Validation of [¹⁸F]FLT as a perfusion-independent imaging biomarker of tumour response in EGFR-mutated NSCLC patients undergoing treatment with an EGFR tyrosine kinase inhibitor

Several studies have evaluated [¹⁸F]FLT as a proliferation biomarker [3, 10, 11, 14]. However, most of these studies have not taken potential perfusion effects into account [10, 11, 14]. During this study, patients will undergo treatment with a TKI, which also intervenes with the vascular endothelial growth factor (VEGF) pathway. This treatment has an inhibiting effect on tumour angiogenesis and its perfusion [20, 21]. Therefore, during this study, it might be possible to assess perfusion changes and to evaluate its effect on [¹⁸F]FLT uptake. As compared to our previous study (3) where we have investigated the relationship between [¹⁸F]FLT uptake and perfusion on whole tumour level, this study will mainly investigate the intra-tumoural relationship between both parameters.

Patients underwent [¹⁵O]H₂O and [¹⁸F]FLT PET/CT scans at three different time points: before treatment and 7 and 28 days after start of treatment with a TKI (erlotinib or gefitinib). The type of TKI was chosen by the treating pulmonary physician: 150 mg of erlotinib or 250 mg of gefitinib orally once a day. In the present study, only those patient studies were included for which all three time points were successfully acquired, i.e. at all three time points, both the [¹⁵O]H₂O and [¹⁸F]FLT PET/CT scans were completed.

Kinetic analysis was performed using in-house developed software in Matlab (version 7.04; MathWorks Incorporation, Massachusetts, USA). For each lesion, the volume of interest (VOI) was defined on the [¹⁸F]FLT PET scan. An averaged image of the last three frames of the scan was produced and lesions were delineated using a 50% threshold of the peak standardised uptake value (SUV_peak).

Image-derived input functions (IDIFs) were generated for both [¹⁵O]H₂O and [¹⁸F]FLT scans, using the early frames in which the first pass of the bolus was best visualised: VOIs of 2 × 2 voxels were positioned in five consecutive axial planes within the lumen of the ascending aorta. These VOIs were then projected onto all image frames to generate a whole blood IDIF. The [¹⁸F]FLT IDIF was calibrated using the radioactivity concentrations in the venous blood samples. Moreover, IDIFs were corrected for both plasma to whole blood ratios and metabolites to obtain calibrated parent [¹⁸F]FLT plasma input functions (detailed description in Frings et al. [3]).

Next, parametric images of tumour VOIs were generated. For perfusion (tumour blood flow (TBF) = K ₁ of [¹⁵O]H₂O), these images were generated using the basis function method [22] and for volume of distribution (V _T ) of [¹⁸F]FLT these images were generated using spectral analysis [23] with optimised settings [24]. Regionally averaged as well as voxel-by-voxel TBF and V _T data were obtained.

The intra-tumoural association between TBF and [¹⁸F]FLT V _T was assessed on a voxel-by-voxel basis by plotting the parametric TBF values against those of [¹⁸F]FLT V _T .

A multiparametric classification was applied to TBF and [¹⁸F]FLT V _T data of all the lesions, using a global threshold for both parameters. The threshold was determined as the average of all tumour values of all the lesions (for all time points), for TBF and [¹⁸F]FLT V _T , respectively. First of all, voxels were classified as low or high TBF and low or high V _T (independently for both parameters), using the threshold (Additional file 1: Figure S1). Voxels with values below the threshold were classified as “low TBF” or “low V _T ” whereas voxels with values above the threshold were classified as “high TBF” or “high V _T ”. This resulted in the distribution of voxels into two categories for both TBF and V _T independently (independent classification).

Next, these initial classifications were combined and voxels were classified into four categories (Additional file 1: Figure S1): lowTBF-lowV _T , lowTBF-highV _T , highTBF-lowV _T and highTBF-highV _T (multiparametric classification) using the method developed in Iqbal et al. [25]. Depending on the number of voxels in each category, several different voxel distribution patterns could be identified.

A quantitative metric, the clustering index (CI), was applied to describe the degree of voxel clustering (spatial grouping) within the tumour [26]. The main purpose for using this method was to gain more insight in the spatial coherence/heterogeneity of tumours and the co-localisation of voxels belonging to the same category (e.g. high FLT and high perfusion). The CI was obtained by measuring the local entropy of the classified voxels and was calculated for three (theoretical) situations to determine the expected range for CI per lesion (Additional file 1: Figure S2). By defining this range, effects of tumour volume and shape were taken into account. The following situations were simulated: (1) a condition where all voxels are classified into one category and maximally clustered within the VOI, resulting in a CI of approximately 0; (2) a condition where voxels are classified into four categories and most optimally clustered, providing a CI of approximately 0.3; (3) a condition where voxels are classified in four categories, showing random distribution and almost no clustering, resulting in a CI of approximately 2. Once the range of the CI, given by difference in CI for minimal and maximal clustering, has been determined for each lesion, the actual observed clustering can be expressed as a percentage using the following equation:

One hundred percent clustering corresponds to the first simulated condition (maximal clustering) whereas 0% clustering corresponds to the third simulated condition (random distribution, no clustering). The CI method has been described and explained in detail by Iqbal et al. [25].

Statistical analyses were performed using SPSS Statistics 20 (IBM Corp.). The correlation between TBF and [¹⁸F]FLT V _T (for each time point) was assessed using Pearson’s correlation coefficient. The ANOVA test was performed to assess statistical differences between tumour clustering at the three different time points.

Initially, ten patients were included in this study. However, four patients did not complete imaging studies with both tracers at all time points, and therefore, these four patients were excluded from the study. Finally, the study included six patients with stage IV NSCLC: one male and five females with an average (±SD) age of 65 ± 6 years (Table 1). A total of six malignant lesions, all of which were primary adenocarcinomas, could be identified in the study group (Table 2). The lesions were analysed at three different time points using 17 [¹⁵O]H₂O and 18 [¹⁸F]FLT PET scans. For one patient, data of the [¹⁵O]H₂O scan obtained 7 days after starting treatment was unavailable due to technical problems. Average TBF (±SD) values (of all lesions) at baseline and 7 and 28 days after treatment were 0.50 ± 0.21, 0.59 ± 0.31 and 0.54 ± 0.30 mL cm^− 3 min^− 1, respectively. Average [¹⁸F]FLT V _T (±SD) values (of all lesions) at baseline and 7 and 28 days after treatment were 5.28 ± 1.71, 3.30 ± 1.09 and 3.16 ± 0.66 mL cm^− 3, respectively. The average lesional values of both parameters are shown in Table 3 and Fig. 1. The global threshold (average ± SD), as used for the multiparametric classification, were 0.53 ± 0.26 mL cm^− 3 min^− 1 and 4.25 ± 1.71 mL cm^− 3 for TBF and [¹⁸F]FLT V _T , respectively.

The association between TBF and [¹⁸F]FLT V _T , as assessed on a voxel-by-voxel basis, is represented for patient one (Fig. 2). The association was assessed at three different time points: baseline (Fig. 2a), 7 days after treatment (Fig. 2b) and 28 days after treatment (Fig. 2c). Pearson’s correlation showed no significant relationship between TBF and [¹⁸F]FLT V _T (baseline, r − 0.14, P = 0.79, 7 days after treatment, r − 0.24, P = 0.69, 28 days after treatment, r 0.41, P = 0.43). At all the three time points, data were scattered over a large range of TBF values. A specific trend, such as decreased tracer uptake with decreased TBF, could not be observed. Over the three time points, a clear decrease in [¹⁸F]FLT V _T could be perceived which was most prominent comparing the data at baseline with the data obtained 7 days after treatment. However, such a trend could not be identified for TBF.

TBF and [¹⁸F]FLT V _T data were evaluated using the independent and multiparametric classification (Fig. 3). Figure 4 represents the voxel distribution patterns for one patient, for both the classifications. The overall evaluation for all patients revealed that across the three time points, changes in TBF were minimal. However, in comparison to TBF, [¹⁸F]FLT V _T underwent a shift in voxel distribution over the three time points. At baseline, most voxels were classified in the “highV _T ” (average 77.1%, range 47.4–99.9%) or “lowTBF-highV _T (average 45.5%, range 14.1–70.1%) and highTBF-highV _T (average 31.5%, range 2.7–59.7%)” category, depending on the classification applied (independent or multiparametric). One week after start of the therapy, most voxels were classified in the “lowV _T ” (average 85.4%, range 65.0–100%) or “lowTBF-lowV _T (average 47.7%, range (18.4–90.4%) and highTBF-lowV _T ” (average 41.9%, range 9.6–73.8%) category. This distribution pattern remained similar afterwards.

Visual inspection of the voxel distribution showed that voxels classified within one category were clustered. Tumours could be visually delineated into several different regions, depending on its TBF and [¹⁸F]FLT V _T . Voxels in the “lowTBF-lowV _T ” category were mainly located at the tumour edges whereas voxels in the “highTBF-highV _T ” category could often be identified at the core of the tumour. Voxels in the remaining categories showed a more dispersed distribution pattern across the lesions.

Quantitative analysis using the CI also showed that voxels were clustered rather than randomly distributed across the lesions. It indicated that the degree of clustering was significantly different from 0% (P < 0.001) for all three time points. Table 4 represents clustering (%) for all lesions at the three different time points (average ± SD; baseline 59.1% ± 9.4%, 7 days post-therapy 66% ± 14.7, 28 days post-therapy 74.8% ± 4.4%). It shows an increase in clustering when comparing data at baseline with data obtained 28 days after treatment (Fig. 5; ANOVA test, P = 0.016).

We hypothesised that TBF could affect tracer uptake: low perfused areas might show low tracer uptake due to perfusion-limited delivery of [¹⁸F]FLT. However, our data reveals that TBF does not seem to affect tracer uptake, as assessed with scatterplots (Fig. 2) and the classifications/cluster index applied. The [¹⁸F]FLT V _T is comparable in low and high perfused areas and does not correlate with TBF. In addition, changes in [¹⁸F]FLT V _T after start of treatment do not seem to accompany the (minimal) changes in TBF. This might indicate that [¹⁸F]FLT distribution and its changes do not depend on perfusion which is a preferable characteristic of the [¹⁸F]FLT PET tracer.

Prior to the multiparametric evaluation, TBF and [¹⁸F]FLT V _T data were evaluated independently. For TBF, most voxels were classified in the “lowTBF” category and this distribution pattern persisted over the three time points. As patients were treated with erlotinib or gefitinib a decrease in TBF was expected. The administered TKIs influence the VEGF signalling pathway through the EGFR pathway. In healthy tissue, activation of the EGFR pathway leads to production of VEGF, the primary inducer of angiogenesis [19, 21]. However, during TKI treatment, both pathways are inhibited. This results in decreased angiogenesis, eventually resulting in decreased TBF and tumour necrosis. De Langen et al. monitored response to antiangiogenic therapy (bevacuzimab and erlotinib) by investigating changes in TBF [9]. They assessed that mean TBF decreased by 11% 3 weeks after starting treatment. Another study also showed a decrease in TBF 2 weeks after treatment with Sunitinib [27]. Mankhoff et al. investigated changes in TBF in patients receiving neoadjuvant chemotherapy [28]. They differentiated between responders and non-responders, the first showing a decrease in TBF whereas the latter showed an increase in TBF. Although our study does not show distinct changes in TBF, it might be possible that these changes occur at a later stage of treatment as angiogenesis and perfusion changes are relatively slow processes.

Furthermore, our data shows that [¹⁸F]FLT V _T decreases after start of TKI treatment with the most prominent decrease observed 7 days after treatment. Kahraman et al. also presented similar results in NSCLC patients, showing decreased [¹⁸F]FLT uptake in early PET images (1 week after starting treatment with erlotinib) [29]. They concluded that early evaluation with [¹⁸F]FLT PET is better than late evaluation (6 weeks after treatment with erlotinib) for response prediction and prediction of progression-free survival. Several other studies also showed a decrease in [¹⁸F]FLT uptake 1 or 2 weeks after therapy, which correlated with progression-free survival [12, 30]. Our study also evaluated survival data to assess a possible relationship between overall survival and the voxel distribution patterns of the multiparametric classification. However, no such relationship could be found, possibly due to the small population size.

In addition, the visual analysis suggested that voxels classified within one category were clustered rather than randomly distributed across the lesions. It was possible to visually delineate a tumour into several different regions, depending on the voxel distribution of TBF and [¹⁸F]FLT V _T , independently and multiparametric. Voxels in the “lowTBF-lowV _T ” category were mainly located at the tumour edges; this might be a partial volume effect: due to tumour movement caused by respiration, the PET signal might be blurred, therefore showing a low PET signal [25]. On the contrary, voxels in the “highTBF-highV _T ” category could often be identified at the core of the tumour, indicating high angiogenesis and proliferation in the core of the tumour [25]. The voxel distribution varied intralesionally as well as interlesionally, indicating spatial heterogeneity of these tumours [31, 32]. This can be the result of heterogeneity at genetic and cellular levels as well as at a tissue level [33‐35]. The first refers to mutational events, which take place during tumour development. These mutations might vary between tumours, which might be one of the mechanisms responsible for differences in response to treatment and resistance to treatment. The heterogeneity at the tissue level indicates differences in vascular density, blood flow, regional metabolism and proliferation. This study assessed the diversity of TBF and proliferation rate within NSCLC tumours and showed that each tumour has its own heterogeneity.

This study has certain limitations, and one of these includes the small sample size (n = 6, with each three scans). To increase the reliability of the study performed, it would be recommended to perform it with a larger sample size in future. Secondly, as during this study patients with lung cancer were scanned (scanning field: thorax), respiratory motion might have led to imaging artefacts (this might have affected further analysis and the results) or reduced spatial resolution. However, the presence of respiratory motion-related artefacts was carefully checked and these were not visually identified. Moreover, the primary tumours in our study were relatively large (≥ 3 cm diameter), and for these tumours, the tracer uptake is usually less affected by motion. Yet, to limit the presence of respiratory motion artefacts, respiratory gating might be a helpful technique for future studies. Finally, while performing the kinetic analysis, lesions on PET scans were delineated using the 50% SUV_peak threshold. This might have led to exclusion of low perfused and/or low tracer uptake areas. However, during our study, visual inspection of the VOI using both PET and CT data did not suggest that tumour tissue was missed and this threshold could be used.

In our study, we evaluated the relationship of tumour perfusion and [¹⁸F]FLT tracer uptake responses in patients suffering from adenocarcinoma and treated with EGFR TKIs. We found that [¹⁸F]FLT responses occurred independent from perfusion (changes), which suggests that [¹⁸F]FLT uptake and its change may be considered as a unique biomarker independent from perfusion. The authors believe that these results are generally applicable to other types of NSCLC patients and therapies. However, for other patients and therapies, a treatment response or change in both perfusion and [¹⁸F]FLT uptake might occur. Our results, however, suggest that [¹⁸F]FLT uptake changes as a response marker can be observed regardless of change in perfusion.

[¹⁵O]H₂O and [¹⁸F]FLT PET are a non-invasive method for the assessment of perfusion and cellular proliferation, respectively. The acquired data was used for generating the multiparametric classification. As it is a new method for analysing changes in perfusion and tracer uptake, further evaluation of this method is necessary to reliably apply this in future studies. Clinical data, such as resection specimens, could be useful for the assessment of perfusion and proliferation (using immunohistochemical staining). It would be interesting to compare pathology with the results of the multiparametric classification and investigate similarities or differences. It would also be interesting to assess perfusion and proliferation at a later stage of treatment. During the follow-up of this study, changes in perfusion were limited, while changes in proliferation were prominent. It might be interesting to see how this evolves when patients are under treatment for a longer time as in that case changes in perfusion might occur or become more prominent as angiogenesis and perfusion changes are relatively slow processes.