Predictive value of quantitative diffusion-weighted imaging and 18-F-FDG-PET in head and neck squamous cell carcinoma treated by (chemo)radiotherapy

Background and purpose: In head and neck squamous cell carcinoma (HNSCC) (chemo)radiotherapy is increasingly used to preserve organ functionality. The purpose of this study was to identify predictive pretreatment DWIand 18F-FDG-PET/CT-parameters for treatment failure (TF), locoregional recurrence (LR) and death in HNSCC patients treated by (chemo)radiotherapy. Materials and methods: We retrospectively included 134 histologically proven HNSCC patients treated with (chemo)radiotherapy between 2012-2017. In 58 patients pre-treatment DWI and 18F-FDG-PET/CT were performed, in 31 patients DWI only and in 45 patients 18F-FDG-PET/CT only. Primary tumor (PT) and largest lymph node (LN) metastasis were quantitatively assessed for TF, LR and death. Multivariate analysis was performed for F-FDG-PET/CT and DWI separately and thereafter combined. In patients with both imaging modalities, positive and negative predictive value in TF and differences in LR and death, were assessed. Results: Mean follow-up was 25.6 months (interquartile-range; 14.0–37.1 months). Predictors of treatment failure, corrected for TNM-stage and HPV-status, were SUVmax-PT, ADCmax-PT, total lesion glycolysis (TLG-LN), ADCp20-LN (P= 0.049, P=0.024, P= 0.031, P= 0.047, respectively). TLG-PT was predictive for LR (P= 0.003). Metabolic active tumor volume (MATV-PT) (P= 0.003), ADCGTV-PT (P < 0.001), ADCSD (P= 0.048) were significant predictors for death. In patients with both imaging modalities SUVmax-PT remained predictive for treatment failure (P=0.049), TLG-LN for LR (P=0.003) and ADCGTV-PT for death (P < 0.001). Higher predictive value for treatment failure was found for the combination of SUVmax-PT and ADCmax-PT, compared to either one separately. Conclusion: Both DWIand F-FDG-PET/CT-parameters appear to have predictive value for treatment failure, locoregional recurrence and death. Combining SUVmax-PT and ADCmax-PT resulted in better prediction of treatment failure compared to single parameter assessment.


Introduction
Head and neck squamous cell carcinoma (HNSCC) is the most common head and neck malignancy with 550,000 cases annually worldwide [1]. Most patients presenting with locally advanced disease are treated with (chemo)radiotherapy in the context of organ preservation. Complete response cannot be achieved in approximately outcome may be uncertain [2]. It is essential to identify tumor characteristics predictive of response to (chemo)radiotherapy to increase treatment efficacy and long-term outcome by treatment intensification or by offering alternative treatment options as early as possible to these patients (e.g. surgery or best supportive care).
Tumor characteristics associated with treatment failure or adverse long-term survival, might be identified with functional imaging [5,7,15,16]. Qualitative assessment is inadequate to measure tumor heterogeneity and often carries a certain extent of subjectivity, resulting in high interobserver variability. Quantitative analysis could provide more reproducible information of tumor heterogeneity by utilizing whole lesion assessment [17,18].
Earlier studies showed that (combinations of) DWI and 18 F-FDG-PET/CT predict treatment response [15] and long-term survival [5]. However, most of these studies included only small patient populations without correction for TNM-stage, HPV-status, tumor volume or location. Furthermore, they used different image acquisition systems, which resulted in limited robustness of predictive and prognostic value [5,16,[19][20][21]. Although some studies use histogram analysis, the predictive accuracy of the combination of DWI and 18 F-FDG-PET/CT for primary tumor and lymph node metastasis was not investigated [20,21]. Finally, quantification of based on robust whole-lesion delineation by multiple independent observers, including partial volume correction [22], was not performed previously [5,16,20,21]. The purpose of this study was to identify predictive pretreatment DWI-and 18 F-FDG-PET/CT-parameters for treatment failure, locoregional recurrence and death in a large cohort of HNSCC-patients treated with (chemo) radiotherapy.

Patients
We retrospectively enrolled 134 consecutive patients treated from 2012-2017. The local ethics committee waived informed consent. This study was reported according to the STARD-criteria [23]. Inclusion criteria were: histopathologically proven HNSCC of primary tumor with or without lymph node metastasis; pretreatment DWI and/or 18 F-FDG-PET/CT and planned (chemo)radiotherapy. Exclusion criteria were: previous locoregional treatment for HNSCC and insufficient image quality for lesion segmentation.
Treatment failure was defined as: presence of residual malignant tissue within the first 6 months post-treatment. For the assessment of treatment failure we used a standardized physical examination and imaging with MRI and PET-CT which was performed between 3-6 months after treatment.
Locoregional recurrence (LR) was defined as presence of a locoregional (primary tumor ( PT ) or lymph node ( LN )) recurrence in the total follow-up period after end of treatment. In case of suspected malignancy on imaging, histological confirmation was acquired.
Death was defined as the time patients were alive after end of treatment. Patients were censored at the date of the last follow-up for surviving patients.

MRI acquisitions
MR imaging was performed using a 1.5 T MR system (Signa HDxt, GE Healthcare, Milwaukee, WI) with a 12-channel neurovascular headand-neck-array coil. The protocol consisted of at least conventional T1weighted Spin-Echo (SE), short-tau-inversion-recovery (STIR) and echoplanar imaging (EPI)-DWI with TR/TE: 4300-5600/59-98 ms; inversion time 160 ms, averages: 3; parallel imaging acceleration factor: 2; acquisition time ranged between 235 and 314 s; flip-angle: 90°; matrix size: 256 × 256 (which was not directly acquired, but zero filled from an acquisition matrix of 86 × 128), voxel size 1 × 1 x 4mm 3 , 21-28 slices. ADC was measured using 2 b-values (0 and 1000s/mm2). The ADC values of patients with a TE of 59 and 98 ms were not significantly different (p = 0.46), and therefore both patient groups were combined in the analysis. 18 F-FDG-PET/low-dose-CT was performed according the EANM guidelines 2.0 on a Gemini TOF-64 PET/CT (Philips Medical Systems, Best, The Netherlands) with EARL accreditation [24]. Low-dose noncontrast CT (120 kV; 50 mAs) was used for attenuation correction and anatomic correlation of 18 F-FDG uptake. Whole-body 18 F-FDG-PET/CT was performed in arms-down-position, with 18 cm axial field of view, from mid-thigh to skull vertex, 60 min after intravenous administration of 2.5 M Bq/kg 18 F-FDG, 2 min per bed-position. The 18 F-FDG-PET images were reconstructed using vendor-provided and EARL-compliant reconstruction protocol with photon-attenuation correction. Reconstructed images had an image matrix size of 144 × 144 and voxel size of 4 × 4 x 4 mm. Post-reconstruction image resolution was 7 mm full-width at half maximum.

Image analysis
Delineation on MRI was performed manually by two independent radiologists with 34 and 10 years of experience in head-and-neck radiology, respectively. Whole-tumor segmentation was performed, including necrotic or cystic areas and excluding voxels, which were filled with air or subject to distortion. 18 F-FDG-PET/CT semi-automated delineations ( Fig. 1), based on a background-corrected 50% of SUVpeak isocontour [25], were assessed and corrected by a resident supervised by a nuclear medicine physician with 5 and 30 years of experience in head-and-neck PET, respectively. Observers were aware of the HNSCC diagnosis, TNM-stage (7th edition) and primary tumor location for delineation of proven malignant lesions. Any discrepancies between the observers were resolved by performing a segmentation in consensus.
Whole-lesion segmentation was performed on MRI on the T1-and ADC-map in accordance with other sequences (Fig. 1) using VelocityTM software (Varian Medical Systems, Inc, Palo Alto, USA). Histograms were generated with Matlab (MathWorks Inc, MA, USA), based on included voxels. The following MRI parameters were extracted: gross tumor volume on T1 (T1 GTV ) and ADC (ADC GTV ), ADC mean , ADC max , ADC min , ADC standard deviation (ADC SD ), ADC percentiles (ADCp 10 to ADCp90; indicating the value below which a given percentage of tumor voxels in a group of voxels fall), ADC kurtosis , and ADC skewness . Consensus values were used for final analyses.
With 18 F-FDG-PET/CT, segmentation was performed using a background-corrected 50% isocontour of tumor SUV peak [26]. SUV was normalized to body weight. SUV peak was defined as the highest uptake in 1 mL spherical volume of interest across all tumor voxel locations. Uptake parameters were derived using in-house developed Accurate software, [24] to quantify lesions on 18 F-FDG-PET/CT were: metabolically active tumor volume (MATV), SUV max , SUV peak , SUV mean . Total lesion glycolysis (TLG) was calculated by the tumor volume multiplied by the SUV mean of the included voxels [17].

Statistical analysis
The correlation of functional parameters between DWI and 18 F-FDG-PET/CT was assessed using the Pearson r test. The difference and correlation of functional parameters between HPV-positive and HPV-negative patients was analyzed performing the Pearson r test. Logistic regression was used to predict locoregional treatment failure. We used Cox regression model to assess locoregional recurrence (LR) and death. P-values of < 0.05 were considered statistically significant in the univariate analysis. Significant single modality parameters were combined in multivariate analysis, for DWI and 18 F-FDG-PET/CT separately, using a backward Wald test (Fig. 2). The remaining significant parameters for DWI or 18 F-FDG-PET/CT, for tumor and for lymph node metastasis were taken together, corrected for TNM-stage and HPV status. This correction was performed by adding T-, N-and HPV-status to the multivariate analysis. Multivariable regression analysis was performed according to the TRIPOD-statement, accepting p-values up to 0.157 to enhance the model applicability to other patient groups [27]. For survival analyses, the end of treatment was used as start of follow-up. For multivariate analysis only patients were included of whom all data was available (i.e. DWI and 18 F-FDG-PET/CT). Subgroup analyses was performed for tumors with a MATV > 4.2 ml [22], to avoid bias from partial volume effect in small tumors.
With receiver operating characteristic (ROC) analysis three optimal cut-offs (three highest Youden indices (YI)) of significant multivariate parameters were determined to predict treatment failure. The positive and negative predictive value (post-test risk and 1-post-test risk, respectively) at each cut-off value was calculated for treatment failure using a prevalence ranging from 10 to 50%. A log-rank test of remaining significant multivariable parameters (divided by the optimal cut-off) was presented as Kaplan Meier survival curves. Analyses were performed using SPSS (version 18.0; SPSS Inc., Chicago, III, USA).
During the total follow-up time, in 28 patients (20.9%) a locoregional recurrence (LR) occurred. The LR rate in both DWI only and 18 F-FDG-PET/CT only group was also similar (27 versus 28 patients (20.1%, 20.9%, respectively).

Correlations of functional imaging and HPV status
The primary tumor ADCGTV and ADCkurtosis were found significantly correlated (supplement 1 and 2) with all PET-parameters (range Pearson r from 0.42-0.91, p < 0.0015 after Bonferroni's correction for multiple testing). In lymph node metastases, only ADCGTV correlated significantly with TLG (r = 0.51, p < 0.0015).
Forty-seven patients were found HPV-positive, 67 HPV-negative and in 20 patients HPV-status was unknown. In the treatment failure group; 5 patients were HPV-positive, 8 HPV-negative, 4 unknown. The LR group consisted of 8 HPV-positive, 16 HPV-negative, 4 unknown status. The patients who died, consisted of 7 HPV-positive, 30 HPV-negative, 5 unknown.

Image and univariate analysis of treatment failure
In the treatment failure group, a wider distribution (higher ADC SD ) and higher ADC of voxels (higher ADC max ) in PT was shown (Fig. 3A).
Univariate DWI (ADC max-PT and ADCSD-PT) and all primary tumor 18 F-FDG-PET/CT parameters were higher in patients with treatment failure (Table 3, Fig. 4). However, subgroup analysis (not tabulated) of tumors with MATV larger than 4.2 ml (to minimize partial volume effects; n = 79 patients), did not result in any significant 18 F-FDG-PET predictor.

Image and univariate analysis of LR and death
A significant higher ADC p70-PT and ADC p80-PT was shown in LR (Fig. 3B). A higher ADC SD-PT and ADC max-PT was found in patients who died (Fig. 3C).

Multivariate analysis of treatment failure
The significant univariate parameters (Table 3) predicting treatment failure, were assessed in a multivariate analysis for DWI and 18 F-FDG-PET/CT separately, and for PT and LN separately (Table 4), corrected for TNM-stage and HPV-status. Primary tumor SUV max-PT (P = 0.049) and ADC max-PT (P = 0.024), and TLG -LN (P = 0.031) and ADC p20-LN (P = 0.047) were significant predictors of treatment failure.
Finally, a head-to-head comparison was made in patients (n = 58) who underwent both imaging modalities using these remaining significant parameters (Table 4), which revealed SUV max-PT as significant predictive factor (P = 0.042) in multivariate analysis with a hazard ratio of 1.302 (95%CI 1.010-1.679). Subgroup analysis of tumors with a MATV larger than 4.2 ml (PT n = 48, LN n = 26), resulted in the same predictive parameter (not tabulated).
In order to determine the additional value of predicting treatment failure using a combination of both modalities, the remaining PT and LN significant multivariate parameters of both cohorts (ADC max of DWIcohort and SUV max of 18 F-FDG-PET/CT-cohort) were analysed in patients with MATV > 4.2 ml in order to minimize partial volume effects (Table 5). First, optimal cut-offs for SUV max-PT were 7.13, 11.3, 13.58; and for ADC max were 1.927, 2.236, 2.528, respectively. The positive predictive value (PPV) and negative predictive value (NPV) were calculated for ADC max and SUV max at the cut-off values described above, showing additional value of combining ADC max and SUV max . For each assumed prevalence, the PPV for treatment failure was the highest when both a high ADC max (> 2.528) and high SUV max (> 13.58) were assessed. Regardless of the ADC max , a SUV max < 7.13 ruled out treatment failure. For each prevalence, the NPV for treatment failure increased as lower SUV max cut-off and ADC max cut-off value was used. The combination of significant multivariate nodal parameters ADC p20-LN and TLG -LN resulted in a slightly better prediction of treatment failure (Supplement 3).

Multivariate analysis of LR and death
The significant univariate parameters for LR prediction, were assessed per modality in a multivariate analysis, corrected for TNM-stage and HPV-status. This resulted in only TLG-PT as significant (P = 0.003) prognostic 18 F-FDG-PET/CT parameter or LR (single modality and modalities together, Table 4). TLG -PT remained significant Subgroup analysis (MATV > 4.2 ml, not tabulated) resulted in TLG -PT (P = 0.039) as predictive parameter for LR. Performing the ROC analysis, an optimal cut-off value for TLG -PT of 36.2 was found which use was significant predictive for LR (Fig. 5A).
The univariate parameters for prognosis of death of 18 F-FDG-PET/ CT (MATV-PT , SUV max -PT , SUV peak -PT , SUV mean -PT and TLG PT ) combined in a multivariate analysis, resulted in MATV -PT (P = 0.003) as remaining significant prognosticator (single modality, Table 4).
In univariate significant DWI parameters for death (ADC GTV -PT , ADC SD-PT and ADC max-PT ) were combined in a multivariate analysis resulted in ADC GTV-PT (P < 0.001) and ADC SD-PT (P = 0.048) as significant prognosticators (single modality, Table 4). ADC GTV-PT (P = 0.009) remained a significant predictor in subgroup analysis (MATV > 4.2 ml) who underwent both imaging modalities (not tabulated).
The combination of significant prognosticators to predict death resulted in only ADC GTV -PT as significant prognosticator (Modalities together, Table 4; P < 0.001). The use of both MATV -PT and ADC GTV-PT (Fig. 5C) resulted in a similar overall survival curve as for single parameter assessment (Fig. 5B), which showed no additional value for the use of both modalities for the prediction of death.

Discussion
Pretreatment DWI and 18 F-FDG-PET/CT were evaluated for their predictive value for treatment failure, locoregional recurrence-free survival and overall survival. The assessment of both SUV max-PT and ADC max-PT showed additional value for treatment failure prediction, compared with single parameter assessment. Furthermore, a high TLG -PT was predictive for locoregional recurrence, and high MATV -PT , ADC GTV-PT , and ADC SD-PT were predictive for death.
In order to identify tumor characteristics with quantitative analyses, whole-lesion delineation could capture tumor heterogeneity and ignores subjective exclusion of necrosis or other potential predictive characteristics [20,28,29]. This characteristic and possible predictive heterogeneity may be caused by areas with high cellularity, necrosis, stroma and areas with increased or decreased vascularity. The mean value of an imaging parameter may be sub-optimal, because when areas with low and high ADC values are included in the ROI, heterogeneity is flattened out.
A wide distribution, as sign of lesion heterogeneity, might be reflected by the ADC SD [30]. Furthermore, low-cellular tumor parts (e.g. necrosis) might be measured with high ADC values (high ADC max or ADC-percentiles), while high-cellular solid viable areas are reflected by low ADC values. In our study primary tumor ADC SD and ADC max parameters were higher in non-responders, which might be due to limited efficacy of radiotherapy in structures with a low cellularity (e.g. low diffusion restriction such as necrosis and fibrosis) [31]. Moreover, ** Locoregional recurrences within the total follow-up time after end of treatment.

Table 2
Differences and correlation of functional parameters divided per HPV-status in PT and LN. Most functional parameters were significant higher (marked green) in HPV negative patients. SD: Standard deviation. *: Difference between HPV-positive and HPV-negative tumours (Mann-Whitney U test). **: Correlation between a high functional parameter value and a positive HPV-status (Pearson r test). Fig. 3. Histogram of ADC voxels of primary tumors (PT) and lymph node metastases (LN) with adverse (blue) and good outcome (orange). A) In treatment failures (TF), wider distribution (ADC SD ) and higher ADC max values was found compared to the responders. B) Locoregional recurrences showed a significant higher ADC max-PT . C) Patients who died showed a wider distribution and higher ADC max-PT . D) TF included more of the lowest ADC -LN voxels. E) LR showed a higher LN peak and a less wide distribution in the control group. F) The ADC -LN parameters showed no significant predictors for death. in this study primary tumor ADC GTV and ADC kurtosis significantly correlated with all FDG-PET parameters, which implies a volume dependency in all FDG-PET parameters as well as ADC kurtosis as a possible surrogate marker for FDG uptake. Last, lymph node ADC p20-LN was found higher in non-responders, which might be caused by necrotic parts, insufficient vascularization and keratin protein contents [19,32].
HPV-negative tumors were correlated with high ADC-values (r = 0.452). This is in line with some previous studies [28,[33][34][35], in which higher ADC histogram parameters in HPV-negative patients were found. This was confirmed by Meyer et al. [36], who found a negative correlation of high ADC max and ADC SD with low P53 expression in HPVnegative tumors, which lead to minimal cell cycle arrest, senescence or apoptosis [37]. This was attributed to a more keratinizing morphology with variable cellularity, cell shape, keratin pearls, large intratumoral necrosis, hemorrhage and stromal cells found in HPV-negative tumors [28,31,33]. However, 2 studies did not find any differences of ADC histogram parameters between p16 status [36] or between the more accurate HPV status [15]. This suggested independent prognostic value of ADC and HPV status. In contrast, HPV-positive tumors have a typical non-keratinizing morphology, with small central necrosis and large Table 3 Univariate parameters of FDG-PET and DWI separately for predicting treatment failure, locoregional failure and death. The odds ratio for treatment failure and hazard ratios for locoregional recurrence and death is shown. The significant univariate parameters were assessed in a multivariate analysis (Table 4). 1,2,3,4 : T-stage Color green: Significant univariate parameter S.E: Standard error *: ADC-percentiles. amounts of infiltrating lymphocytes, low stromal volume (i.e low ADC), sufficient vascularization and, therefore, more responsive to (chemo) radiotherapy [15,28,31]. Hypoxia may inhibit dividing tumor cells or lead to cell death (increased ADC), due to insufficient oxygen supply, but it may directly induce tumoral progression (high FDG-uptake with reduced apoptosis, high production of VEGF, glycolic enzymes and signaling molecules [12,38]) with radioresistance. This however, was dependent on tumor grading with different relations between parenchyma, stroma and microvascular density [14,39]. Surov et al. [14] described that SUV max and SUV mean were correlated with cellularity in advanced stage tumours (G3) and only HIF-1α tended to correlate with FDG-uptake. HPVnegative tumours exhibit increased HIF1α-induced glucose metabolism as evidenced by increased glycolysis and proliferation (e.g. Ki 67), production of lactate and hypoxic microenvironment [9,14]. In contrast, HPV-positive cells effectively utilize mitochondrial respiration as evidenced by increased oxygen consumption [40][41][42] and low expression of HIF-1α [36].We confirmed this hypothesis by finding significant higher 18 F-FDG-PET parameters (MATV -PT , SUV max-PT , SUV peak-PT , SUV mean-PT and TLG-PT in univariate analysis) in HPV-negative than HPV-positive patients. This underlines the necessity for correction for HPV status in multivariate prognostic analysis.
The prediction of treatment failure was previously described performing single DWI [16,20,43,44] or FDG-PET [4,11,45]. Nakajo et al. [5] compared both imaging techniques and found a similar predictive potential for ADC as for SUV predicting LRF in a small sample size. Limited data on the combination of imaging techniques was published. Kim et al. [46] found that volumetric MATV, TLG corrected by tumor cellularity (ADC min and ADC mean ) and were prognostic for LR [46]. However, their patient population underwent surgery, and sub-optimal delineation of only one imaging technique was performed. Although in advanced tumors SUV max /ADC min , SUV max /ADC mean and TLG/ADC min correlated with cellularity, Ki67 and HIF-1a [47], no prognostic tests were performed. Chan et al. [21] found that a combination of PET heterogeneity (uniformity), DCE-MRI parameters and clinical risk factors was most predictive for LR and OS. However, in their study DWI was not applicable in the prognostic tests and HPV-status was lacking.
To our knowledge, the prediction of treatment failure and locoregional recurrence was not assessed before by combining PET and DWI together (including histogram parameters capturing tumor heterogeneity and clinical parameters including HPV-status) in a multivariate prognostic analysis and presented with a clinical applicable prediction model (Table 5). We showed that the combination of ADC max and SUV max was more accurate than using a single imaging technique for predicting treatment failure (i.e. during the first 6 months after treatment). The prediction of locoregional recurrence showed TLG-PT as main prognostic parameter. This was in line with other studies, in which identification of necrosis or hypoxia and a higher glycolytic and increased tumor heterogeneity in HPV-negative tumors was predictive for treatment failure [16,20]. TLG has been suggested to reflect global metabolic activity in whole tumors better than MATV, as TLG represents both functional tumor burden and biological aggressiveness [48]. Regarding death, in this study only volumetric parameters (high MATV -PT and ADC GTV-PT ) were found to be significantly predictive for death. This was confirmed in previous studies. However, high SUV mean was also described as prognostic factor [4,49].
The identification of predictive tumor characteristics predictive of resistance to (chemo)radiotherapy, such as a high intratumoral heterogeneity (ADC SD , ADC max , ADC -P20-LN ), tumor aggressiveness (SUV max ) and a negative HPV-status, could help stratify risk groups, which might implicate tailored treatment (e.g., intensifying concurrent chemotherapy or offering alternative treatment options such as surgery as early possible to minimalize the side effects of ineffective chemoradiation). Patients with high TLG -PT and ADC GTV-PT may be considered at high risk of locoregional recurrence and may benefit from intensifying post-treatment monitoring.
Our findings suggested that treatment failure prediction benefits from combining DWI with 18 F-FDG-PET/CT imaging, by measuring the ADC max-PT and SUV max-PT (Table 5). However, the drawback of performing both imaging must be considered, e.g. a heavier burden for the patient and higher costs. Optimal cut-off values of ADC max-PT and SUV max-PT showed their diagnostic potential to considerably improve the prediction of treatment failure. In contrast, no additional value of combining imaging modalities was found for predicting LR and death (Fig. 5).
Some limitations must be acknowledged. First, 18 F-FDG-PET/CT delineation bias could have occurred using a 50% SUV peak threshold, which excludes necrotic areas with low SUV [15]. This results in overall higher SUV and lower MATV, although it is more reproducible. Secondly, the maximum and minimum values of DWI and SUV parameters were single voxel dependent and therefore susceptible to noise. In this study, readers excluded voxels in areas of distortion or air, whereas the rest was considered characteristic for tumor phenotype. Furthermore, in this study the percentiles ranging from 10 to 90 were assessed, which are less effected by extreme values. Thirdly, the assessment of the largest LNs only could falsely ignore the prognostic adverse effect of Table 4 Multivariate analysis of significant univariate parameters (marked green in Table 3) assessed per DWI and PET separately, corrected for TNM-stage and HPV-status. Hereafter, these significant multivariate parameters of the single modality section are combined in the 'modalities together' section to predict treatment failure, locoregional recurrence and death.
Abbreviations: NS: Not significant CI: confidence interval PET*: The lymph node metastasis analysis consisted of 48 patients.

Conclusion
Both DWI and 18 F-FDG PET/CT parameters appear to have predictive value for treatment failure, locoregional recurrence and death. Combining SUV max-PT with ADC max-PT improved treatment failure prediction compared to single parameter assessment. In contrast, no additional value of combining imaging modalities was found for predicting locoregional recurrence and death. Stratification of patients with DWI and 18 F-FDG PET/CT parameters to predict favorable and unfavorable outcomes might help to tailor patient's individualized treatment.

Role of the funding source
This work was (partly) supported by grant 10-10400-98-14002 from the Netherlands Organisation for Health Research and Development.

Table 5
Positive (A) and negative (B) predictive value for treatment failure using both ADC max (horizontal) and SUV max (vertical) with the most optimal cut-off values for each prevalence. A red colour marks a high predictive value for treatment failure.
*The prevalence of treatment failure in HNSCC in patients [1].