[¹⁸F]Fluciclovine PET discrimination between high- and low-grade gliomas

Patients with biopsy proven glial neoplasms were recruited from February 19, 2008, to April 21, 2015, under a National Institutes of Health grant (R01CA121320). The recruitment protocol was approved by the Institutional Review Board (IRB) and complied with the Health Insurance Portability and Accountability Act (HIPPA). The radiotracer was administered under FDA Investigational New Drug (IND) 72,437 and was synthesized either via automated synthesis [14] or the FastLab Cassette System (GE Healthcare). Safety monitoring during the drug infusion was performed, and no adverse events were recorded. In total, 51 patients were recruited with 37 eventually receiving at least one fluciclovine PET study post-histologic confirmation as required by the IRB. For this analysis, patients were sub-selected under the inclusion criteria of biopsy proven disease either after stereotactic brain biopsy or excisional biopsy/partial resection. Patients with systemic chemotherapy or radiation before the fluciclovine PET in the current episode of care were excluded, resulting in a total of 16 patients included in this study (Table 1). We excluded that group of patients from this analysis as we felt that the previously treated high-grade glioma patients were too heterogenous in their prior therapy which may introduce too many confounding errors into the image analysis. Additionally, no patients were imaged with recurrent previously treated low-grade glioma (LGG). Two of the included 16 patients had two discrete lesions each that underwent tissue sampling and PET lesion analysis resulting in a total of 18 lesions. In 1 patient, a bis-chloroethylnitrosourea (BCNU) wafer had been employed during excisional biopsy/partial resection as per existing standard of care.

PET imaging of the head was performed using a high resolution research tomograph (HRRT) scanner (Siemens Medical Solutions, Knoxville, TN, USA). Patients were administered 366–399 MBq (9.9–10.8 mCi) fluciclovine (¹⁸F) intravenously via pump over 4 min. Continuous dynamic PET imaging of a single bed position was begun at the beginning of injection and ended at the 65 min time point. Attenuation scanning was performed in single mode with a 30-mCi ¹³⁷Cs point source. The data was collected in list mode and later histogrammed into 15 time points (6 × 30, 4 × 180, and 5 × 600 s). PET emission data were corrected for attenuation, random, and scatter and reconstructed with an ordinary Poisson ordered–subset expectation maximization (OP-OSEM) algorithm (6 iteration, 16 subsets) to a 256 × 256 × 207 image matrix (voxel size 1.2 mm³) [15]. Image data was post-reconstruction smoothed with a 24 mm FWHM Gaussian filter to reduce noise and improve contrast in the brainstem. The final isotropic resolution was 4.6 mm and matched that reported in our previous work. Data was transferred to a MIM workstation (MIM Software, OH) for further analysis.

A board-certified radiologist using the Absolute Threshold Contouring Tool (MIM Software, OH, USA) drew regions of interest (ROIs) over the tumors and background ROIs (i.e., contralateral brain and venous confluence) for all time points.

Fluciclovine PET images were co-registered to T1 post contrast MRI and FLAIR sequences. Tumor regions of interest (ROIs) were defined by creating a spherical PET ROI to include the volume of tissue demonstrating hyperintense FLAIR signal corresponding to the biopsy proven glioma. Within this PET ROI, the voxels with peak activity were used to derive a tumor maximum standardized uptake value (SUV_max). A 15-mm spherical ROI was placed over the contralateral normal brain, including both gray and white matter, to provide a normal mean standardized uptake value (SUV_mean). The SUV_mean of the contralateral normal brain was utilized to select a threshold for defining metabolically active tumor within the aforementioned PET ROI. Minimum thresholds of 1.3*, 1.6*, and 1.9* of the contralateral normal parenchymal SUV_mean were used to define the tumor SUV_mean (Figs. 1 and 2). These thresholds were selected according to similar work done with FET and other amino acid PET tracers [16, 17]. Careful consideration when drawing ROIs over the tumor was used to exclude blood pool or adjacent choroid plexus which could falsely contribute to metabolic tumor volume.

Time activity curves for each lesion, normal contralateral parenchyma, and venous confluence were obtained by averaging PET metrics such as SUV_max or tumor: background (TB_mean) over all lesions at each time point using Excel (Microsoft, WA, USA). Differences between HGGs, LGGs, blood pool, and normal brain parenchymal SUVs, along with assessment of apparent TAC equilibrium kinetics, were detected using visual inspection with statistical comparison.

SUV_max and SUV_mean for each tumor lesion, contralateral normal background, and venous confluence were recorded at all time points. Tumor to background ratios for each lesion were calculated as TB_max = (SUV_max tumor)/(SUV_mean background) and TB_mean = (SUV_mean tumor)/(SUV_mean background) at all imaged time points and at all thresholds relative to normal contralateral brain resulting in TB_mean1.3, TB_mean1.6, and TB_mean1.9.

The optimal threshold for differentiating HGGs from LGGs utilizing tumor SUV_max was calculated using a receiver operator characteristic curve (ROC) for each tumor SUV_max measurements from 30 to 50 min post-injection, when there was an apparent plateau of the SUV_max. A single patient with anaplastic astrocytoma was included within the high-grade group for the purposes of this analysis. Sensitivity and specificity for identifying HGGs is reported based on the optimal threshold. A similar approach using ROC curves was applied to each TB_mean dataset at least 30 min post-injection to obtain a TB_mean threshold to distinguish HGG from LGG.

Five micron sections of formalin-fixed, deparaffinized tissue from 13 of the 18 lesions were immunostained using antibodies against Ki-67 (clone MIB1, 1:160 dilution, DAKO Corp, Carpinteria, CA, USA) using normal brain as the control. Five lesions did not undergo proliferative analysis. Negative controls were run simultaneously and had primary antibody replaced with buffer. Antigen retrieval was conducted in citrate buffer at pH 6.0 under a pressure of 15 lbs/in^{^2} for 3 min. Envision+ Dual Link Kit (DAKO Corp) was used as the detection system, with diaminobenzidine as the chromogen and hematoxylin as the counterstain. Staining was performed with the DAKO Autostainer.

Statistical analysis was performed with MATLAB R2017a prerelease (Mathworks, MA) using the Statistics and Machine Learning Toolbox. Four semiquantitative [¹⁸F]-fluciclovine PET metrics were assessed as potential discriminators of HGG vs LGG: SUV_max, SUV_mean, TB_max, and TB_mean. Each metric was analyzed at every time point demonstrating apparent equilibrium kinetics (i.e., 30, 40, 50, and 60 min post-injection). Relative importance of each predictor was made using a lasso regularization of a logistic regression model to identify relevant predictor variables (see Additional file 1). Unbalanced 2-way ANOVA followed by post hoc Tukey’s test were also used to compare lesion SUV_max, normal brain parenchyma SUV_mean, and venous confluence SUV_mean at equilibrium time points.

Unbalanced statistical design was used since only 6/10 HGG patients and 5/6 LGG patients completed the full 65-min acquisition protocol. The remaining studies were prematurely halted due to patient request. This patient attrition skewed the graphs at 60 min, so although the full data set was analyzed, only data up to 50 min is displayed in Figs. 3 and 4.

Statistical analysis on single time point data using scatter plots of SUV_max and TB_mean was performed using unequal variance two-tailed t tests of World Health Organization (WHO) II and WHO IV tumor uptake at equilibrium time points. Only one lesion was consistent with WHO III classification and was not included in this statistical analysis. However, this lesion was included in the high-grade tumor class for purposes of ROC analysis and with the high-grade group for time activity curve analysis as previously discussed.

For expression of percent positive Ki-67 cells in the malignant lesions correlated with lesion SUV_max and TB_mean at 30 min post-injection, the linear correlation coefficient (R value) for strength of correlation and p value to test the null hypothesis of a zero slope was performed.

Sixteen patients were included per inclusion criteria (Table 1). All 16 patients (8 M and 8 F with mean age of 49.6 years) completed the fluciclovine PET scan after either having undergone a stereotactic biopsy (8 patients) or excisional biopsy/partial resection (8 patients) prior to the fluciclovine PET scan. PET imaging was performed an average of 19.4 days (3–61 days) after surgical intervention. The confirmed histopathology for each patient were distributed as follows: 4 oligodendrogliomas (WHO grade II), 2 diffuse astrocytomas (WHO grade II), 1 anaplastic astrocytoma (WHO grade III), and 11 glioblastomas (WHO grade IV) with two patients having multifocal disease involving 2 distinct lesions. One of the multifocal glioblastoma patients previously had a known oligodendroglioma which had been treated with external radiation 2 years prior to the fluciclovine PET study. In total, 18 tumors were analyzed from 16 patients.

Logistic regression found that SUV_max and TB_mean were good discriminators of HGG versus LGG, whereas TB_max and SUV_mean were relatively irrelevant predictors of tumor grade (see Additional file 1). Using ROC analysis on all HGG and LGG SUV_max levels obtained at least 30 min after injection demonstrated an optimal discrimination threshold of SUV_max > 4.32. This SUV_max threshold provides a sensitivity of 90.9% and specificity of 97.5% (AUC 0.985; Table 2).

TB_mean of HGGs and LGGs was evaluated utilizing minimum threshold levels of 1.3*, 1.6*, and 1.9* contralateral normal brain parenchyma SUV_mean, resulting in TB_mean1.3, TB_mean1.6, and TB_mean1.9, respectively (Fig. 4). Each TB_mean metabolic threshold level resulted in a unique discriminator of HGG and LGG as expected. For example, setting a TB_mean1.3 threshold to 2.15 resulted in a sensitivity of 97.5% and specificity of 95.5% (AUC 0.989) with a statistically significant difference in uptake between WHO II and WHO IV lesions (p = 0.0036 at 30 min, p = 0.00041 at 40 min, and p = 0.0088 at 50 min). Similar results for TB_mean1.6 show an optimal threshold of 2.56 with sensitivity of 95.0% and specificity of 100% (AUC 0.989). Finally, ROC analysis shows an optimal threshold for TB_mean1.9 of 2.58 with a sensitivity of 95.0% and specificity of 100% (AUC = 0.991). This analysis found that all metrics based on SUV_max and TB_mean have high discriminatory power, but SUV_max and TB_mean1.3 had the highest sensitivity (lowest false negative) for detection of HGG, with TB_mean1.3 showing higher specificity. Due to our small sample size, these metrics will need to be externally validated.

Despite the small sample size, statistically significant differences between HGG and LGG SUV_max as well as between either HGG or LGG SUV_max and normal or venous SUV_mean were noted (for example between HGG and LGG, p < 0.00001) (Fig. 5). No statistically significant difference between venous blood pool and normal parenchymal SUV_mean was noted (p = 0.29).

The correlation of SUV_max and TB_mean with Ki-67 from the evaluable 13 out of 18 lesions was examined. While SUV_max and TB_mean1.3, TB_mean1.6, and TB_mean1.9 were all found to have statistical correlation with Ki-67 levels (p < 0.05), the strongest correlation was with TB_mean1.3 (Fig. 6). TB_mean1.3 has a robust correlation with Ki-67 (R = 0.81, p = 0.00081) with a slightly less robust correlation between SUV_max and Ki-67 values (R = 0.62, p = 0.0227). While there was a range of Ki-67 values observed in the HGG group, a very narrow range of Ki-67 values was observed in the LGG group. Of the sampled LGG, 4/5 lesions had a Ki-67 of 3 or 4 and no isolated analysis of LGG or HGG was thus performed.