Multiparametric functional MRI and ¹⁸F-FDG-PET for survival prediction in patients with head and neck squamous cell carcinoma treated with (chemo)radiation

For this prospective single-center study, approved by our ethical committee (Trial NL3946, NTR4111), written informed consent was obtained from all patients. Previously untreated patients with histologically proven HNSCC, planned for curative (chemo) radiotherapy, and those who underwent ¹⁸F-FDG-PET/low-dose CT and DWI-DCE-MRI were consecutively included between 2013 and 2018. Exclusion criteria were nasopharyngeal tumors, age < 18 and inadequate image quality. Within 5 weeks after baseline imaging, treatment was initiated consisting of pre-determined radiotherapy (70Gy in 35 fractions in 7 weeks, or accelerated with 70Gy in 35 fractions in 6 weeks) with/without concomitant chemotherapy (3-weekly 100 mg/m² cisplatin), or cetuximab (400 mg/m² before radiotherapy initiation and then weekly 250 mg/m² for 7 weeks). HPV status was determined by p16-immunostaining followed by DNA-PCR on p16-immuno-positive cases. In clinical practice, for lesions outside the oropharynx, HPV positivity is not causally associated with HNSCC [25] and not routinely tested for HPV status, therefore excluded in the survival analyses. Qualitative variables were transformed into numbers: gender (female = 0, male = 1), T-stage (T2 = 2, T3 = 3, T4 = 4), N-stage (N0 = 0, N1 = 1, N2 = 2), HPV (negative = 0, positive = 1), location PT (oropharynx = 1, hypopharynx = 2), smoking (pack years), alcohol (< 3 units/day = 0, ≥ 3 units/day = 1), intoxications (none = 0, smoking < 10 pack years = 0, alcohol < 3 units/day = 0, smoking (> 10 pack years) or alcohol (> 3 units/day) = 1, smoking and alcohol use = 2) [26].

MRI was performed on a 3.0T Ingenuity MR scanner (Philips Healthcare) utilizing a 16-channel neurovascular coil. DWI was performed using fat-suppressed single-shot spin-echo echo-planar imaging (SS-SE-EPI); TR = 500 ms, TE = 105 ms; echo-planar imaging factor = 35; sensitivity encoding factor = 3.5; field of view = 230 × 230 mm; slice thickness = 2 mm; intersection gap = 0.3 mm; matrix = 128 × 128; receiver bandwidth = 2735.7 Hz per pixel. Ten b values were used: 0, 10, 25, 50, 75, 150, 300, 500, 750, and 1000 s/mm². The ADC map was produced by vendor-provided software.

DCE-MRI was performed, using 3-dimensional T1-weighted fast field echo (FFE); TR/TE = 4.8/2.4 ms; flip angle = 12; FOV = 230 × 230 × 180 mm; matrix = 144 × 144; 75 dynamic acquisitions of 4.16 s; signal averages = 2. The dynamic scan was preceded by scans with variable flip angles (2, 5, 10, 15, and 20) to estimate quantitative native T1 maps, which were later used to convert signal intensity of the DCE scan into contrast agent concentration curve map [27]. Intravenous bolus injection of 0.2 ml/kg of body weight Gd-DOTA (Dotarem, France) was administered after 3 dynamic acquisitions (3 ml/s followed by 25 ml saline flush).

¹⁸F-FDG-PET/low-dose CT was performed according to EANM guidelines 2.0 on a Gemini TF-PET/CT (Philips Healthcare) with EARL accreditation [28]. Low-dose CT (120 kV; 30 mAs) was performed. Whole-body ¹⁸F-FDG-PET/CT was performed in arms down position in radiotherapy mask, from mid-thigh to skull vertex, 60 min after intravenous administration of 2.5 MBq/kg ¹⁸F-FDG, 2 min per bed position. ¹⁸F-FDG-PET images were reconstructed using vendor-provided reconstruction protocol with photon attenuation correction, matrix size = 144 × 144, and voxel size = 4 × 4 × 4 mm. Post-reconstruction resolution was 5 mm full width at half maximum.

¹⁸F-FDG-PET/CT delineation was performed by semi-automatic delineation by a nuclear medicine specialist (B.Z.) using 50% of tumor-specific SUVpeak threshold, corrected for blood glucose level. Details on this method were published previously [29].

Imaging parameters were extracted from both PT and LNM whole-lesion ROIs of each observer. Anatomical total lesion volume, i.e., gross tumor volume (GTV), was calculated for each ROI on each imaging map (ADC_GTV, DCE_GTV, and metabolic active tumor volume (MATV)). The following quantitative imaging features were calculated per observer by averaging all voxels included in the whole-lesion ROI.

IVIM feature extraction of perfusion fraction (f), perfusion coefficient (D*), and diffusion coefficient (D) was performed with Olea Sphere (Olea Medical, La Ciotat, France), after motion correction.

DCE-MRI analysis was processed with in-house built software (Dynamo; [27]), performing quantitative pharmacokinetic analysis using the 2-compartment Tofts model [8] with patient-specific arterial input function (AIF) obtained from manual delineating the most cranial part of the external carotid artery. The following features were extracted: K^trans (transfer rate of contrast agent from plasma to extravascular, extracellular space); V_e (fractional volume of extracellular extravascular space); K_ep (transfer rate of contrast agent from extravascular, extracellular space to plasma).

¹⁸F-FDG-PET/CT in-house built software (Accurate; [28]) automatically calculated SUV_max, SUV_mean, and SUV_peak (i.e., peak value of 8 highest voxels) based on all included voxels of the ROI, and whole-lesion MATV, and total lesion glycolysis (TLG = SUV_mean × MATV).

The average of the above written extracted parameters for both observers was used for analyses.

Correlations were assessed between parameters for PT and LNM separately (Pearson correlation coefficient). Differences in imaging parameters among T-stages, N-stages, and intoxications were assessed with the Kruskal-Wallis tests. In order to capture HPV status-specific tumoral characteristics, associations between parameters of patients with locoregional control (LRC) and failure (i.e., recurrence; LRF), and survival and death (univariate Cox regression analysis) were assessed. Bonferroni’s correction for multiple testing was applied.

Secondly, multivariable Cox regression analysis was performed of all PT parameters for each modality separately with a backwards Wald test (p value significance threshold of 0.157 according to the Tripod statement [30, 31]). All quantitative parameters per modality were corrected for significant clinical parameters (gender, age, T-stage, N-stage, PT location, intoxications) by combination in the backwards Wald elimination analysis.

Thereafter, a CoxBoost analysis was performed to fit a Cox proportional-hazards model by component-wise likelihood-based boosting, to deal with the amount of features relative to the events. Internal validation was performed, using bootstrap cross-validation with 500 bootstrap samples. Due to lacking of LNM parameters in N0 patients, these LNM parameters were excluded in the multimodality CoxBoost analysis in order to remain statistically robust.

All predictive PT parameters were given a score 1 when they were higher than the parameter’s median value, which was based on all included patients. By summing up the points, a risk stratification system was constructed. Thereafter, RFS and OS were assessed, stratified for T-stage, AJCC (7th edition), and risk scores (log-rank test; Kaplan-Meier curves).

Between 2013 and 2018, 81 patients were consecutively recruited (Fig. 1). Nine patients were excluded because of non-curative or surgical treatment and 2 because of significant low image quality.

The final study population consisted of 70 patients (Table 1) with a PT located in the oropharynx (n = 56) or hypopharynx (n = 14). Among the oropharyngeal tumors, the HPV status of 24 patients was positive (43%). Fifty-four patients received concurrent cisplatin-based chemoradiotherapy. Ten patients received weekly cetuximab with concurrent radiotherapy (70Gy). Six patients received radiotherapy only.

The mean follow-up was 22.1 months (IQR 14.3–29.4). Seventeen patients (24.3%) developed locoregional recurrence. Twenty (28.6%) patients died during follow-up, all deaths being related to HNSCC (Table 1).

Seventy PT ROIs were drawn and 59 lymph node metastasis ROIs (largest LNM) on each modality (Table 1). The comparison of both observers resulted in no significant different values and a high interobserver correlation (Supplement 1). A Dice index in primary tumors of 0.88 at the DWI/IVIM and 0.85 at DCE delineation was found (not tabulated). For LNM, a Dice index of 0.97 at DCE and 0.92 at DWI/IVIM delineation was found (not tabulated). Primary tumor ADC_GTV, D, f, and D*, DCE_GTV, and all ¹⁸F-FDG-PET values (Supplement 2) were significantly higher in advanced T-staged tumors (all p ≤ 0.02). In advanced N-staged tumors, PT ADC_GTV and V_e were significantly higher (p = 0.021 and p = 0.023, respectively). In HPV-negative tumors, ADC_mean, D, D*, SUV_max, SUV_mean, and SUV_peak were significantly higher than HPV-positives (p < 0.043). In patients with intoxications, ADC_mean, D, and D* were significantly different among the different categories (all p < 0.027).

In LNM (Supplement 3), K_ep and all ¹⁸F-FDG-PET parameters were significantly higher in advanced N-stages (p = 0.025, p ≤ 0.016, respectively). In HPV-negative tumors, D was found to be significantly higher (p = 0.002) and D* lower (p = 0.007) than in HPV-positive tumors. In patients with intoxications, f was found to be significantly lower (p = 0.026).

The inter-modality correlation in PT between ¹⁸F-FDG-PET-, DWI/IVIM-, and DCE-derived parameters (Supplement 4) was only significant among GTV parameters (ADC_GTV, DCE_GTV, TLG, and MATV), SUV_peak, and SUV_mean (range: r = 0.434–0.915). In LNM (Supplement 5), only volume parameters of LNM ADCGTV, MATV, and DCEGTV correlated significantly (range: r = 0.399–725). The intra-modality correlation for PT and LNM (Supplement 6) resulted in significant internal moderate correlation of ADC parameters (range: r = 0.432–0.794) and DCE parameters (r = 0.637–0.741) and high correlation of ¹⁸F-FDG-PET parameters (r = 0.409–0.995).

The univariate analysis (Table 2) showed that HPV-negative status and the combined intoxications were associated with locoregional recurrence (LRF; p = 0.036, p = 0.031, respectively). High ADC_GTV, DCE_GTV, K^trans, V_e, and TLG values of primary tumors were significantly associated with LRF (all p ≤ 0.047).

The multivariate analysis per modality (Table 2), corrected for significant clinical parameters (i.e., HPV and intoxications), showed that high primary tumor ADC_GTV, DCE_GTV, K^trans, V_e, and MATV remained predictive for LRF (all p ≤ 0.048). For LNM, only DCE_GTV remained significantly predictive for LRF (p = 0.018). The subgroup analysis in HPV-negative patients is shown in Supplement 7.

The multivariable CoxBoost analysis (Table 4), combining all modalities and clinical parameters, showed that HPV status, intoxications, ADC_GTV, K^trans, and V_e remained predictive for LRF (C-index of 0.546). The log-rank test (Fig. 2) showed that these risk factors were significantly predictive (p = 0.023) for LRF (Fig. 2b), whereas risk stratification per T-stage (Fig. 2a) was not significantly predictive (p = 0.92).

Primary tumor univariate analysis (Table 3) showed that clinical parameters HPV status, PT location, intoxications (p ≤ 0.047), and imaging parameters ADC_GTV, ADC_mean, D*, D, DCE_GTV, V_e, MATV, and SUV_max, were significantly associated with OS (all p ≤ 0.047). For LNM, SUV_max, SUV_mean, and SUV_peak were associated with OS (all p ≤ 0.015).

In multivariate analysis per single modality (Table 3), corrected for clinical parameters (i.e., HPV status, hypopharyngeal PT location, intoxications), ADC_GTV (p = 0.004), D* (p = 0.016), DCE_GTV (p = 0.001), V_e (p = 0.019), MATV (p = 0.088), and SUV_max (p = 0.001) remained predictive for OS. In LNM, only SUV_max (p = 0.055, HR = 0.563) and SUV_mean (p = 0.005, HR 3.536) remained predictive for OS. The subgroup analysis in HPV-negative patients is shown in Supplement 7.

The multivariable CoxBoost analysis combining all PT parameters of all modalities, including all clinical parameters (Table 4), shows that N-stage, HPV status, PT location, intoxications, PT ADC_GTV, ADC_mean, D*, K^trans, V_e, SUV_max, and TLG remain predictive for OS, with a C-index of 0.664.

Predictive parameters scored as risk factors for OS (Fig. 2) were significantly predictive (p = 0.046) in the log-rank test (Fig. 2d) when combined, whereas risk stratification per T-stage (Fig. 2c) was found not significant (p = 0.188).

Advanced stage tumors (high T-stage) and HPV-negative status had significant higher diffusion (high ADC_mean, D), higher permeability (K^trans, V_e), and lower perfusion (low f and D*), implying different tumor characteristics than early stage and HPV-positive tumors. These parameters were also found to be associated with an adverse outcome. This is in line with literature, which described the decrease of cellularity due to apoptosis/necrosis (increased ADC_mean and D) to be associated with treatment resistance and thereby with poor prognosis [18]. In contrast, in studies which excluded areas of necrosis in the ROI, lower ADC values were found in high-grade tumors with high cellularity. In the current study, HPV-negative patients had a higher ADC value than HPV-positive patients, which was in line with other studies regardless of including [10] or excluding [32‐35] necrotic areas. An increase of permeability (increased K^trans) is possibly due to tumor neoangiogenesis, which increases immature incompetent vessel leakage, thereby increasing the fraction in the extracellular extravascular space (increased V_e), causing higher interstitial fluid pressure and lower flow [16, 36]. The reduced perfusion (low blood flow and volume; low D* and f, respectively) results in worse access to chemotherapeutic drugs and oxygen for radiosensitivity, and is associated with an adverse outcome. This reduced perfusion was found in larger, more advanced stage tumors, and is indicative for low microvessel density, low velocity and hypoxia, due to the incompetent microvessels and increased interstitial pressure [16, 18, 37, 38]. A high/or increased metabolism was also associated with adverse outcome, which might be due to a high/increased glucose demand of advanced staged tumors [39], due to proliferating malignant cells and stromal tissue. In contrast, reduced metabolism in the tumoral center due to diminished access of nutrition and oxygen supply, leading to necrosis with hypoxia, was also associated with adverse outcome [40]. These tumor characteristics might be used to target subvolumes for dose-paint RT [2, 3, 15].

In the present study, the combination of HPV status, tumor volume (ADC_GTV), high K^trans, and high V_e showed more predictive potential for locoregional control than the clinically used risk stratification per T-stage. The more of these adverse factors, the worse the locoregional-free survival was. The previously described predictive value of K^trans [21, 23] and V_e was confirmed in this study [22, 23]. In contrast, Ng et al [41] found K_ep, SUV_max, TLG, and LNM V_e, and ADC_mean as predictive parameters for LRFS. However, their chemotherapy scheme was uncommon, delineation was performed on the single-slice largest diameter, and HPV status was not assessed. In this study, ADC_mean was not found to be predictive for LRFS, which was also confirmed by King et al [19]. However, in small studies (N = 17 patients, [42], N = 40, [43], N = 32, [18]) with single modality predictive assessment, a low ADC_mean was found predictive for LRFS. Furthermore, we found that ¹⁸F-FDG-PET/CT parameters did not remain predictive when combining modalities, which was in line with Ng et al [22]. However, in single modality studies [9, 23], SUV parameters were found predictive for LRFS. Such discrepancies in predictive value may be explained by factors such as sample sizes, treatment protocols, and multivariable Cox regression analysis with or without inclusion of important clinical parameters [22].

The combination of all modalities showed that N-stage, HPV-negative status, hypopharyngeal PT location, and intoxication were risk factors for adverse overall survival. This was in line with other studies who found hypopharyngeal PT location [44], alcohol use [23], and HPV status [24] as predictors.

Besides, a large tumor volume (ADC_GTV, DCE_GTV, MATV), K^trans, and V_e (as were predictive for adverse LRFS), also high ADC_mean, D*, SUV_max, and TLG, were predictive for adverse OS. The more of these adverse factors, the worse the overall survival was. Previous studies performing multivariable analysis were in line with these findings and found that V_e [22] and K^trans were predictive for OS. The pretreatment finding of low ADC_mean was associated with highly cellular tumors including rapidly dividing cells, which are more sensitive to subsequent chemotherapy and radiotherapy and therefore associated with a more favorable prognosis [18, 19]. A possible explanation for the extra predictors for OS compared with the predictors of LRFS is that certain tumor tissue architecture, e.g., heterogeneous tissue with low diffusion restriction (high ADC_mean) and aggressive high metabolism (high SUV_max and TLG), is less sensitive to (chemo)radiotherapy, which additionally decreases tumor control and survival.

In previous studies, smoking, K^trans, K_ep [23, 41], and a heterogeneity ¹⁸F-FDG-PET/CT parameter (18F-FDG-PET uniformity) were reported to be predictive for OS. In contrast, the significant predictive value of the various clinical parameters combined with significant parameters from the whole spectrum of DWI with IVIM, DCE, and ¹⁸F-FDG-PET/CT was not described previously. The aforementioned risk factors for adverse OS were found significantly predictive, whereas stratification per T-stage was not significantly predictive. This implies an additional predictive value of functional imaging to clinical staging based on morphology.

In order to improve prediction accuracy, the complementary value of each imaging modality is of importance to capture the whole spectrum of predictive tumoral characteristics, such as tumoral cellularity, necrosis, vascularity, and metabolism [11‐13, 16, 24]. Previously described hypothetical overlapping parameters, such as DWI, IVIM (e.g., ADC_mean and D, both capturing tissue cellularity indirectly), and DCE (e.g., K^trans with K_ep) [11, 12, 45], correlated not evidently in this study. Different heterogeneous tumor architecture (inflammation, fibrosis, necrosis, and hypoxia) and/or HPV status in more advanced staged tumors might have caused the loss of correlations. Further optimization of protocols and selection and evaluation of qualitative and quantitative parameters is necessary in future studies.

The current study underlines the superiority of combining MRI and ¹⁸F-FDG-PET/CT, which allowed combining significant predictive clinical parameters, such as HPV-negative status, intoxication (smoking/alcohol), hypopharyngeal tumor location, and N-stage with predictive quantitative imaging parameters: ADC_mean (DWI) and D* (IVIM), K^trans, V_e (DCE) and SUV_max, and TLG (¹⁸F-FDG-PET/CT) for OS. In this way, risk stratification on a patient level was shown to be possible. Furthermore, this might improve patient care and pave the road for personalized treatment options by identification and targeting tumoral subvolumes which are predictive for adverse outcome [7, 46].

There was a relatively low incidence of events in our cohort; therefore, this study should be considered as hypothesis-generating. Also, selection bias might have occurred by excluding surgical treatment at the prospective selection of patients with curative (chemo)radiotherapy.

Secondly, although T-stage is dependent on the gross tumor volume, they were both included in the predictive analysis, which might have caused confounding bias. Although in this study the GTV was determined on functional imaging maps (ADC and DCE maps), it should be evaluated in future studies whether GTV determined on anatomical MRI sequences is more accurate in the predictive analyses.

Thirdly, we performed pharmacokinetic modeling analysis by using a patient-specific AIF, measured in the external carotid arteries. Flow artifacts as well as high concentrations of contrast agent can result in incorrect amplitude of the arterial concentration. This can affect final calculation of K^trans and V_e which, as a consequence, are over-estimated (e.g., V_e can be larger than 1). We have decided to leave the results as we obtained them, but in future studies, we intend to correct the AIF for flow artifacts.

Finally, the LNM parameters were based on the ROIs of the largest lymph node metastasis, which might falsely ignore the adverse effect of having multiple metastases and consisting of necrotic tumoral areas, which reduced the average tumoral FDG uptake. The LNM parameters were excluded in the multimodality CoxBoost analysis in order to remain statistically robust, which might have limited the predictive value. Moreover, only internal validation by bootstrap cross-validation was feasible. These limitations were managed by performing a well setup internal validation by bootstrap cross-validation to test limited parameters repeatedly in a subset.