Predictive potential of dynamic contrast-enhanced MRI and plasma-derived angiogenic factors for response to concurrent chemoradiotherapy in human papillomavirus-negative oropharyngeal cancer

We consecutively included patients with previously untreated locally advanced HPV-negative OPSCC who were scheduled for curative cCRT. Patients were treated with the intensity-modulated radiotherapy (RT) technique to a radiation dose of 70 Gy in 35 fractions over 7 weeks. Cisplatin 40 mg/m2/week IV or carboplatin 1.5 AUC/week were administered concurrently with RT. HPV status was determined by p16 immunohistochemistry and in situ hybridization studies. Patients with concurrent malignancies and those unable to undergo MRI examination were excluded from the study.

Patients underwent DCE-MRI before treatment initiation, after receiving 20 Gy of irradiation, and 10–12 weeks after cCRT completion. Peripheral venous blood samples were obtained at each MRI timepoint.

MRI scans were conducted using a 3.0T MAGNETOM Trio, A Tim System scanner (Siemens Healthineers®, Forchheim, Germany), equipped with a head and neck receive coil. The diagnostic imaging protocol included an axial T2-weighted sequence with short tau inversion recovery (STIR) (TR/TE 5010/71 ms, TI 170 ms, matrix size 256 x 256, slice thickness 3 mm, 10% gap, and field of view (FOV) 18 x 18 cm) and a contrast-enhanced (CE) axial T1-weighted volumetric interpolated breath-hold examination (VIBE) sequence (TR/TE 3.26/1.26 ms, voxel size 1.1 x 0.9 x 1.5 mm, matrix size 288 x 288, and FOV 250 cm²). DCE-MRI was performed using a 3D fast low-angle shot (FLASH) sequence optimized for spatial and temporal resolution (TR/TE 5/1.16 ms, matrix size 220 cm², slice thickness 4 mm, temporal resolution 9 s, total acquisition time 6 min). T1 mapping was utilized to convert signal intensities into gadolinium concentration. The T1 map was derived from two pre-contrast flip angle images (6° and 15°). Baseline images were acquired for 27 seconds. Intravenous administration of the paramagnetic contrast medium gadobutrol (Gadovist®, Bayer HealthCare Pharmaceuticals) was performed at a dose of 0.1 mmol/kg body weight with a flow rate of 3.5 mL/s, followed by a 20 mL saline flush. The diagnostic imaging protocol (axial STIR and CE axial T1 VIBE) was employed for pre-treatment tumour evaluation and radiation treatment planning.

Patients were stratified into two groups based on clinical evaluation. Responders demonstrated complete disappearance of the primary tumour within 10–12 weeks following cCRT. Non-responders encompassed all other patients, including those with partial response, stable disease, or progressive disease.

We conducted the post-processing of DCE 3D datasets using a commercially available Tissue 4D workflow within the syngo.via medical imaging software (Siemens Healthineers, Forchheim, Germany). We utilized a population-based arterial input function with an intermediate pre-set, and motion correction was automatically conducted using the software algorithm.

Regions of interest (ROI) delineation was verified by two experienced head and neck radiologists (A. L., K. Š. P.). ROI were delineated across VIBE multiple images, resulting in the generation of volumes of interest (VOI). VOI delineation was conducted twice for each examination. Initially, the entire tumour volume was delineated, excluding larger vessels and necrotic regions. These were assessed visually based on their appearance and enhancement characteristics, with large necrotic areas showing little to no enhancement. Subsequently, a smaller tumour volume was delineated, characterized by homogeneous or near-homogeneous enhancement, with a substantial margin from surrounding heterogeneities. The Tofts two-compartment pharmacokinetic model was utilised to derive a set of perfusion parameters, which included the volume transfer constant between blood plasma and extracellular extravascular space (k_trans), the total extracellular extravascular space (EES) volume fraction (V_e), the reflux constant (k_ep) and the initial area under the curve (iAUC) for the selected VOI. The software generated parameter maps depicted as colour-coded images. The mean values of perfusion parameters from the two delineated VOIs were used for analysis.

Blood samples underwent processing to yield 0.5 ml plasma samples, which were immediately stored in the Protein LoBind® Tubes (Eppendorf, Hamburg, Germany) at −80°C. Plasma concentrations of VEGF, CTGF, PDGF-AB, ANG, END, and THBS1 were measured at selected time-points using the enzyme-linked immunosorbent assay (ELISA) technique. The following ELISA Kits were used for the analysis: Human VEGF ELISA Kit (KHG0111), Human PDGF-AB ELISA Kit (EHPDGFAB), Human Angiogenin ELISA Kit (EHANG), Human Endostatin ELISA Kit (EHCOL18A1), Human Thrombospondin 1 ELISA Kit (EHTHBS1), all Invitrogen (Life Technologies Corporation, Carlsbad, CA, USA); and Human CTGF Mini ABTS ELISA Development Kit Catalog# 900-M317 Range (PEPROTech, EC Ltd., London, VB, part of Thermo Fisher Scientific Inc., Waltham, MA, USA).

All steps were performed following the manufacturers' protocols, and each assay was conducted in duplicate or triplicate to ensure accuracy and reliability. To determine the levels of ANG, END, PDGF-AB, and THBS1, the plasma samples underwent dilution at ratios of 8000-, 100-, 2-, and 400-fold using 1x Assay Diluent A or D. The absorbance measurements were conducted using Synergy H4 (Biotek) at 450 or 405 nm with wave-length correction at 650 nm. Standard curves were calculated in GraphPad Prism V. 10 for Windows (GraphPad Software, Boston, MA, USA, www.graphpad.com) using curve fitting analyses, followed by the calculation of results for plasma samples. Sample concentrations were expressed in ng/mL or pg/mL.

Shapiro-Wilk tests were used to assess the normality of data distributions. Depending on the results, either Student's t-tests for independent samples or Mann-Whitney U tests were employed to investigate disparities in DCE-MRI perfusion parameters and plasma levels of angiogenic factors between responders and non-responders. P-value below 0.05 was considered significant, while a p-value between 0.05 and 0.10 was regarded as indicative of a trend. Spearman correlation analysis was performed to analyse correlations between DCE-MRI perfusion parameters and biochemical parameters. Predictive models for treatment outcomes were developed using logistic regression analysis and a classification tree model. These models were validated using the leave-one-out strategy. The data analysis was carried out using Python extension libraries, including Orange¹⁰, Pandas¹¹, NumPy¹¹, Matplotlib¹², and Scipy¹³.

Twenty-five patients met the inclusion criteria. Initial DCE-MRI was performed on average 11 days (SD 4.2; range 3–18) prior to cCRT initiation, and the intra-treatment DCE-MRI was performed after receiving an average of 21 Gy (SD 2.9; range 18–34) of irradiation. Venous samples were collected an average of 11 days (SD 6.4; range 2–19) prior to cCRT initiation, and the intra-treatment samples were obtained after receiving an average of 23 Gy (SD 5.7; range 18–40) of irradiation. Patient and tumour characteristics are summarised in Table 1.

Significant discrepancies were noted in the DCE-MRI parameter values for one patient, indicating a possible technical issue. As a result, we decided to exclude this patient with stage III OPSCC, leaving us with 24 patients for analysis. Additionally, there was missing data in the biochemical analysis, with two patients failing to attend the follow-up visit at 10–12-week post-cCRT. Furthermore, some data were missing due to values falling below the detection limit, including 8 samples for ANG and 1 sample for THBS1.

The median follow-up for 24 patients was 30 months (range 17–53 months). At the last follow-up visit, 17 patients were alive, 3 of them with active malignant disease. 7 patients died, with index cancer and infection being the cause of death in 5 and 2 patients, respectively. The latter 2 patients were free of treated cancer at the time of death. Among the participants, 15 were classified as responders (complete response), whereas 9 were categorised as non-responders (incomplete response). All non-responders experienced disease progression during the follow-up period. Of these, one individual was eligible for salvage surgery and remained recurrence-free throughout the follow-up period, while two others developed distant metastases, involving the lung and skin, respectively. Conversely, among the responders, one patient experienced local recurrence and subsequently underwent surgery, while another developed distant metastasis involving the bone and lung. The remaining 13 responders remained disease-free throughout the follow-up period.

Spearman correlation analysis revealed no strong correlations between DCE-MRI perfusion parameters and biochemical parameters. However, moderate positive correlations were observed between k_ep and both VEGF and ANG before treatment, as well as between VEGF and both k_trans and iAUC after receiving 20 Gy of RT. On the other hand, moderate negative correlation was observed between k_trans and THBS1 and also between iAUC and CTGF after receiving 20 Gy of RT. The remaining parameters exhibited weak correlations, including k_trans and VEGF before treatment (Spearman rho = 0,166) (Table 2, Figure 1).

Figure 1.

A correlation heatmap visualising perfusion parameters alongside plasma concentrations of angiogenic factors before concurrent chemoradiotherapy (cCRT) initiation (left) and after receiving 20 Gy radiotherapy (right).

Representative co-registered anatomic MR images and colour-coded pharmacokinetic parametric maps of a responder and a non-responder patient before and after 20 Gy RT are shown in Figures 2 and 3, respectively.

Figure 2.

Co-registered volumes of interest (VOI) and color-coded k_trans maps, together with concentration curves alongside corresponding k_trans values before treatment (upper section) and intra-treatment after receiving 20 Gy radiotherapy (bottom section) in a representative responder patient.

Figure 3.

Co-registered volumes of interest (VOI) and color-coded k_trans maps, together with concentration curves alongside corresponding k_trans values before treatment (upper section) and intra-treatment after receiving 20 Gy radiotherapy (bottom section) in a representative non-responder patient.

Compared to non-responders, responders exhibited higher k_trans values before treatment (0.270, SD 0.087 min-1 vs. 0.169, SD 0.062 min-1, p = 0.006) as well as after receiving 20 Gy of RT (0.289, SD 0.067 min-1 vs. 0.215, SD 0.027 min-1, p = 0.007) (Table 3). After receiving 20 Gy of RT, an increase in k_trans was observed in both groups. iAUC was larger in responders compared to non-responders before treatment (0.334, SD 0.109 vs. 0.220, SD 0.092, p = 0.015) but not after receiving 20 Gy of RT (0.361, SD 0.070 vs. 0.336, SD 0.099, p = 0.468). Before treatment, responders exhibited a larger V_e compared to non-responders (0.306, SD 0.082 vs. 0.237, SD 0.108, p = 0.092); however, this difference was only observed as a trend. After receiving 20 Gy of RT, there was an increase in V_e in all patients, but the fractional increase in V_e between responders and non-responders was insignificant.

Responders exhibited a trend of higher k_ep values compared to non-responders after receiving 20 Gy of RT (0.819, SD 0.229 vs. 0.709, SD 0.086, p = 0.065). Before treatment, k_ep values did not distinguish between responders and non-responders.

Plasma levels of VEGF were consistently higher in responders compared to non-responders across all three time points – before treatment (p = 0.011), after receiving 20 Gy of RT (p < 0.001) as well as 10–12 weeks following the completion of cCRT (p = 0.010) (Table 4). There were no significant differences observed in the dynamics of plasma VEGF levels between the two groups (p = 0.462).

Plasma levels of PDGF-AB showed a decreasing trend across the three time points for all patients, with no significant difference observed in the dynamics between responders and non-responders. After receiving 20 Gy of RT, plasma levels of PDGF-AB were higher in responders compared to non-responders (p = 0.021), whereas no significant difference was observed between the two groups before treatment (p = 0.431) and 10–12 weeks after CRT completion (p = 0.919).

Apart from several values falling below the detection limit, there was considerable variation in the measured levels of plasma ANG among all patients. No notable dynamics or significant differences were observed between responders and non-responders. Plasma levels of THBS1 remained stable across the three time points, with no discernible differences observed between responders and non-responders. An insignificant increase in plasma levels of END was observed across the three time points in all patients, with no noticeable differences between responders and non-responders. Plasma levels of CTGF displayed considerable variability, with no notable differences observed between responders and non-responders (Table 4).

Based on Spearman correlation analysis and descriptive statistics for DCE-MRI and biochemical parameters between responders and non-responders, we selected two variables for inclusion in statistical modelling to predict treatment outcomes, namely k_trans and VEGF as measured after receiving 20 Gy of RT. The multivariate logistic regression (LR) analysis, which incorporated two predictor variables, k_trans and plasma VEGF after receiving 20 Gy of RT, yielded a classification accuracy (CA) of 0.708 and area under the ROC curve (AUC) of 0.859. Notably, the variable VEGF achieved statistical significance (p = 0.033), whereas the variable k_trans approached significance (p = 0.056).

A univariate LR analysis utilizing only the variable k_trans, measured after receiving 20 Gy of RT, exhibited a higher CA of 0.792, but a significantly lower AUC of 0.625. Notably, the variable k_trans demonstrated statistical significance (p = 0.019).

The classification tree model, using the same two variables, achieved CA of 0.833 and AUC of 0.881. The classification tree (Figure 4) structure consists of one terminal node on the first branch and two terminal nodes on the second. Variable k_trans was found to be the most influential in determining treatment response with a cut off value of 0.259 min⁻¹, followed by variable VEGF with a cut off value of 62.5 pg/mL.

Figure 4.

Graphical presentation of the classification tree model.

cCRT = concurrent chemoradiotherapy; VEGF = vascular endothelial growth factor