Background

Tumor hypoxia is a major cause of treatment failure in patients with locally-advanced cervical carcinoma (LACC) given cisplatin-based chemoradiotherapy [1,2,3]. The outcome of LACC may be improved by personalizing the treatment, and personalized therapy of LACC requires novel biomarkers of treatment outcome [4]. Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) may provide information on the extent of hypoxia in LACC [4,5,6,7], and the outcome of LACC has been shown to be associated with parameters derived from DCE-MRI data [8,9,10,11,12]. However, these associations are fairly weak, and a sound DCE-MRI strategy for predicting the outcome of LACC has yet to be developed.

We have analyzed DCE-MRI data of LACC voxelvise and identified two biomarkers of disease-free survival (DFS) and overall survival (OS): low-enhancing tumor volume (LETV) [13] and tumor volume with increasing signal intensity (TVIS) [14]. LETV refers to the tumor volume showing low contrast enhancement during the first 60 s of the recordings, and TVIS represents the tumor volume showing increasing contrast enhancement during a 6-min-long interval in the late-phase of the DCE-MRI series [13, 14]. The calculation of LETV and TVIS was not based on any pharmacokinetic model.

Two pharmacokinetic models are used frequently to analyze clinical DCE-MRI series: the Brix model [15, 16] and the Tofts model [17, 18]. These models are based on physiological properties of tumors, and consequently, it may be preferable to analyze DCE-MR recordings of tumors using model-based strategies rather than non-model-based strategies [19]. In an earlier study of LACC, we compared the prognostic power of the parameters of the Brix model with that of LETV and TVIS and concluded that biomarkers derived from Brix analysis are not superior to LETV and TVIS [20].

In the present investigation, the DCE-MR recordings of the LACC patients included in our studies of LETV and TVIS were analyzed using the Tofts model. The Tofts model has the advantage to the Brix model that the analysis can be based on the concentration of contrast agent in the tumor tissue rather than signal intensity [15,16,17,18]. The objective of the investigation was to determine whether the parameters of the Tofts model may be clinically useful biomarkers of the outcome of LACC and to compare the prognostic value of these biomarkers with that of LETV and TVIS.

Materials and methods

Patients

Eighty patients with untreated LACC (FIGO stage IB through IVA) admitted to the Norwegian Radium Hospital were included in the study. The demographics of the patients have been described earlier [13, 14]. After DCE-MRI, the patients were treated with concurrent cisplatin-based chemoradiotherapy with curative intent. Briefly, external beam radiation therapy was given in 25 fractions during a period of 5 weeks to a total dose of 50 Gy to the primary tumor, parametria, and ajacent pelvic wall and 45 Gy to the rest of the pelvic region. In addition, 5–6 fractions of intracavitary brachytherapy with a dose of 4.2 Gy per fraction were given to Point A. Chemotherapy with cisplatin (40 mg/m2) was given weekly with a maximum of 6 courses during the radiation therapy period.

The primary endpoints were DFS, defined as the time from diagnosis to local or distant relapse or death, and OS, defined as the time from diagnosis to death. The study was approved by the regional committee of medical research ethics in southern Norway and was carried out in accordance with the Declaration of Helsinki. Informed consent was obtained from all patients.

MRI

The MRI protocol has been described in detail elsewhere [13]. Briefly, the pelvis was scanned with axial T2-weighted fast spin echo sequences, and tumor volume and lymph node status were assessed by examining the T2-weighted images in the dicom viewer Osirix. A gradient echo sequence was used to generate axial proton density-weighted images prior to and after DCE-MRI. DCE-MRI was conducted by using an axial T1-weighted spoiled gradient recalled sequence [repetition time: 160 ms; echo time: 3.5 ms; flip angle: 90°; number of excitations: 1; field of view: 20 × 20 cm2; image matrix: 256 × 256; slice thickness: 5 mm; slice spacing: 6 mm; voxel size: 0.78 × 0.78 × 5.0 mm3; temporal resolution: 29 s; sampling time: 10 min]. Three precontrast images were acquired, and contrast agent was injected manually in a bolus dose of 0.1 mmol/kg Gd-DTPA during a period of ~ 3 s. To enable calculation of the concentration of Gd-DTPA in the tumor tissue, a calibration tube with two chambers, one filled with saline and the other with a 0.5-mmol Gd-DTPA solution, was inserted into the vagina prior to the imaging. The tube was positioned as close to the tumor as possible, and acceptable positioning was verified by MRI.

Image analysis

Voxel-by-voxel analysis of DCE-MR images was conducted by using software developed in Matlab. Minor movements of the tumor tissue during image acquisition were corrected for by coordinate mapping. To avoid significant influence of body cavities and tumor necrosis, voxels showing signal intensities in T2-weighted images consistent with the presence of air or water were excluded from analysis. Gd-DTPA concentrations were calculated from signal intensities using the method of Hittmair et al. [21], as described in detail elsewhere [22].

Plots of Gd-DTPA concentration versus time after contrast injection were generated, and the Levenberg-Marquardt least squares minimization method was used to fit curves to the data, using the standard Tofts equation [18]:

$$ {C}_{\mathrm{t}}(T)={K}^{\mathrm{t}\mathrm{rans}}\bullet {\int}_0^T{C}_{\mathrm{p}}(t)\bullet {e}^{-\frac{K^{\mathrm{t}\mathrm{rans}}\bullet \left(T-t\right)}{V_e}}\ dt. $$

Here, Ct(T) is the tumor tissue concentration of Gd-DTPA at time T, Ktrans is the volume transfer constant of Gd-DTPA, ve is the fractional distribution volume of Gd-DTPA in the tumor tissue, and Cp(t) is the arterial input function:

$$ {C}_{\mathrm{p}}(t)=\mathrm{A}\bullet {e}^{-\mathrm{B}t}+\mathrm{C}\bullet {e}^{-\mathrm{D}t}, $$

where Cp(t) is the plasma concentration of Gd-DTPA at time t, and A = 5.10 mM, B = 14.2 s− 1, C = 0.99 mM, and D = 0.159 s− 1. Frequency distributions and parametric images of Ktrans and ve were generated by using SigmaPlot software.

Statistical analysis

Because standard first-line treatment fails in 60–70% of the patients diagnosed with LACC [23], associations between Tofts parameters and DFS or OS were studied by dividing the patient cohort into two groups consisting of one-third and two-thirds of the patients [13]. Thus, the DFS and OS of the 26 patients with the lowest Ktrans values were compared with the DFS and OS of the 54 patients with the highest Ktrans values, and similar comparisons were carried out for ve. Kaplan–Meier curves were compared by using the log-rank test. Univariate and multivariate Cox proportional hazard analyses were used to evaluate the prognostic power of clinical and DCE-MRI–derived parameters. The Spearman rank order correlation test was used to search for correlations between parameters. Probability values of p < 0.05 were considered significant.

Results

Relevant anatomical MR images of a representative LACC patient are presented in Fig. 1. The position and signal intensities of the calibration tube are illustrated in a sagittal and two axial precontrast T1-weighted images (Fig. 1a), and a proton density-weighted image, a precontrast T1-weighted image, and a postcontrast T1-weighted image are shown to illustrate the signal intensities of the tumor tissue (Fig. 1b).

Fig. 1
figure 1

Representative anatomical MR images. a A sagittal and two axial T1-weighted scans showing the signal intensities of the two-chamber calibration tube. The dashed horizontal lines indicate the positions of the axial scans. b A proton density-weighted image, a precontrast T1-weighted image, and a postcontrast T1-weighted image showing the signal intensities of the tumor tissue. Scale bars: 2 cm

Full size image

Tumor signal enhancement after Gd-DTPA administration differed substantially among individual patients, and Fig. 2 presents Ktrans data for two representative patients, one with a high-enhancing tumor (Fig. 2a) and the other with a low-enhancing tumor (Fig. 2b). The Tofts model was seen to be well suited for analyzing the DCE-MRI series. Single-voxel plots of Gd-DTPA concentration versus time were well fitted by the model, and the numeric values of Ktrans were generally higher in high-enhancing than in low-enhancing tumors. However, the intratumor heterogeneity in Ktrans was substantial, as illustrated by the Ktrans frequency distributions and parametric maps in Fig. 2.

Fig. 2
figure 2

Representative Ktrans data. Single-voxel plots of Gd-DTPA concentration versus time after contrast injection, Ktrans frequency distribution, parametric Ktrans image, and binary Ktrans image of a a high-enhancing tumor and b a low-enhancing tumor. The dark grey regions in the frequency distributions and binary images represent RV-Ktrans. Scale bars: 1 cm

Full size image

Detailed analyses of the Ktrans frequency distributions were carried out to investigate whether Ktrans may provide valid prognostic information on DFS and OS. First, the Ktrans values at each integer percentile of the frequency distributions were determined, and a log-rank test was conducted for each percentile to examine whether the Ktrans values at any of the percentiles could discriminate between the two patient groups. Plots of the log-rank p value versus Ktrans percentile were established, and the plots revealed that the Ktrans values at percentiles above 30 provided p values < 0.05 for both DFS and OS (Fig. 3a). Receiver operating characteristic (ROC) analysis was carried out to determine the optimal Ktrans percentile. The area under the ROC-curve is plotted versus Ktrans percentile in Fig. 3b, and the optimal percentile, corresponding the largest area under the ROC-curve, was found to be the 35 percentile for DFS as well as OS. The Ktrans value at this percentile was termed 35p-Ktrans.

Fig. 3
figure 3

Log-rank and ROC analyses. The values of Ktrans at each percentile of the frequency distributions and the tumor volumes (RV-Ktrans) with Ktrans values below a wide range of threshold values were calculated for the 80 tumors included in the study, and for each Ktrans percentile and each Ktrans threshold value, the outcome of the patients with high values of Ktrans or RV-Ktrans was compared with the outcome of those with low values, using the log-rank test with DFS and OS as endpoints. ROC analysis was carried out to indentify the Ktrans percentile and Ktrans threshold value with the highest discriminative power. a Log-rank p value versus Ktrans percentile. b Area under ROC-curve versus Ktrans percentile. c Log-rank p value versus Ktrans threshold value. d Area under ROC-curve versus Ktrans threshold value

Full size image

Next, we calculated the number of voxels and the corresponding tumor volumes with Ktrans values below a wide range of threshold values up to 0.50 min− 1, using increments of 0.0025 min− 1. A log-rank test was carried out for each threshold value to investigate whether the tumor volumes at any of the threshold values could discriminate between the two patient groups. Plots of the log-rank p value versus Ktrans threshold value were generated for DFS and OS, and these plots showed that Ktrans threshold values between 0.07 min− 1 and 0.24 min− 1 provided p values < 0.05 for both endpoints (Fig. 3c). ROC analysis revealed that the optimal Ktrans threshold value was 0.13 min− 1, regardless of whether DFS or OS was considered (Fig. 3d). The tumor volume with Ktrans values below 0.13 min− 1 was presumed to represent the risk volume (RV) and was termed RV-Ktrans.

Binary tumor maps showing the distribution of voxels with Ktrans values above or below 0.13 min− 1 are included in Fig. 2. In general, the voxels constituting RV-Ktrans were contiguous in tumors having a large RV-Ktrans (Fig. 2b).

Kaplan–Meier plots for DFS and OS based on 35p-Ktrans and RV-Ktrans are presented in Fig. 4. The patients with low Ktrans values did worse than those with high Ktrans values. The 5-year survival rates of the patients with low 35p-Ktrans values and those with high 35p-Ktrans values were 44 and 72%, respectively (DFS: p = 0.032) and 46 and 76%, respectively (OS: p = 0.0065). Similarly, the patients with large RV-Ktrans and those with small RV-Ktrans showed 5-year DFS rates of 41 and 73%, respectively (p = 0.0044) and 5-year OS rates of 42 and 78%, respectively (p = 0.0034).

Fig. 4
figure 4

Treatment outcome. Kaplan–Meier curves for DFS and OS of LACC patients stratified by 35p-Ktransa,b and RV-Ktransc,d. p values: log-rank test

Full size image

Univariate Cox regression analysis showed that 35p-Ktrans and RV-Ktrans had significant impact on OS, and that RV-Ktrans but not 35p-Ktrans had significant impact on DFS (Table 1). A similar analysis of clinical parameters revealed that DFS and OS were influenced significantly by tumor volume and FIGO stage, but not by lymph node status, tumor histology, and patient age (Table 1). Furthermore, RV-Ktrans was found to be the only independent prognostic factor of DFS and OS in multivariate Cox regression analysis involving tumor volume, FIGO stage, lymph node status, and 35p-Ktrans or RV-Ktrans (Table 1).

Table 1 Cox regression analysis of clinical and DCE-MRI–derived parameters
Full size table

The DCE-MR recordings of the patients included in this study have been subjected to non-model-based analyses previously, and LETV and TVIS were identified as independent prognostic factors [13, 14]. LETV and TVIS represent tumor RVs, and plots of RV-Ktrans versus LETV and TVIS revealed strong correlations between the RV derived from the Tofts analysis and the RVs derived from the non-model-based analyses (p < 0.0001; Fig. 5). The horizontal and vertical lines in Fig. 5 represent the border between small and large RVs, and show that the majority of the patients were stratified into the same risk group by the model-based and non-model-based RVs.

Fig. 5
figure 5

Model-based versus non-model-based RVs. Log scale plots of RV-Ktrans versus LETV a and TVIS b for patients with LACC. The patient cohort was divided into two groups consisting of one-third and two-thirds of the patients, and the discrimination levels are indicated by horizontal lines for RV-Ktrans and by vertical lines for LETV and TVIS. Points: individual tumors. Curves: linear regression. p values: Spearman rank order correlation test

Full size image

By subjecting the ve frequency distributions of the tumors to analyses similar to those described above for the Ktrans frequency distributions, it was seen that ve did not provide valid prognostic information on DFS or OS. Parametric ve images and ve frequency distributions of the tumors used as examples in Fig. 2, and plots of log-rank p value versus ve percentile and log-rank p value versus ve threshold value are presented in Additional file 1: Figure S1.

Discussion

Imaging plays an important role in the diagnostics, radiation therapy planning, and treatment monitoring of LACC [4]. In addition to MRI, positron emission tomography (PET)/computed tomography (CT) has shown promise for hypoxia imaging. Current studies are investigating the potential of several PET hypoxia tracers, including 18F-fluoromisonidazole, 18F-fluoroerythronitroimidazole, 18F-fluoroazomycin-arabinoside, and 60Cu-diacetyl-bis(N-methylthiosemicarbazone), and interesting observations have been reported [24]. It has also been hypothesized that 18F-fluorodeoxyglucose (18F-FDG) may be a clinically useful surrogate hypoxia tracer [24]. If this hypo-thesis proves to be valid, 18F-FDG PET/CT may turn out be a powerful imaging modality for LACC, since 18F-FDG PET/CT already is used successfully to detect pathological lymph nodes and high-risk radiation therapy targets [25, 26].

The potential of DCE-MRI as an imaging modality for providing hypoxia-associated biomarkers of the outcome of LACC was investigated in the study reported in this communication. DCE-MRI series of LACC were analyzed by using the Tofts pharmacokinetic model, and the prognostic power of biomarkers derived from the Tofts analysis was compared with that of biomarkers derived from non-model-based analyses. Tofts analysis has the advantage to non-model-based analyses that it is based on the physiology of the imaged tumor, but to utilize this advantage in full, information on the concentration of contrast agent in the tissue is required. This information was provided in the present study by using a two-chamber vaginal calibration tube.

Significant associations were found between p35-Ktrans or RV-Ktrans on the one hand side and DFS or OS on the other, as revealed by univariate Cox regression analysis. Multivariate analysis including clinical parameters and p35-Ktrans or RV-Ktrans showed that RV-Ktrans was a strong independent prognostic factor of DFS and OS, whereas p35-Ktrans, tumor volume, FIGO stage, and lymph node status were not independent prognostic factors. These observations suggest that Tofts analysis of DCE-MRI data can provide biomarkers of the outcome of LACC. Moreover, the prognostic power of RV-Ktrans is strong compared with that of the Ktrans value at any percentile of the Ktrans frequency distribution.

Poor treatment outcome was associated with low values of Ktrans and, consequently, large values of RV-Ktrans. Clinical investigations have revealed that low Ktrans values reflect poor vascularization and blood supply in LACC [9, 27, 28]. Preclinical studies of human cervical carcinoma xenografts have shown strong inverse correlations between median Ktrans and fraction of hypoxic tumor tissue, regardless of whether radiobiological or pimonidazole-based immunohistochemical assays were used to quantify tumor hypoxia [5, 29,30,31]. Therefore, it is reasonable to assume that RV-Ktrans mirrors the hypoxic subvolume in tumors of the uterine cervix. This assumption is consistent with investigations having shown that poor outcome of LACC is associated with low oxygen tension in the primary tumor, as measured with polarographic Eppendorf pO2 electrodes [1,2,3].

RV-Ktrans correlated strongly with the non-model-based DCE-MRI–derived parameters LETV and TVIS, and the vast majority of the patients were stratified into the same risk group by RV-Ktrans, LETV, and TVIS. Furthermore, the RV-Ktrans-based Kaplan-Meier plots for DFS and OS reported here were similar to the LETV-based and TVIS-based Kaplan-Meier plots reported previously [13, 14]. Consequently, the prognostic power of biomarkers identified by Tofts analysis does not appear to be stronger than that of biomarkers identified by non-model-based analyses.

RV-Ktrans, LETV, and TVIS have specific disadvantages and advantages as biomarkers of LACC. RV-Ktrans has the disadvantage that an arterial input function and long scanning times are needed for its assessment. The advantage of RV-Ktrans is that it is related to the extent of tumor hypoxia. The assessment of LETV requires signal intensity-dependent threshold values [13], and because signal intensities vary depending upon MR protocol and scanner, radiology departments have to establish their own threshold values. The advantage of LETV is that short scanning times of only 1–2 min are needed, and this advantage is of significant importance because scanning time is a major limiting factor in many centers. Similar to the calculation of RV-Ktrans, the calculation of TVIS requires long scanning times, but the advantage of TVIS is that the calculation is independent of signal intensity, thus facilitating comparisons of results across institutions.

The current investigation has some limitations. First, T2-weighted images were used to delineate tumors and to exclude image regions influenced by body cavities and tumor necrosis, and despite the fact that these tasks were carried out by two experienced radiologists, there may have been some uncertainties in the assessment of the regions of interest. Second, the MR images were recorded several years ago when the spatial and temporal resolutions of the recordings were poorer than today’s standard due to scanner limitations. Third, the Tofts analysis was performed by using a population-based arterial input function rather than individual arterial input functions. Fourth, only 80 patients were included in the study, and consequently, it requires validation in an independent patient cohort. Despite these limitations, important conclusions can be drawn from our investigation.

Conclusions

By subjecting DCE-MRI series of LACC to pharmacokinetic analysis using the Tofts model, prognostic factors independent of well-established clinical prognostic factors can be provided. Furthermore, RV-Ktrans has stronger prognostic power than the Ktrans value at any percentile of the Ktrans frequency distribution. However, the prognostic value of RV-Ktrans is not necessarily superior to the prognostic value of RVs assessed by non-model-based analyses.