Voxel-based analysis of 201Tl SPECT for grading and diagnostic accuracy of gliomas: comparison with ROI analysis

Purpose The aim of this retrospective study was to assess the utility of a voxel-based analysis (VBA) method for 201Tl SPECT in glioma, compared to conventional ROI analysis. Methods We recruited 24 patients with glioma (high-grade 15; low-grade 9), for whom pre-operative 201Tl SPECT and MRI were performed. SPECT images were coregistered with MRI. The uptake ratio (UR) images of tumor to contralateral normal tissue were measured on early and delayed images, and the 201Tl retention index (RI) map was calculated from the early and delayed uptake ratio maps. In the ROI analysis, tumors were traced on a UR map, and the mean and maximal uptake ratio values on the early images were, respectively, defined as the mean and maximal UR. The mean and maximal RI values (mean and maximal RI) were calculated by division of the mean and maximal UR, respectively, on the delayed image by the mean and maximal UR on the early image. For the RI map calculated voxel by voxel, the maximal RI value was defined as VBA-RI. We evaluated sensitivity and accuracy of differential analysis with the mean and maximal UR, RI, and VBA-RI. Results The high- and low-grade groups showed no significant difference in mean and maximal RI (0.98 ± 0.12 vs. 1.05 ± 0.09 and 0.98 ± 0.18 vs. 1.05 ± 0.14, respectively). The AUC and accuracy of the mean and maximal RI were 0.681 and 66.7 %, and 0.622 and 62.5 %, respectively. In contrast, VBA-RI was higher in high-grade than in low-grade glioma (1.69 ± 0.27 vs. 0.68 ± 0.66, p < 0.001). The AUC and accuracy of VBA-RI were 0.963 and 95.8 %, which are higher than those obtained for mean (p < 0.05) and maximal RI (p < 0.01). There was no significant difference in ROC between the VBA-RI and the mean UR (0.911, p = 0.456) and maximal UR (0.933, p = 0.639); however, the AUC, sensitivity, and diagnostic accuracy of VBA-RI were all higher than those of the mean and maximal UR. Conclusion The voxel-based analysis method of 201Tl SPECT may improve diagnostic performance for gliomas, compared with ROI analysis.


Introduction
Thallium-201 chloride single photon emission computed tomography ( 201 Tl SPECT) is widely available for brain tumors. This technique has been used for the grading of tumor malignancy [1][2][3][4] and differentiation of tumor recurrence from radiation necrosis [5][6][7]. 201 Tl behaves like potassium, and relies on the sodium-potassium adenosine triphosphatase membrane transport system, as well as regional brain blood flow, reflecting the permeability of the blood-brain barrier, tissue viability, and cellular uptake [8][9][10]. For differential diagnosis and malignancy grading, however, although the important findings of 201 Tl images have conventionally been considered to be high uptake on early phase 201 Tl images and high retention indices (RI) for 201 Tl uptake in the delayed phase relative to that in the early phase, the results have been controversial. Some authors have found that there is no statistically significant difference in 201 Tl retention for histological malignancies [11,12]. In contrast, other authors have shown the 201 Tl RI to be correlated with the histological grading of brain tumors [3,4,6,13]. These studies recruited tumors of differing histology, including gliomas as well as metastatic tumors, meningiomas, etc. Especially, meningiomas should be evaluated separately from gliomas, as extra-axial tumors may display different 201 Tl kinetics from intra-axial tumors, due to their unique hemodynamics which proceeds without a blood-brain barrier [12].
For the semiquantitative analysis of 201 Tl images with brain tumors, the region of interest (ROI) analysis has been widely discussed in previous reports; however, this method presents some problems. Firstly, ROI definition on tomographic images has poor reproducibility, especially for small lesions, due to the low resolution of 201 Tl images. Secondly, in heterogeneous tumors (e.g., containing necrotic and cystic components) an ROI analysis may underestimate the activity of the tumor tissue, and misinterpret its grading [14,15]. Finally, ROI analysis is not suitable for coregistration of two phase images (early and delayed images) for calculation of the retention indices, and cannot assess the 201 Tl distribution in the whole brain.
On the other hand, a voxel-based analysis (VBA), which is essentially voxel-by-voxel comparison, may resolve these problems. In functional images (PET, SPECT), or structural images (3-dimensional MRI), this method is usually used for statistical analysis in the whole brain, without a priori assumptions [16][17][18][19]. This method has automated coregistration, and supports comparison with different images in the whole brain, on a voxel-by-voxel basis. Furthermore, even if the tumor uptake is heterogeneous, the voxel-based analysis method reveals the activities of the tumor tissue on a per-voxel basis, and evaluates its malignancy, selectively. The purpose of this study was to estimate the 201 Tl retention index using a voxel-based analysis method for the grading and diagnostic accuracy of supratentorial gliomas, and to compare the results with those of the ROI analysis method.

Patients
In this study, from October 2005 to March 2009, 24 consecutive patients (16 males and 8 females) with pre-operative supratentorial gliomas received 201 Tl SPECT imaging and MR imaging at our institution (Table 1). All patients were given informed consent, and we obtained institutional approval for the retrospective analysis. Both techniques were performed within 1 week. Histological diagnosis was obtained by open surgery (n = 20) or biopsy (n = 4). Twenty-four tumors consisted of 10 glioblastomas, 5 anaplastic gliomas (2 anaplastic astrocytomas, 2 anaplastic oligoastrocytomas, and 1 anaplastic ependymoma), 2 diffuse astrocytomas, 2 oligodendrogliomas, 1 oligoastrocytoma, 1 pilocytic astrocytoma, and 3 astrocytomas which were consistent with Grade II astrocytoma; however, their specific histologies were unknown. To facilitate comparison, the various histological glioma entities were assigned to two groups: a high-grade group (n = 15) consisted of glioblastoma (Grade IV) and anaplastic glioma (Grade III), and a low-grade group (n = 9) consisted of diffuse astrocytoma (Grade II), oligoastrocytoma (Grade II), oligodendroglioma (Grade II), astrocytoma (Grade II), and pilocytic astrocytoma (Grade I).
Image acquisition 201 Tl SPECT SPECT was performed using a triple-headed camera (PRISM 3000; Philips) equipped with a low-energy, fan beam collimator. Imaging was started from 10 min (early imaging) and from 4 h (delayed imaging) after intravenous injection of 111 MBq of 201 Tl-chloride. We performed SPECT data collection (60 s/step, 60 steps) for 24 min. The matrix size was 128 9 128, and the collection window was 74 keV centered with 20 % width. Raw data were reconstructed with 2.2 9 2.2 9 6.7 mm 3 in voxel size after decay correction using OS-EM (four iterations).
Secondly, we selected a slice corresponding to the location of the brain tumor on MRI (Fig. 1a) [1]. Next, a normal brain ROI was drawn on the contralateral normal brain tissue in early and delayed images, with the same voxel size as each tumor ROI (Fig. 1b). Consequently, the coregistered early and delayed SPECT images were divided by the mean value of the contralateral normal tissue ROI, on a per-voxel basis in the whole brain. The obtained images were the early and delayed ''uptake ratio maps'' (Fig. 1c).
Finally, for the retention index map, the delayed uptake ratio map was divided by the early uptake ratio map, on a per-voxel basis (Fig. 1d). To remove the background uptake and/or noise, uptake values lower than 1.5 among whole brain voxels on the early and/or delayed uptake ratio maps were then replaced with zero ( Fig. 1e). This cutoff value was determined by visual assessment in all cases (mean ± SD 1.50 ± 0.06; range 1.4-1.7), so as to keep the tumor uptake as well as remove the background uptake and/or noise (Fig. 2).

ROI analysis of 201 Tl SPECT
ROI analysis was performed according to previous reports [1]. The coregistered early and delayed 201 Tl images were used (Fig. 1a). The ROI for the brain tumor was manually placed over the areas of abnormal 201 Tl uptake determined according to the MRI findings. As mentioned above, a normal brain ROI was drawn on the contralateral brain tissue with the same voxel size as for the corresponding tumor ROI.
The ratio of the mean tumor count to normal brain count, and that of the peak tumor count to normal count were calculated as the mean and maximal uptake ratios (mean and maximal URs), respectively. The ratios of the mean and maximal URs on the delayed image to those on the early image were calculated as the mean and maximal retention indexes (mean and maximal RIs). Briefly, the four parameters in the ROI analysis were defined, as follows: Note that the voxel with the peak count in the tumor ROI on early image does not necessarily correspond to that on delayed image.

Voxel-based analysis of 201 Tl SPECT
Voxel-based analysis was performed using previously obtained ''the retention index map''. The retention index map was created as follows: firstly, early and delayed 201 Tl SPECT images were registered on their corresponding MR images. Thus, all the voxels between early and delayed 201 Tl SPECT images had the same anatomical locations in the whole brain. Secondly, ''uptake ratio maps'' were calculated by normalizing to mean uptake in the contralateral ROI. Thirdly, ''retention index map before cutoff'' was created from two uptake ratio maps (delayed uptake ratio map divide by early uptake ratio map on a per-voxel basis). Finally, ''retention index map'' was completed by the replacement of voxel value under 1.5 with zero for the elimination of the background uptake and/or noise.
The resulting parameter, the ''VBA-RI'' was defined as the peak value for a tumor in the ''retention index map'' obtained after the image processing. Subsequently, we confirmed that the majority of VBA-RI locations corresponded to tumors, with the exception of 5 cases. In 3 of these cases, tumor accumulations were not clear (Pt 20-24), while in the other cases, tumors existed near the scalp (Pt 12, 18). For this reason, these 5 cases required interpretation as a brain tumor using coregistered MRI in a complementary manner. All of the UR and RI parameters were presented as mean ± standard deviation and statistically compared to the high-and low-grade groups with the Mann-Whitney U test. Receiver operating characteristic curves were employed to differentiate the high and low-grade glioma groups, and identify the cutoff values of the five indices. A p value \0.05 was considered significant.

ROI analysis
The mean and maximal UR values were significantly higher in high-grade group, relative to the low-grade group (mean UR: 2.36 ± 0.76 vs. 1.25 ± 0.38, p \ 0.001; maximal UR: 3.92 ± 1.42 vs. 1.91 ± 0.61, p \ 0.001) (Fig. 3). In the ROC analysis, mean and maximal URs had cutoff values of 1.42 and 2.27, with AUC values of 0.911 and 0.933, respectively (Fig. 4a). The sensitivity and accuracy were 86.7 and 83.3 % for the mean UR, and 93.3 and 87.5 % for the maximal UR.

Voxel-based analysis
The VBA-RI was significantly higher in the high-grade group, relative to the low-grade group (1.69 ± 0.27 vs. 0.68 ± 0.66; p \ 0.001) (Fig. 3). In the ROC analysis, the VBA-RI had a cutoff value of 1.25 and an AUC of 0.963 (Fig. 4a, b), with a sensitivity and accuracy of 100 and 95.8 %, respectively.
Both the AUC and accuracy of VBA-RI were higher than those obtained for mean RI (p \ 0.05) and maximal RI (p \ 0.01). There was no significant difference in the AUC between VBA-RI and mean UR (p = 0.456), or between VBA-RI and maximal UR (p = 0.639). However, VBA-RI displayed higher AUC, sensitivity, and accuracy Discussion 201 Tl uptake and its retention index (which reflects 201 Tl affinity) have often been evaluated for the diagnosis of brain tumors using the ROI method. However, the diagnostic criteria of these parameters for the diagnosis of brain tumors have not been established adequately using the ROI method. Some reports have mainly assessed the histological grading of glioma using only the 201 Tl uptake ratio in early imaging using ROI analysis [9,[21][22][23][24][25]. Our results also demonstrated that the maximal and mean UR, using ROI analysis, show high accuracy (mean UR 83.3 %; maximal UR 87.5 %). The diagnostic accuracy, sensitivity, and AUC obtained with VBA-RI were superior to those obtained using the other parameters, although there were no significant differences among the accuracies of VBA-RI (95.8 %) and the mean UR and maximal UR. Therefore, the use of VBA-RI yielded improved diagnostic performance for gliomas.
Furthermore, a voxel-based analysis presents several additional advantages, in that does not require the definition of a tumor margin, and reveals voxel-by-voxel activity for inhomogeneous tumors. In this study, we attempted to calculate the retention index on a voxel-by-voxel basis Fig. 3 Box plot of each of the indices in the high-grade group vs. the low-grade group. The mean and maximal UR shows a significant difference between the high-and low-grade groups. Although the VBA-RI also differed for the two groups, the mean and maximal RIs did not a Receiver operating characteristic curves (ROC) for VBA-RI and mean and maximal URs. b ROCs for VBA-RI and mean and maximal RIs. The AUC for the mean and maximal URs, and VBA-RI were 0.911, 0.933, and 0.970, respectively, which were higher than those of mean and maximal RIs (0.681 and 0.622). The AUC and accuracy of VBA-RI were significantly higher than those of the mean RI (p \ 0.05) and maximal RI (p \ 0.01). The AUC of VBA-RI did not significantly differ from those of the mean and maximal URs. However, the AUC, sensitivity, specificity, and diagnostic accuracy were all higher for VBA-RI, relative to both the mean UR and maximal UR using anatomical coregistration and the voxel-based analysis techniques which a statistical image analysis employs. We assumed that the maximal retention index reflects the 201 Tl affinity of a brain tumor, which is related to tumor malignancy, even when the tumor has inhomogeneous components.
Interestingly, in our study, the voxel of maximal uptake in a brain tumor on an early image did not correspond to that on the delayed image in all cases. Thus, the RI obtained using the ROI method does not reflect the exact change in 201 Tl uptake, in particular, for a tumor with an inhomogeneous distribution. Prieto et al. [26] reported the successful utilization of the voxel-based analysis of dualtime-point 18 F-FDG PET images for brain tumors. They also considered that the location of the voxel of maximal uptake in an early image did not match the voxel of maximal uptake in the corresponding delayed image. A similar result was also obtained in our study using 201 Tl SPECT. For maximal UR, we pointed out that the variation was computed between two different voxels in the tumor, which does not represent the real change in the voxel of maximal uptake in the early image. Therefore, the voxelbased retention index would be more meaningful for the characteristic 201 Tl behavior in a brain tumor. In our results, the maximal RI underestimates the differences in maximal uptake between the early and delayed images, compared with the change estimated by RI using VBA analysis, and the diagnostic accuracy of maximal RI was inferior to that of VBA-RI (62.5 vs. 95.8 %). VBA-RI was significantly higher in the high-grade group than in the low-grade group, and displayed a very high accuracy for classification of histological grading, while the mean and maximal RI with an ROI analysis showed no significant differences between the high-and low-grade groups.
There are some limitations in this study. Firstly, we ignored the voxels with a lower uptake ratio in the early and delayed images for the elimination of background uptake and/or noise. It might also exclude a portion of the brain tumor. While the assessment of uptake ratios without a cutoff process was reported using 18 F-FDG brain tumor PET [26], it was necessary to evaluate 201 Tl SPECT via processing, because 201 Tl SPECT displays more nonspecific accumulation than 18 F-FDG PET. Secondly, there may be a minor error in the coregistration of 201 Tl SPECT to non 3-dimensional MRI using SPM. In this study, we found no errors in any of the acquired fusion images of 3-dimensional 201 Tl SPECT and MRI. Finally, all of the recruited cases had brain tumors in our study. Therefore, the present study included no true negative case, which could possibly have an effect on the diagnostic accuracy.

Conclusion
We have evaluated the affinity of 201 Tl for brain tumors by its retention and uptake using a voxel-based and ROI-based analysis. The voxel-based retention index demonstrated a significant improvement in the differential analysis of tumor malignancy, compared to ROI-based uptake and retention indices. Although this preliminary study should encourage the use of 201 Tl SPECT for brain tumors, further study is required with a larger prospective series of patients. Fig. 6 Case 2 is a 13-year-old male patient who presented with a low-grade (grade I) in left occipital pilocytic astrocytoma. a Gadolinium-DTPA enhanced T1 weighted image; b mean UR image; c maximal UR image; d VBA-RI image. Since the tumor existed near the scalp, the MRI was used for reference. The mean and maximal URs of the tumor demonstrated high values of 1.95 and 2.55, respectively, which were suggestive of high-grade glioma. The VBA-RI of the tumor, however, showed a low value of 0.98, which was suggestive of, and in agreement with low-grade glioma Open Access This article is distributed under the terms of the Creative Commons Attribution License which permits any use, distribution, and reproduction in any medium, provided the original author(s) and the source are credited.