Background

Nasopharyngeal carcinoma (NPC) is a common head and neck cancer, with over 133,000 newly diagnosed cases in 2020, which was approximately 0.7% of all cancers, and over 80,000 confirmed deaths reported worldwide [1]. However, NPC shows a geographically unbalanced incidence across countries, with > 70% of new cases diagnosed in Southeast Asia and East Asia [2]. The primary treatment methods for NPC include chemotherapy and radiotherapy. Radiation therapy in patients with early-stage NPC and concurrent chemoradiation or induction chemotherapy followed by concurrent chemoradiation treatment in patients with locally advanced-stage NPC have proven effective according to the latest National Comprehensive Cancer Network guidelines [3]. However, the 5-year locoregional failure rate is approximately 15–19%, and the isolated distant metastasis rate is 18% [4]. This may be because the response to treatment varies between patients; therefore, the ability to predict the treatment response can provide useful information for clinicians to adjust the treatment protocol.

Computed tomography (CT) and magnetic resonance imaging (MRI) are routinely used for detection, staging, radiotherapy planning, and patient monitoring after therapy [5]. MRI is an excellent imaging modality in the oncological medical framework. Currently, advanced chemoradiation therapy requires accurate MRI findings for characterizing, contouring, and providing quantitative functional parameters and for monitoring treatment, thereby possibly providing alternate treatment protocols for the best response to treatment [6]. Diffusion-weighted imaging (DWI) is an advanced MRI technique that measures the random Brownian motion of water molecules within the tissue voxels. The apparent diffusion coefficient (ADC) can be calculated from diffusion-weighted images after obtaining different diffusion-weighted values. Thus, DWI is a valuable tool for monitoring changes in ADC produced by tumor cell proliferation, density, and apoptosis [7] and, therefore, could be used as a pre-treatment imaging biomarker for treatment response [8]. Accordingly, the pre-treatment and treatment-related changes in ADC value have been documented as potential biomarkers for predicting treatment response in NPC and other head and neck squamous cell carcinomas [9,10,11,12,13,14], with the changes in ADC value demonstrating a trend toward effective outcome prediction comparable to RECIST criteria [13, 14]. However, the predictive accuracy and clinical utility of DWI require further validation and standardization.

NPC is characterized by tumor cells exhibiting high cellularity, cystic components, and necrotic tissues, which contribute to tumor heterogeneity. Therefore, the measured average ADC represents the overall tumor but does not account for intratumoral heterogeneity. Consequently, a method called the parametric response mapping (PRM) has been developed to address this limitation. This analysis evaluates the response by quantifying voxel-by-voxel changes based on the ADC. Previous studies have investigated the use of PRM analysis in predicting cancer treatment response and demonstrated that it provides valuable prognostic information by detecting localized changes in tumor microstructure that may not be detectable by the conventional methods, which include whole-tumor evaluation [15,16,17]. In a previous study on patients with brain tumors, PRM showed superior prognostic potential for radiological responses, time to progression, and overall survival [17].

Furthermore, PRM analysis based on changes in the ADC has shown promising results in predicting treatment responses [18]. Consequently, improving imaging techniques by obtaining a biomarker that predicts treatment response can potentially increase the individual benefit to patients, such as an early intensified treatment protocol for patients with poor response and avoiding treatment side effects in patients with good response. Therefore, in this study, we aimed to evaluate the use of ADC and PRM to predict treatment outcomes in patients with NPC.

Methods

Patients and treatment

This retrospective and prospective cohort study was approved by our institutional review board (IRB number 932/63), and written informed consent was obtained from all included patients. Patients were consecutively enrolled during the defined study period. Retrospective data were collected from all eligible patients treated between February 27, 2018, and March 9, 2021. Prospective enrollment then continued consecutively from March 10, 2021, to January 10, 2022.

The inclusion and exclusion criteria were identical for both the retrospective and prospective cohorts, as follows:

Inclusion criteria

  • Age > 18 years.

  • Pathological diagnosis of NPC.

  • Complete staging with bone scan, ultrasound, CT, or MRI with or without positron emission tomography/CT results.

  • Complete demographic records.

  • Epstein–Barr virus (EBV) viral load measurement.

  • No history of prior treatment.

Exclusion criteria

  • Loss to follow-up or incomplete treatment.

  • Contraindications for MRI.

  • Previously treated NPC.

  • Poor image quality.

All patients were evaluated using DWI before (pre-treatment) and 5 weeks (intra-treatment) after starting radiation or chemoradiotherapy, with or without induction chemotherapy, according to the treatment protocol provided by the Radiation Oncology Division of our institution.

All demographic and clinical data of the patients, including age, sex, tumor–node–metastasis (TNM) stage, EBV status, pathological results, and treatment protocols were recorded. NPC staging was performed following the American Joint Committee on Cancer (AJCC) guidelines, eighth edition [19]. Patients were categorized into three major groups: (1) low and high AJCC stage groups (stage I–II versus vs. III–IV, respectively); (2) low and high T-stage groups (T1–T2 vs. T3–T4, respectively); and (3) low and high N stage group (N0–N1 vs. N2–N3, respectively). Additionally, based on the pathological results, the patients were classified into undifferentiated and differentiated squamous cell carcinoma groups. Plasma EBV status was recorded as detectable (plasma EBV ≥ 316 copies/mL) and undetectable (plasma EBV < 316 copies/mL). The cutoff of 316 copies/mL corresponds to the lower limit of quantification established by our institutional laboratory validation protocol for plasma EBV DNA measurement.

The prescription dose for the high-risk planning target volume was 70 Gy (2.12 Gy/fraction). The dose administered to the low-risk planning target volume was decreased to 54 Gy (1.64 Gy/fraction) in 33 fractions using the intensity-modulated radiation therapy technique. Concurrent chemotherapy consisted of weekly cisplatin 40 mg/m² for at least five cycles. Adjuvant chemotherapy comprised cisplatin (80 mg/m²) plus 5-fluorouracil (1,000 mg/m²) for 24–96 h continuous infusion at 4-week intervals for three cycles. The induction chemotherapy regimen was as follows: docetaxel (60 mg/m²) for day 1, cisplatin (60 mg/m²) for day 1, and 5-fluorouracil (600 mg/m²) for days 1–5 at 3-week intervals for three cycles.

The treatment outcomes were followed up until July 31, 2022, or until the patients experienced events that were counted as poor outcomes. Patients were divided into two groups based on outcomes: good and poor outcomes. Patients were defined as having a poor outcome if they experienced an event considered an endpoint in the progression-free survival (PFS) analysis, such as locoregional recurrence, distant metastasis, and death. Conversely, patients were defined as having a good outcome if they did not experience any event, which was considered an endpoint and was followed up until July 31, 2022, and further analyzed in the PFS analysis.

Magnetic resonance imaging

All MRI examinations were performed with a 1.5-Tesla MRI scanner (Signa HDxt, GE Medical Systems, Chicago, IL. USA) using a six-channel flex coil with a routine MRI simulation protocol and an additional axial diffusion-weighted sequence called the “periodically rotated overlapping parallel lines with enhanced reconstruction” (PROPELLER) technique. MRI was performed at two time points for each patient: before (pre-treatment) and approximately 5 weeks after initiating chemoradiation therapy (intra-treatment, approximately 49 Gy), in accordance with the standard patient visit schedule. This imaging protocol was followed regardless of whether induction chemotherapy had been administered prior to concurrent chemoradiation therapy (CCRT). An immobilization mask was created for each patient to reduce head and neck movements. The acquisition parameters were as follows: TR/TR, 5000/79.81 ms, b-value of 0 and 800 s/mm2; receiver bandwidth, 650.78 Hz/pixel; matrix size, 256 × 256; slice thickness, 5 mm; gap, 5 mm; field of view (FOV), 260 mm2; and total acquisition time, 5.04 min. The FOV covered the entire primary tumor volume and organ of interest during pre- and intra-treatment. The ADC images were generated from the DWI images using two b-values. We utilized the INLINE auto postprocessing ADC/eADC software provided by the vendor (GE Medical Systems) to calculate the ADC maps. The ADC was calculated as follows to quantify the diffusion motion: ADC = 1/b*(ln Sb/S0), where S0 and Sb are the DWI values at the b-value of 0 and 800 s/mm2, respectively.

Data analysis

Change in tumor volume

Two board-certified neuroradiologists with 12 (NJ) and 4 (TP) years of experience manually delineated the regions of interest (ROI) on all slices that contained the primary NPC tumor on diffusion-weighted images before and after treatment, creating the volume of interest (VOI). This process was performed independently, and the neuroradiologists were blinded to clinical information and treatment outcomes. Subsequently, an interobserver analysis between the two experts was performed. The software used for the drawing process was the 3D Slicer version 5.0.2 (www.slicer.org). The size reduction for each tumor volume for individual patients was calculated from the VOI as the percentage change in the volume at intra-treatment compared with the pre-treatment images using the equation: percentage change of the volume (%\(\:\varDelta\:\)Vol) = 100 × (Vp – Vm)/Vp, where Vp is the volume of the tumor before treatment, and Vm is the volume of the tumor during treatment.

Change in ADC value

We also calculated the change in the ADC value between intra-treatment and pre-treatment ADC values using the equation: percentage change in ADC (%∆ADC) = 100 × (ADCm – ADCp)/ADCp, where ADCp is the ADC value of the tumor at pre-treatment and ADCm is the ADC value of the tumor at intra-treatment.

PRM analysis

We used PRM analysis, which is based on voxel-wise subtraction between the aligned DWI/ADC images during pre- and intra-treatment, to fully capture the spatial and temporal changes of the tumor during treatment. Image co-registration between the pre- and intra-treatment images was performed using monomodal affine mutual information in MATLAB software (MathWorks, Inc., Natick, MA, USA), using a two-step registration process. First, we used the diffusion-weighted image at intra-treatment to co-register with the pre-treatment diffusion-weighted image, resulting in a geometric transformation matrix for registration. A board-certified neuroradiologist visually verified the registration of each pair of pre-treatment and intra-treatment diffusion-weighted images to ensure the accuracy and acceptability of the registration. To address potential distortion in DWI images, the PROPELLER DWI sequence was used to reduce susceptibility artifacts. Additionally, all co-registered image pairs were visually reviewed on a slice-by-slice basis by an experienced neuroradiologist to confirm accurate alignment. We also used the resulting transformation matrix to warp the intra-treatment ADC map into a pre-treatment frame to match the pre-treatment ADC map. Subsequently, the difference between the co-registered image and the pre-treatment ADC maps (∆ADC map) for each voxel within the tumor VOI at pre-treatment was observed. Each voxel was classified according to its corresponding change in ADC (∆ADC) and a threshold value indicating the significance of ∆ADC.

∆ADC was categorized into three classes: red, blue, and green. The voxel that showed a significant increase in the ADC value beyond the pre-defined threshold value was displayed in red (∆ADC > threshold). Whereas the voxel showing a significant decrease in the ADC value below the threshold was displayed in blue (∆ADC < –threshold). Furthermore, the voxel without significant changes in ∆ADC value was displayed in green (–threshold < ∆ADC < threshold). The PRM analysis focused on voxels with ADC values above the threshold based on the hypothesis. The proportion of voxels with a significant increase in the ADC value, defined as PRM+, was calculated as follows: PRM+ = (N+/Ntotal) × 100, where N+ is the number of voxels with increased ADC values, and Ntotal is the total number of voxels within the tumor.

Scatter plots show the distribution of PRM changes throughout the entire tumor VOI. In this study, five different thresholds (threshold = 250, 500, 750, 1,000, and 1,250 × 10− 6 mm2/s) were used in the PRM analysis. The PRM threshold was selected based on a statistical method to determine the most significant threshold with the least effect on other variables.

Statistical analysis

The Shapiro–Wilk normality test was used to determine whether the data were normally distributed. The demographic and clinical characteristics of the patients were described. Continuous variables are expressed as median (interquartile range [IQR]) and categorical variables as percentages. Differences in continuous and categorical variables between the two groups (good and poor outcomes) were assessed using the Wilcoxon rank-sum test and the chi-square or Fisher’s exact test, respectively.

The cumulative PFS rate was calculated using the Kaplan–Meier and log-rank tests to compare between the two groups. Maximally selected rank statistics were used to identify the cut-off points for PRM+ that would best predict poor treatment outcomes. Cox regression analysis was used to determine factors associated with poor treatment outcomes. Multivariate models were developed by adjusting for covariates (p < 0.1) in the univariate models and stepwise backward logistic regression to select the final model. The predictive ability of the final model was evaluated using Harrell’s concordance index (C-index). The interobserver reliability for continuous measures obtained by two experts was assessed using the concordance correlation coefficient, which was calculated based on Lin’s (1989, 2000) concordance correlation. Statistical analyses were performed using STATA version 15.1 A (STATACorp, College Station, TX, USA). Statistical significance was set at p < 0.05. There was no missing data for the primary imaging or clinical variables analyzed in this study.

Results

Patient data

A total of 80 patients with pathologically confirmed NPC who underwent MR examinations participated in the study. Of these, 10 patients were excluded because of poor image quality. Finally, 70 patients were included in the study (Fig. 1). Table 1 presents the patient demographic data. Among the 70 patients, 5 (7.1%) received RT, 21 (30%) received CCRT, 13 (18.6%) received CCRT with adjuvant chemotherapy, and 31 (44.3%) received induction chemotherapy followed by CCRT. The median number of concurrent chemotherapy cycles was 5 (range 4–7). The median number of adjuvant chemotherapy cycles was 3 (range 1–3). Among patients receiving induction chemotherapy, the median number of cycles was 3 (range 2–3). Overall, 53 patients (75.7%) had good outcomes and 17 (24.3%) had poor outcomes. Among patients with poor outcomes, seven (10%) had local recurrence with pathological confirmation, 10 (14.3%) had distant metastases, and nine (12.9%) died. Follow-up was conducted between February 27, 2018, and July 31, 2022, with a median follow-up duration of 31.2 months (IQR = 19.2–37.2 months).

Fig. 1
figure 1

Flowchart of patient inclusion; Abbreviation: MRI, magnetic resonance imaging

Full size image
Table 1 Demographic data and clinical characteristics of the patients
Full size table

Response prediction

The initial ADC value was significantly higher in the poor outcome group than that in the good outcome group in the pre-treatment images (9595 × 10− 6 vs. 8715 × 10− 6 mm2/s, p = 0.005). Additionally, the tumor volume was significantly higher in the poor outcome group than in the good outcome group in the pre- and intra-treatment analyses (pre-treatment: 6.802 vs. 2.588 cm3, p = 0.013; intra-treatment: 0.872 vs. 0.386 cm3, p = 0.016). PRM+ at every threshold was significantly lower for poor outcomes than that for good outcomes (thresholds: 250, 500, 750, 1,000, and 1,250 × 10− 6 mm2/s: 24.0% vs. 70.1%, 23.2% vs. 68.3%, 23.0% vs. 67.0%, 22.8% vs. 65.8%, and 22.4% vs. 64.1%, p < 0.001 [at all thresholds]). The ADC value in the intra-treatment image was higher for poor outcomes than for good outcomes, with no statistical significance (12,604 × 10− 6 vs. 11,486 × 10− 6 mm2/s, p = 0.169). The %∆Vol was lower in the poor outcome group than that in the good outcome group, with no statistical significance (77.4% vs. 86%, p = 0.114). The %∆ADC in the poor outcome group was lower than that in the good outcome group, without statistical significance (17.9% vs. 33.2%, p = 0.108). The comparisons of ADC, tumor volume, and PRM+ with treatment outcomes are presented in Table 2.

Table 2 Comparison of %∆ADC, %∆Vol, and PRM+ between good and poor outcomes
Full size table

The scatter plot in Fig. 2 depicts the correlation between %∆ADC and %∆Vol for good and poor outcomes, which were not significantly different (p = 0.108, 0.114, respectively). However, only PRM+ was significantly higher in the good outcome group than that in the poor outcome group at every threshold (p < 0.001). The PRM threshold was set at 1,250 × 10− 6 mm2/s, as it showed the least correlation and effect to the other variables. This threshold selection was used to determine the cut-off point and was used in the univariate and multivariate analyses.

Fig. 2
figure 2

Scatter plot comparing %∆ADC, %∆Vol, and PRM+ of good and poor outcomes Sections A and B do not show significant differences for %∆ADC and %∆Volume and treatment outcomes. Section C shows a significantly higher PRM+ in patients with good outcomes at every threshold

Full size image

Figs. 3 and 4 show representative cases of the PRM analysis in patients with good and poor outcomes, respectively. The scatter plot depicts the quantification and distribution of the pre-treatment (y-axis) and intra-treatment ADC values (x-axis) for the entire tumor VOI. The PRM scatter plot had more red voxels for patients with good outcomes than for those with poor outcomes.

Fig. 3
figure 3

Representative case of the good outcome. A patient with nasopharyngeal carcinoma (NPC) presented with an infiltrative mass involving the left-sided nasopharynx and left posterior nasal cavity in a pre-treatment image (arrow) (A). This mass shows restricted diffusion on the pre-treatment apparent diffusion coefficient (ADC) image with ADC value = 7,248 × 10− 6 mm2/s (arrow) (B). Intra-treatment ADC image of the mass shows decreased size and degree of restricted diffusion with ADC value = 9,824 × 10− 6 mm2/s (arrow) (C). The percentage change in ADC (%∆ADC) is 35.54%. The color parametric response map shows the voxels with significantly increased ADC value (PRM+) beyond, between, and below the threshold (depicted in red, green, and blue, respectively), with most of the voxels in the tumor depicted in red (D). Parametric response mapping (PRM) scatter plot of the mass shows the distribution of changes in PRM+ over the entire tumor with PRM+ (threshold 1250) = 80.9% (E). No recurrence was detected after 17 months of follow-up (F)

Full size image
Fig. 4
figure 4

Representative case of the poor outcome. A patient with nasopharyngeal carcinoma (NPC) presented with an infiltrative mass involving the right-sided nasopharynx and right posterior nasal cavity in a pre-treatment image (arrow) (A). This mass shows restricted diffusion on the pre-treatment apparent diffusion coefficient (ADC) image with ADC value = 11,186 × 10− 6 mm2/s (arrow) (B). Intra-treatment ADC image of the mass shows a much-decreased size and degree of restricted diffusion with ADC value = 14,739 × 10− 6 mm2/s (arrow) (C). The percentage change in ADC (%∆ADC) is 31.76%. The color parametric response map shows the voxels with significantly increased ADC value (PRM+) beyond, between, and below the threshold (depicted in red, green, and blue, respectively) with an equivocal number of red and blue voxels (D). Parametric response mapping (PRM) scatter plot of the mass shows the distribution of change in PRM+ over the entire tumor with PRM+ (threshold 1250) = 24.04% (E). At 19 months after treatment initiation, the patient developed a new enhancing mass at the right-sided nasopharynx (arrow) (F), with the pathological result confirmed as a tumor recurrence

Full size image

Table 3 displays the performance of ADC, tumor volume, and PRM+ to predict a poor outcome. ADCp ≥ 9500 × 10− 6 mm2/s, Vp ≥ 3.6 cm3, %∆Vol ≥ − 85%, and PRM+ < 35% were identified as variables that are significantly associated with poor treatment outcomes.

Table 3 Performance of ADC, tumor volume, and PRM+ to predict poor outcome
Full size table

Table 4 presents the results of the univariate and multivariate analyses. Male sex, high initial T-stage (T3–T4), high stage (III–Iva), EBV positive, Vp ≥ 3.6 cm3, %∆Vol < 85%, ADCp ≥ 9500 × 10− 6 mm2/s, and PRM+ (threshold 1250) < 35% were the risk factors associated with a poor outcome in the univariate model (p < 0.1). The multivariate model showed that only high initial T-stage (T3–T4) and PRM+ < 35% were significant predictors of poor outcome.

Table 4 Univariate and multivariate analyses of risk factors that associated with poor outcome
Full size table

The 3-year PFS rates for PRM+ at threshold 1,250 × 10− 6 mm2/s at cut-off value ≥ 35% and PRM+ < 35% were 93.6% vs. 16.8%, respectively; log-rank test: p < 0.001 (Fig. 5).

Fig. 5
figure 5

Progression-free survival rate by PRM+ at threshold 1,250 × 10− 6 mm2/s

Full size image

Interobserver reliability

A high interobserver reliability was observed for PRM+ between the two investigators, with a concordance correlation coefficient (95% confidence interval) of 0.829 (0.576–1), 0.839 (0.596–1), 0.832 (0.582–1), 0.864 (0.659–1), and 0.869 (0.670–1), at the thresholds 250, 500, 750, 1,000, and 1,250 × 10− 6 mm2/s, respectively. Lin’s concordance correlation coefficient of 0.71–0.80, 0.81–0.90, 0.91–0.95, and > 0.95 indicate moderate, fairly good, very good, and excellent, respectively.

Discussion

In this study, we evaluated the use of ADC and PRM to predict treatment outcomes in patients with NPC. We found that patients with poor outcomes had significantly higher pre-treatment ADC values and significantly lower PRM+ values than those with good outcomes. Multivariate analyses identified the cut-off values of PRM+ < 35% and initial T-stage 3–4 as the only risk factors associated with poor outcomes.

Our results showed that patients with poor treatment outcomes had a significantly higher tumor volume on pre- and intra-treatment images. The high initial T-stage (T3–T4) and high tumor volume reflect a high tumor burden, resulting in poor treatment response, poor survival rate, and a higher rate of local failure and tumor recurrence [20, 21]. The percentage change in tumor volume indicates how much the tumor shrunk after the beginning of treatment compared with that at pre-treatment; therefore, it might reflect the future response [15, 22]. Previously, Galban et al. [15] did not find any correlation between tumor volume and response to treatment. However, Vandecaveye et al. [22] found that the change in tumor volume 2 weeks after treatment initiation was significantly lower in patients with local recurrence. Thus our results, which indicated that a decrease in tumor volume of < 85% was associated with poor treatment outcomes, are consistent with those of Vandecaveye et al. This discrepancy between our results and those of Galban et al. may be explained by the different contouring processes they used, which included the primary tumor and pathologic lymph nodes. Furthermore, the interval imaging period in the previous study was 3 weeks after treatment initiation, compared with 5 weeks in our study.

Previous studies aimed to evaluate the correlation between ADC before treatment and the response to treatment and reported varying results. Some studies indicated that a high pre-treatment ADC value is associated with a higher local failure rate and a lower 3-year survival prognosis [23,24,25,26]. This explanation is based on the fact that areas with high pre-treatment ADC values may reflect the necrotic portion and cellular hypoxia [25]. Our results are consistent with those of previous studies, in that a high pre-treatment ADC was associated with poor treatment outcomes. In contrast, various studies have found that low pre-treatment ADC values are associated with poor treatment outcomes [27, 28]. Garbajs et al. [29] indicated that a low pre-treatment ADC reflects high cellular tissue, such as tumors, causing limited diffusion capacity of water molecules, resulting in low ADC values. The different results reported by these studies are probably due to tumor heterogeneity, which alters ADC values.

The correlation between the treatment outcome and ∆ADC has been reported in previous literature [9, 13, 14, 22, 30]. According to Hong et al. [30], the ∆ADC values of patients without residual tumors at 3 months were significantly higher than those of patients with residual tumors. Tangyoosuk et al. [9] also indicated that the ∆ADC was significantly lower in the poor response group than that in the good response group, with a follow-up time of at least 6 months. These two studies delineated ROIs based on the largest tumor-bearing slice resulting from a two-dimensional analysis, which is different from that in our study. Furthermore, Vandecaveye et al. [22] showed that the ∆ADC between the 2- and 4-week periods was significantly lower for lesions with post-treatment recurrence than for lesions with complete response, and patients with a higher ADC showed a higher locoregional control rate. They delineated ROIs on whole slices of solid primary tumors and pathological lymph nodes for three-dimensional analysis. Recent studies by Liu et al. and Ai et al. further suggest that ∆ADC measured from pre- and post-induction chemotherapy could predict prognosis, with higher ∆ADC after induction chemotherapy associated with better survival outcome [13, 14]. These studies supported the hypothesis that a higher ∆ADC value between pre- and intra-treatment reflected a significant decrease in tumor cellularity. This is because the number of tumor cells would decrease after treatment initiation, thus increasing the ADC value. However, our study did not find a significant correlation between ∆ADC and treatment response, possibly due to differences in the study protocol since previous studies drew ROI on the single largest tumor dimension slice [9, 30]. Furthermore, some studies included solid primary tumors and pathologic lymph nodes with the exclusion of the necrotic component [22] compared with our study, which delineated the ROI on every tumor-bearing slice, resulting in analysis of the entire primary tumor that may include the cystic and necrotic components, which better represents the tumor and its heterogeneity. In addition, the heterogeneity of treatment regimens in our cohort and variability in timing of imaging related to treatment may have contributed to the differences observed in our study compared with previous reports.

The PRM analysis of head and neck squamous cell carcinoma (HNSCC) has previously been performed using parameters such as ADC and dynamic contrast-enhanced MR perfusion [15, 16]. Studies on PRM in other types of cancers, such as breast cancer and glioblastoma have also been reported with promising results showing that PRM can predict the response to treatment better than conventional methods [31, 32]. Galban et al. [15] compared pre-therapy and intra-therapy (3 weeks after pre-therapy) using DWI, ADC, and PRM of ADC (PRM+) to monitor treatment-induced tissue alteration in HNSCC. The results showed that %∆ADC and %∆Vol did not correlate with tumor control at 6 months. Only PRM+ showed significant differences among patients with different outcomes. Their study also showed that changes in ADC in the complete response group were higher than those in the partial response group. Furthermore, they observed a drastic decrease in tumor volume in the complete response group compared with that in the partial response group. However, none of these results were statistically significant. Our findings are consistent with these results, which suggest that PRM+ is a better predictive factor of poor outcomes than %∆ADC and %∆Vol. We found the same trend of increased ADC values and decreased tumor volume after treatment. However, we did not observe a significant correlation between the %∆ADC and treatment response. Additionally, we observed a %∆Vol, which correlates with poor outcomes in univariate analysis but is not significant in the multivariate analysis. These differences may be because of some differences between our study and the previous study; for example, (1) the previous study focused on various HNSCC whereas our study specifically focused on NPC, (2) different time intervals between pre- and intra-treatment image acquisition with a 3-week interval in the previous study and a 5-week interval in our study, and (3) a longer follow-up time in our study that better reflects the long-term outcome. Moreover, our study demonstrated a correlation between treatment outcomes and PRM+ at a cut-off value of 35%. The PRM+ < 35% was significantly associated with poor treatment outcomes, whereas PRM+ ≥ 35% was associated with a better PFS rate. This may indicate that PRM+ is superior in treatment outcome prediction compared to ∆Vol and ∆ADC. ADC reflects the average values of the whole tumor and may not be comparable to PRM+, which analyzes the tumor based on voxel-by-voxel methods.

PRM analysis demonstrates considerable potential as an adjunctive tool within the post-processing and reporting workflow for radiologists involved in the management of NPC. Its voxel-wise analytical approach provides additional quantitative information that complements conventional imaging assessments. With further clinical validation, PRM may be integrated into standard imaging protocols to enhance the evaluation of treatment response. Moreover, PRM-derived metrics could serve as valuable biomarkers to support multidisciplinary tumor board deliberations by providing early indicators of therapeutic efficacy or disease progression, thereby facilitating more informed and individualized treatment planning.

Nonetheless, our study has some limitations. First, the small sample size and imbalance between poor and good responders may have introduced bias and limited the statistical power to detect significant differences, particularly in subgroup analyses. Second, variations in tumor volumes and treatment protocols, including differences in the use of induction chemotherapy and concurrent chemoradiation regimens, may have influenced individual responses and PRM+ values, introducing potential confounding effects that could affect the interpretation of the imaging biomarkers and may limit the generalizability of our findings. Future research with larger sample sizes should stratify patients by tumor volume, such as small and large volume disease, and by treatment regimens to enable subgroup analyses and better assess the consistency of PRM+ as a predictive marker. Third, the follow-up time was rather short, and longer studies could provide insights into long-term outcomes. Fourth, the DWI sequence may have susceptibility artifacts between different tissue interfaces. We reduced these artifacts using the PROPELLER DWI pulse sequence with non-echoplanar imaging, which enhance spatial resolution and reduced susceptibility artifact. Furthermore, the co-registration process for DWI studies was evaluated qualitatively. Future studies should follow Quantitative Imaging Biomarkers Alliance (QIBA) recommendations for precision in multicenter trials. Finally, external validation of the predictive tool using the established PRM+ cut-off and application in diverse geographic and clinical settings is essential for confirming the generalized applicability of the results. 

Conclusions

Our results demonstrated that low PRM+ and high initial T-stage were associated with poor treatment outcomes. Therefore, PRM+, which is superior to the ADC value, can be used as a predictive tool for treatment outcomes in patients with NPC. Intensive follow-up and/or adjuvant treatment may be beneficial in these patients.