for 6 different PET systems.

The article by Kadrmas and Christian (1), which appears in this issue of The Journal of Nuclear Medicine, is a well-written account of a comprehensive study that compares lesion detectability of 6 different commercial PET imaging systems, including 2 state-ofthe-art high-resolution dedicated bismuth germanate (BGO) systems, an older dedicated BGO system, a dedicated NaI(Tl) system, and 3 NaI(Tl) hybrid systems. The important contributions of the study include a realistic whole-body phantom with simulated lesions of different sizes and contrasts, a well-designed experimental protocol, and the use of a human observer performance study and the localization receiver operating characteristic (LROC) study paradigm in the evaluation of lesion detectability. The unique whole-body phantom is comprised of a 3-dimensional (3D) Hoffman brain phantom (Data Spectrum Corp., Hillsborough, NC), an anthropomorphic thorax phantom (Radiology Support Devices Inc., Long Beach, CA), and an elliptic cylinder pelvis phantom (Data Spectrum). The organ compartments are filled with relative activity concentrations to simulate 18F-FDG distribution in patients. Twenty-seven spheric lesions filled with 22Na activity and with inner diameters of 7, 8, 12, and 16 mm are placed at various locations throughout the whole-body phantom. The lesion-tobackground activity concentration ratios are 4, 6, 10, and 16 to simulate different lesion contrasts. The elaborate and realistic whole-body phantom provides a realistic simulation of a normal-size patient found in tumor PET studies. Because the PET systems included in this study came from several categories of PET systems with very different costs and performance characteristics, comparison among them is a complex issue. To minimize experimental variations over time, data from all PET systems were acquired over a period of several weeks. The same 2-dimensional (2D) data acquisition mode was used in all PET systems except for the C-PET system, where the 3D mode was used. To conform to clinical practice, manufacturer-supplied image reconstruction software and manufacturer-suggested default processing parameters in data processing were used. Also, similar data acquisition times were adopted for the same patient throughput and for patient comfort. However, as shown in Table 2 (1), the different system sensitivities result in very different total counts in the acquired images. The lesion detectability performance of the PET systems was evaluated using a localization ROC, or LROC, study design with 11 human observers. Different from the simpler ROC study design, the LROC study paradigm offers a closer resemblance to the lesion detection task in clinical studies. The observer data were analyzed using the LROCFIT program by Swensson (2) to obtain LROC curves and the probabilities of correct lesion localization in images obtained from the different PET systems by the human observers. The LROC results were analyzed using several statistical tests. The results of the study were summarized in Figures 4 and 5 (1). They clearly indicate the general superior performance of the dedicated PET systems compared with the NaI(Tl)-based hybrid PET systems in terms of lesion detectability. Among the dedicated PET systems, the performance of the state-of-art BGO-based systems is superior to that of the older BGO-based system, which in turn is superior to that of the NaI(Tl)-based system. A point of interest shown in Figure 6 (1) is that for the largest (16 mm) lesions, lesion detectabilities among the different PET systems are almost identical. This suggests that for larger lesions, where statistical noise fluctuations are less important, the similar spatial resolution characteristics of the different PET systems rendered similar lesion detectability. Another point of interest is shown in Figure 7 (1), where the same iterative ordered-subsets expectation maximization (OSEM) reconstruction algorithm with different iterative numbers, when applied to data obtained from very similar BGO-based dedicated PET systems, can provide reconstructed images with markedly different quality. Specifically, data from the HR PET system (CTI PET Systems, Knoxville, TN), when processed with 7 iterations of the OSEM algorithm, yielded image quality that was superior to that obtained using 2 iterations and was comparable with that from the Advance PET system (General Electric Medical Systems, Milwaukee, WI). An important note, as indicated by Kadrmas and Christian (1), is the exercise of caution in the interpretation of the results of this study. The lesion detectability performance of a PET system depends on a variety of factors, including the performance characteristics of the system—for example, system sensitivity, spatial resolution and counting rate capability, 2D versus 3D Received Jul. 25, 2002; accepted Jul. 25, 2002. For correspondence or reprints contact: Benjamin M.W. Tsui, PhD, Department of Radiology, Johns Hopkins University, JHOC 4263B, 601 N. Caroline St., Baltimore, MD 21287-0859. E-mail: [email protected]