Repeatability of quantitative ¹⁸F-FLT uptake measurements in solid tumors: an individual patient data multi-center meta-analysis

To identify all relevant publications, a systemic search was performed in PubMed, EMBASE.​com and the Cochrane Library (via Wiley) from inception to October 20, 2016 (last elicitation). A combination of the search terms comprising ‘FLT-PET’ and ‘neoplasms’ was used. This included MeSH terms and controlled terms from EMtree for PubMed and EMBASE.​com, respectively, as well as free-text terms. We only used free-text terms in the Cochrane Library (see supplemental data). All potentially relevant titles and abstracts were screened for eligibility. Full-text articles were checked for eligibility criteria where necessary. References of eligible publications were checked for relevant publications. We have also checked ClinicalTrials.​gov and The European Union Clinical Trials Register for ongoing and unpublished studies.

Studies were included if they met the following criteria:

Studies were excluded if they met the following criteria:

Sites from all identified cohorts were contacted, and permission was requested to re-analyze the original ¹⁸F–FLT PET repeatability scans. All datasets consisted of 60- or 95-min dynamic test and retest ¹⁸F–FLT PET scans. Where permission was granted, original ¹⁸F–FLT scans of all individual patients were supplied in DICOM or Analyze format. Prior to re-analysis, all scans were checked for technical issues and artifacts. If any technical issues or artifacts were present, data were cross-checked with the original research teams. After checking of the scan data, static standard uptake value (SUV) images were generated from the dynamic images: 40–65 or 45–60 min post-injection, depending on the original frame definition. A 5-mm Gaussian filter was applied to the non-smoothed reconstructed images to match the spatial resolution between existing datasets and with previously published data. New volumes of interest (VOI) were defined by segmenting tumors using a 50% isocontour of the SUV_peak (1.2 cm in diameter sphere positioned to maximize its mean value), adapted for local background (in-house developed software) [12, 18]. For each VOI, SUV_max, SUV_mean, SUV_peak, proliferative volume (50% threshold of SUV_peak corrected for local background) and total lesion uptake (TLU, product of SUV_mean and proliferative volume) was determined. These quantitative ¹⁸F–FLT uptake metrics were checked for outliers and discrepancies with the original data, however no important issues were identified. In addition, tumor-to-blood ratios (TBR) were calculated by normalizing tumor SUVs to the bloodpool SUV_mean of a large vascular structure (2 × 2 voxel VOI in five consecutive planes) [19]. ¹⁸F–FLT uptake in the tumor was normalized to the SUV_mean of the carotid artery in HNC data and to the ascending aorta for all other lesions. All SUVs were calculated by normalizing the radioactivity concentrations by the injected ¹⁸F–FLT dose and body weight and were corrected for physical decay.

Repeatability of the quantitative uptake and volume metrics was determined by calculating the mean and standard deviation (SD) of the percentage differences between the two baseline scans:

Normality of the data was assessed using histogram analyses and quantile-quantile plots (data not shown). The repeatability coefficients (RC) were calculated as 1.96 × SD of the percentage differences. A paired t test was performed to test for significant differences in mean uptake between both baseline scans. To assess the significance of differences in RC, the Levene’s test was used. Moreover, the intra-class correlation coefficient (ICC) using a two-way mixed model, model II regression analysis [20] and Bland–Altman plots were used to evaluate correlations and biases between the test-and-retest scans. The effect of various lesion selection strategies on repeatability was evaluated: lesions ≥ 4.2 ml (diameter ≥ 20 mm) [18], SUV_max ≥ 4.0 [10, 11], hottest lesion per scan (highest SUV_max) or primary lesions only. In addition, the uptake values of individual lesions were averaged per patient to obtain the averaged uptake and assess repeatability on a patient level. All statistical analyses were performed using SPSS 22.0 (SPSS, Chicago, IL, USA).

The literature search generated 1728 results: 630 in PubMed, 1076 in EMBASE.​com and 22 in the Cochrane Library. No ongoing or unpublished trials were identified. After removing duplicates, 1172 references remained (Fig. 1). Out of 1172, four articles (five patient cohorts) were considered eligible [14‐17]. We obtained permission to re-analyze the original ¹⁸F–FLT repeatability data from four of these cohorts, comprising data of 30 patients and 52 individual lesions (ten in breast cancer [14], nine in head and neck squamous cell carcinoma [15], and 33 in non-small cell lung cancer patients from two cohorts [15, 16]; Fig. 2). All patients were included in this individual patient data meta-analysis and no scans had to be excluded. An overview of the cohorts can be found in Table 1.

SUV metrics were lower in the lung cancer dataset from Trigonis et al. [16] compared to the other three datasets (average SUV_mean: 2.4 vs. 3.5, respectively; p < 0.05). In addition, the SUV_max and SUV_peak values in the breast cancer dataset from Kenny et al. [14] were higher compared to those from de Langen et al. [15]. Proliferative volumes and TLU were significantly smaller in the HNC group and the NSCLC lesions in the dataset from Trigonis et al. [16] were also significantly smaller than in the de Langen et al. dataset [15]. Despite overall proliferative volumes of the retest scan being significantly larger than the test scans (MATV: 14.5 vs. 15.6 ml, p = 0.02), no differences were found between the SUV metrics from test-and-retest scans (Table 2). When assessed per site, a small but significant difference in proliferative volume and TLU was only found in the dataset from Trigonis et al. (mean difference −2.3 ml and −4.2 ml respectively, p < 0.01) [16].

Correlations between test-and-retest scans were strong for all uptake metrics per lesion as well as averaged per patient (R² ≥ 0.93 and ICC ≥ 0.96, Fig. 3). Moreover, no systematic bias was present between both scans as revealed by the correlation plots (slope, 0.98–1.04, Fig. 3) and the Bland–Altman plots (Fig. 3). Overall, the best repeatability of quantitative ¹⁸F–FLT PET/CT was obtained using SUV_peak (RCs 23.1%, Table 3). No differences in RCs were found between the individual SUV metrics.

Variability of proliferative volume and TLU (RCs 36.0 and 36.4%, respectively) were significantly worse than for SUV metrics, with an average increase in RC of 9.6 ± 6.6% (p ≤ 0.02)(Fig. 4). When the datasets were evaluated individually, variability of SUV_peak and SUV_mean within the de Langen et al. [15] cohorts was significantly smaller compared to those of the breast cancer dataset, the only one reconstructed with FBP (p < 0.02) [14]. In general, the largest variability was seen in the latter dataset. When comparing only the OSEM reconstructed datasets, RCs for SUV_max, SUV_peak, and SUV_mean changed to 25, 20, and 17% respectively, but RCs of proliferative volumes and TLU remained close to 35%. An overview of the absolute repeatability coefficients for each quantitative uptake metric can be found in supplemental Tables 4 and 5.

Assessment of repeatability on a patient level improved repeatability in general (Table 4). Improvement of repeatability weighted for lesions number was < 2% compared to unweighted averaging per patient. For the SUV metrics, a decrease in RC was largest in the de Langen dataset [15]. Only SUV_mean showed a slight increase in variability, which was caused by one lesion with a 53% difference (4 SDs) between both scans from the breast cancer dataset. If excluded, repeatability of SUV_mean improved to 19%, while other SUV metrics remained unaffected. RCs of proliferative volume and TLU also decreased to < 30%, with the exception of the breast dataset [14].

Assessing repeatability including only lesions with SUV_max ≥ 4.0 decreased variability of volumetric metrics but did not influence RCs of SUV metrics (RCs 26–28%, Fig. 5). The former is mainly caused by a large decrease of RCs in the Trigonis dataset (−20%). If only lesions larger than 4.2 ml were included in the analysis, no significant change in variability of SUV, proliferative volume or TLU was seen (RCs 22–25% and 34–36%, respectively). Similar results were observed when only the hottest or primary lesions were assessed. Combining the two selection criteria SUV_max ≥ 4.0 and lesions ≥ 4.2 ml did not further improve results. No significant change in repeatability of SUV metrics was seen when analyzing cohorts individually. In addition, applying lesions selection criteria to the per-patient analysis did not decrease variability of SUV and volumetric ¹⁸F–FLT uptake measures.