Diffusion weighted imaging-based tumor growth rate for predicting long-term survival in nasopharyngeal carcinoma

Abstract

Background

To explore the prognostic value of tumor special growth rate (TSGR) in patients with NPC and investigate the potential of diffusion-weighted imaging (DWI) habitat analysis as a surrogate marker.

Methods

Patients who underwent two MRI scans before any type of treatment at two institutes were included. Radiologists manually delineated the volume of the primary tumor, and the doubling time and TSGR were calculated. Correlation between TSGR and long-term survival was evaluated using multivariate Cox analysis. Moreover, a radiologist delineated the tumor on b = 0 s/mm² images during the first MRI scan. Each tumor was divided into two subregions based on the individual mean apparent diffusion coefficient value (ADC_Mean), and the ADC value (ADC_down [mm²/s] and ADC_up [mm²/s]) of each subregion was calculated. ADC values for diagnosing TSGR and predicting survival were evaluated and validated using 5-fold cross-validation (iterations = 1000) and in an independent cohort.

Results

A total of 98 patients were included (median age: 51 years; 24 females; training cohort: 64). The median doubling time was 131.9 days. Patients in the rapid-growth group (TSGR ≥ 1.4%) had significantly poorer survival than those in the slow-growth group (TSGR < 1.4%) in both the training and validation cohorts (p ≤ 0.006). ADC_down was a valuable factor in predicting TSGR (area under the curve: 0.792). Notably, the addition of ADC_down further improved the predictive value of TNM stage (0.663–0.848 vs. 0.631–0.785).

Conclusion

Patients with high ADC_down value may have the rapid-growth potential and were more likely to have a poor prognosis.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s40644-026-01041-9.

Background

Nasopharyngeal carcinoma (NPC) primarily affects residents of East and Southeast Asia and southern China [1]. Early curative treatment after the cancer diagnosis is important as the prolonged wait time for treatment constitutes an independent predictor of worse NPC mortality [2, 3]. However, some unexpected socioeconomic situations including pregnancy, Covid-19, and examinations may create dilemmas regarding whether to pursue timely treatment. Although such patients are relatively few, with the increasing trend toward younger populations affected by NPC, this situation should receive attention from clinicians. Therefore, conducting growth rate prediction for NPC before treatment may provide a certain basis for patient decision-making.

Studies in oropharyngeal and laryngeal carcinomas have demonstrated that tumor-specific growth rate (TSGR) calculated using diagnostic and radiotherapy-planning computed tomography (CT) is an important factor in long-term survival [4–6]. Only one study conducted in 2011 preliminarily explored the NPC doubling time based on two CT scans; however, low soft tissue resolution and large slice intervals may have impacted the accuracy of the results. Magnetic resonance imaging (MRI) has higher soft tissue structure resolution and, therefore, plays a crucial role in the evaluation of primary tumor invasion and metastatic lymph nodes, with accurate and reliable size measurements. The tumor growth-rate evaluation is typically performed using two pretreatment imaging examinations, and it is considered unethical to postpone treatment for additional imaging. Thus, an effective, robust predictive method to ascertain the TSGR based on the first imaging is needed.

Routinely acquired diffusion-weighted imaging (DWI) can noninvasively assess water-molecule diffusion and tissue cellularity [7–9]. Fu et al. [10] found that the additive irradiation of a subregion (areas with apparent diffusion coefficient [ADC] values below the ADC_Mean value) of the primary tumor in patients with NPC significantly improved patient survival without increasing acute toxicity. Imaging habitat analysis enables voxel classification based on comparable tissue properties to characterize tumor subregions [11]. However, research on the prognostic and predictive value of DWI habitat analysis in NPC is limited. This study aimed to explore the correlation between TSGR and treatment outcomes. A secondary objective was to assess the potential of DWI-based habitat analysis as an alternative to TSGR and in predicting long-term survival.

Materials and methods

Patients

This study was approved by the institutional ethics committee (No.23/242–3984) and conducted in accordance with the Declaration of Helsinki. The requirement for written informed consent was waived because of the retrospective nature of the study. Consecutive patients were recruited from one Chinese tertiary institute (Beijing [2014/01/01 ~ 2022/12/31]) and one US tertiary institute (Institute 2: Stanford [2007/1/1 ~ 2022/12/31). The inclusion criteria included: (1) pathologically confirmed NPC at the time of the initial MRI; (2) at least two MRI scans before treatment at the same institute; (3) confirmed treatment per standard of care. The exclusion criteria were as follows: (1) presence of a second primary tumor; (2) the primary tumor was not fully visualized on MRI; (3) without adequate MRI quality; (4) the patient was loss to follow-up. A total of 64 patients from institute 1 were included in the training cohort, and 34 patients from institute 2 were included in the validation cohort, respectively. The patient flowchart is shown in Fig. 1.

Fig. 1.

Patient flowchart

MRI technique

All MRI scans were acquired using 3.0 or 1.5 T MR scanners (Siemens, Erlangen; Discovery MR 750, GE Healthcare) with an 8-channel head and neck phase array coil. Details of the MRI scan parameters were listed in the Table S1.

MRI tumor volume and growth rate analysis

Image quality of all the MRIs was rated excellent by two senior radiologists. Two radiologists (1 and 2, with 5 and 3 years of experience in head and neck imaging, respectively) manually delineated the volume of the primary tumor slice-by-slice on axial T2WI using ITK-snap software (version 3.8.0). The radiologists used available anatomic images to carefully differentiate the primary tumor from the retropharyngeal lymph nodes. One month later, radiologist 1 delineated 30 randomly selected patients to measure the intra-observer agreement. In the validation cohort, radiologist 1 delineated all the patients twice at a 1-month interval. When patients underwent more than two MRI scans, only the first and last MRI scans obtained before treatment were selected for analysis. With reference to previous studies [4–6], the time interval (t) between two index MRI scans was recorded, and the doubling time (DT) was calculated using Eq. (1):

1

The TSGR was calculated by considering the inaccuracy of the growth rate measurement within a short time interval using Eq. (2):

2

Where V1 and V2 are the tumor volumes on the first and last MRI scans, respectively. The representative patients were shown in Figure S1.

DWI habitat analysis

The ADC map of DWI was obtained using ADW 4.7 software. Radiologist 1 manually delineated the volume of the primary tumor on b = 0 s/mm² images in random order, with reference to conventional MRI sequences. One month later, radiologist 1 re-delineated all the patients in the random order. Based on pixel level, each ADC region was divided into two different subregions (down subregion [lower than the ADC_Mean] and up subregion [higher than the ADC_Mean]) based on their own ADC_Mean. Therefore, each patient had three ADC parameters: ADC_down (the mean ADC value of “down” region), ADC_up (the mean ADC value of “up” region), and f.

Treatment and outcomes

The TNM stage of all patients was determined using the 9th edition of the AJCC staging system [12]. In the training cohort, two senior radiologists (M.L. and H.M.Z.) independently evaluated the TNM stage based on first or last MRI scan, and recorded patients with progression in T and N stages. Any disagreement was resolved by discussion. Patients who developed new neck region metastases were recorded. In addition, 64 and 25 (39.1%) patients underwent EBV DNA detection at first and last MRI scan, respectively in the training cohort, and the dynamic change of EBV DNA level was also recorded.

All patients underwent curative treatment strategies after the last MRI. All patients underwent regular examinations at regular intervals (every 3–4 months during the first two years, every 6 months for third to fourth year, and annually thereafter). In cases in which tumor progression is suspected, biopsy or positron emission tomography (PET-CT) scan was performed.

The primary endpoint was distant metastasis-free survival (DMFS), defined as the time from initial diagnosis to the first occurrence of distant metastasis or death from any cause. The secondary endpoints were progression-free survival (PFS; defined as the time from the initial diagnosis to disease progression or death from any cause) [13]. The follow-up endpoints were set for December 31, 2024.

Statistical analysis

All analysis were performed using R studio (version 4.3.3; R Foundation, Vienna, Austria; https://www.r-project.org/), SPSS (version 26.0; IBM Corp., Armonk, NY, USA), Prism (version 9.4.1; GraphPad Software, Boston, MA, USA), and X-tile software. Consistency was evaluated using an intraclass correlation analysis (ICC). Continuous variables were presented as median and interquartile range (IQR), and categorical variables as percentage. A two-tailed P value < 0.05 was considered statistically significant.

Independent sample t-test, Mann-Whitney U test, or Kappa test was used, as appropriate. Univariate and multivariate Cox regression analysis were used to evaluate the correlations between clinical factors, TSGR and multiple survival endpoints. The optimal cut-off value of TSGR was determined using X-tile software, and patients were divided into Rapid growth (RG) and Slow growth (SG) groups. Kaplan-Meier curves with log-rank test were plotted for both the training and validation cohorts.

As TSGR is rarely acquired in clinical settings, the potential of ADC maps combined with habitat analysis for predicting TSGR was evaluated. The area under the curve (AUC) was calculated, and the optimal cut-off value of ADC parameters were determined using the receiver operating characteristic curve (ROC) analysis. The K-M curves were also plotted. Two types of models were constructed: a TNM stage model and a combined model (TNM stage + ADC map), and the C-index was calculated using 5-fold cross-validation (iteration = 1000) in the training cohort and validated in an independent cohort. The net reclassification index (NRI) and integrated discrimination improvement (IDI) were used to evaluate the additional value of the ADC map, and a nomogram and calibration curves were also plotted. Sample size was calculated using PASS software (version 2025 v25.0.2) with a power of 90% and one-sided α of 0.05.

Results

Patient characteristics

There were no significant differences between the training and validation cohorts (p = 0.058 ~ 0.963), except for the treatment regimen (p = 0.013, Table 1). Eleven (32.4%) patients in the validation cohort were not Asian. There was good agreement in volume measurement and ADC value (all ICC ≥ 0.825, Table S2). The TNM stages of first and last MRI scans for all the included patients are shown in Table S3. The reasons for patients delayed treatment included socioeconomic factors (family matters, Covid-19, graduate admission test) and health-related factors (e.g., pregnancy, hypertension, and diabetes).

Table 1.

Patients characteristics

Characteristics	Training Cohort (n = 64)	Validation Cohort (n = 34)	P value
Age			0.963
< 45	21 (32.8%)	11 (32.4%)
≥ 45	43 (67.2%)	23 (67.6%)
Sex			0.101
Male	45 (70.3%)	29 (85.3%)
Female	19 (29.7%)	5 (14.7%)
WHO			0.214
I	0	1 (2.9%)
II	20 (31.3%)	14 (41.2%)
III	44 (68.7%)	19 (55.9%)
Race			-
Asian	64 (100%)	23 (67.6%)
Not Asian	0	11 (32.4%)
T stage			0.589
T1 + T2	30 (46.9%)	14 (41.2%)
T3 + T4	34 (53.1%)	20 (58.8%)
N stage			0.150
N0 + N1	27 (42.2%)	19 (55.9%)
N2 + N3	37 (57.8%)	15 (44.1%)
M stage			0.338
0	62 (96.9%)	31 (91.2%)
1a+1b	2 (3.1%)	3 (8.8%)
EBV DNA	266.5 (0, 835.5)	85 (50, 1566.5) *	0.943
Treatment			0.009
RT-Only	15 (23.4%)	1 (2.9%)
RT-Plus	49 (76.6%)	33 (97.1%)
MRI Interval (day)	38.5 (25.0, 33.3)	19.0 (11.8, 33.3)	0.006
DT (day)	131.9 (62.0, 249.8)	109.6 (56.6, 209.6)	0.633
TSGR	0.53% (0.28%, 1.12%),	0.63% (0.33%, 1.22%)	0.633

^*, Only 24 (70.6%) patients underwent the EBV DNA level detection. Continuous variables were presented as median and interquartile range (IQR), and categorical variables as percentage

DT, doubling time; TSGR, tumor special growth rate; RT, radiotherapy

Natural NPC tumor growth rate

In the training and validation cohorts, the median DT was 131.9 (IQR, 62.0, 249.8) days and 109.6 (IQR, 56.6, 209.6) days and median TSGR was 0.53% (IQR, 0.28%, 1.12%) and 0.63% (IQR, 0.33%, 1.22%), respectively. For all patients (n = 98), the median DT was 131.9 (60.2, 214.8) days. The univariate analysis results showed that N stage was correlated with DMFS (p = 0.032), and that TSGR with DMFS (HR: 2.401, p = 0.015) and PFS (HR: 2.015, p = 0.043). Multivariate Cox analysis showed that TSGR was significantly correlated with DMFS (HR [95%CI]: 2.648 [1.299, 5.397], p = 0.007) and PFS (HR [95%CI]: 2.362 [1.188, 4.696], p = 0.014, Table 2).

Table 2.

Univariate and multivariate Cox analysis with DMFS and PFS in training cohort

	DMFS				PFS
	Univariate		Multivariate		Univariate		Multivariate
Parameters	HR (95% CI)	P	HR (95% CI)	P	HR (95% CI)	P	HR (95% CI)	P
Age			-	-			-	-
< 45	Reference				Reference
≥ 45	1.925 (0.714, 5.192)	0.196			1.463 (0.613, 3.492)	0.392
Sex			-	-			-	-
Female	Reference				Reference
Male	0.736 (0.312, 1.740)	0.485			0.920 (0.400, 2.118)	0.845
WHO			-	-			-	-
II	Reference				Reference
III	1.110 (0.456, 2.702)	0.818			1.116 (0.485, 2.570)	0.797
T stage *							-	-
T1 + T2	Reference		Reference		Reference
T3 + T4	2.207 (0.907, 5.370)	0.081	2.394 (0.971, 5.900)	0.058	1.752 (0.780, 3.932)	0.174
N stage*							Reference 2.726 (1.127, 6.595)	0.026
N0 + N1	Reference		Reference		Reference
N2 + N3	2.950 (1.094, 7.952)	0.032	3.978 (1.453, 10.893)	0.007	2.280 (0.958, 5.427)	0.063
EBV DNA			-	-			-	-
< Median	Reference				Reference
≥ Median	1.811 (0.781, 4.200)	0.166			1.646 (0.753, 3.601)	0.212
Treatment			-	-			-	-
RT-Only	Reference				Reference
RT-Plus	1.388 (0.472, 4.084)	0.551			1.202 (0.453, 3.190)	0.711
TSGR	2.401 (1.187, 4.855)	0.015	2.648 (1.299, 5.397)	0.007	2.015 (1.024, 3.968)	0.043	2.362 (1.188, 4.696)	0.014

^*, The TNM stage was evaluated based on last MRI scan

DMFS, distant metastasis-free survival; PFS, progression-free survival; TSGR, tumor specific growth rate; HR (95%CI), hazard ratio (95% confidence interval)

The X-tile software identified 1.4% as the optimal TSGR cut-off value, corresponding to 49 days (DT; Figure S2) in the training cohort. According to the cutoff value, nine (14.1%) and seven (20.6%) patients were categorized into the RG group in the training and validation cohorts, respectively. The K-M curves showed that, compared with patients in the SG group, patients in the RG group had significantly lower DMFS and PFS in both the training (p = 0.001 and 0.006, Fig. 2) and validation (p = 0.003 and = 0.002, Fig. 2) cohorts.

Fig. 2.

Survival analysis results categorized by TSGR. Patients in RG group showed poorer survival than those in SG group both in both the training (a and b) and validation cohorts (c and d). DMFS (a and c); PFS (b and d). TSGR, tumor special growth rate; RG, rapid growth; SG, slow growth; DMFS, distant metastasis-free survival; PFS, progression-free survival

We undertook subgroup analysis based on the time interval between the first diagnosis and the initiation of treatment (≤ median vs. > median). In the postponed treatment group (> median), patients in the SG group had significantly favorable DMFS (p = 0.002) and PFS (p = 0.003) than their counterparts, whereas there was no significant difference among patients who underwent RT immediately (all p ≥ 0.091, Table S4). A total of 7/64 (10.9%), 8/64 (12.5%), and 16/25 (64%) patients experienced TNM stage progression, emerging cervical lymph node metastasis region, and evaluated EBV DNA level, respectively. The K-M curves showed there were no significant differences (all p ≥ 0.075, Figure S3).

ADC maps habitat analysis

Sixty-four (100%) and 26 (76.5%) patients underwent DWI scan during the first MRI examination in training and validation cohorts, respectively. Patients in the RG group had higher ADC_up and ADC_down values than those in the SG group (p ≤ 0.005, Table 3), with the AUC (95% CI) of 0.818 (0.668, 0.969) and 0.792 (0.639, 0.945). There was no significant difference in parameter f (p = 0.860). The optimal cut-off value of ADC_down using the ROC curve was 7.79 × 10^− 4 mm²/s. The multivariate Cox analysis showed that ADC_down was an independent predictor of DMFS (HR: 3.320, p = 0.006) and PFS (HR: 2.895, p = 0.008, Table S5). Additionally, patients with high ADC_down had poor survival (DMFS: p = 0.004, PFS: p = 0.003, Fig. 3) than those with the low ADC_down.

Table 3.

The comparison between two groups

Characteristics	RG group (n = 9)	SG group (n = 55)	P value	AUC (95% CI)
ADCup	13.9 (12.8, 15.2)	11.8 (10.0, 12.8)	0.002	0.818 (0.668, 0.969)
ADCdown	8.3 (7.2, 8.9)	6.8 (6.1, 7.8)	0.005	0.792 (0.639, 0.945)
f	56.8% (55.5%, 59.7%)	57.8% (55.8%, 60.0%)	0.860	-
Age			1.000#	-
< 45	3 (33.3%)	18 (32.7%)
≥ 45	6 (66.7%)	37 (67.3%)
Sex			0.713#	-
Female	2 (22.2%)	17 (30.9%)
Male	7 (77.8%)	38 (69.1%)
WHO			0.443#	-
II	4 (44.4%)	16 (29.1%)
III	5 (55.6%)	39 (70.9%)
T stage *			0.729#	-
T1 + T2	5 (55.6%)	26 (47.3%)
T3 + T4	4 (44.4%)	29 (52.7%)
N stage *			0.463#	-
N0 + N1	2 (22.2%)	23 (41.8%)
N2 + N3	7 (77.8%)	32 (58.2%)
EBV DNA *			1.000#	-
< Median	5 (55.6%)	30 (54.5%)
≥ Median	4 (44.4%)	25 (45.6%)
Treatment			1.000#	-
RT-Only	2 (22.2%)	13 (23.6%)
RT-Plus	7 (77.8%)	42 (76.4%)

^*, The TNM stage was evaluated based on the first MRI scan; ^#, Fisher exact test. Continuous variables were presented as median and interquartile range (IQR), and categorical variables as percentage

RG, rapid growth; SG, slow growth; AUC (95%CI), area under the curve (95% confidence interval); RT, radiotherapy

Fig. 3.

Survival analysis results categorized by ADC habitat analysis. Patients with higher ADC_down values had worse long-term survival than those with lower ADC_down values with respect to DMFS (a) and PFS (b). The unit of ADC value is 10^− 4 mm²/s. DMFS, distant metastasis-free survival; PFS, progression-free survival

Our study further evaluated the additional value of ADC_down, two models including TNM stage model and the combined model (TNM stage + ADC map; Table 4) were constructed. The combined model demonstrated higher performance than the TNM stage model in predicting DMFS (0.728 and 0.669), PFS (0.717 and 0.638), with NRI ≥ 0.283 and IDI ≥ 0.065 (Table 5). These results were consistent when the models were validated in the independent cohort (Combined model: 0.711 ~ 0.728, and TNM stage model: 0.649, Table 5). The 5-fold cross-validation confirmed model robustness (Combined model: 0.675 ~ 0.700, and TNM stage model: 0.614 ~ 0.649, Figure S4). The nomogram, calibration curves, and K-M curves were shown in Fig. 4.

Table 4.

The TNM stage model and the Combined model

Parameters	DMFS		PFS
	HR (95% CI)	P	HR (95% CI)	P
TNM stage model
T stage
T1 + T2	Reference		Reference
T3 + T4	2.669 (1.085, 6.563)	0.032	2.046 (0.901, 4.642)	0.087
N stage
N0 + N1	Reference		Reference
N2 + N3	3.475 (1.276, 9.466)	0.015	2.582 (1.072, 6.216)	0.034
Combined model
T stage
T1 + T2	Reference		Reference
T3 + T4	2.330 (0.947, 5.734)	0.066	1.839 (0.813, 4.158)	0.144
N stage
N0 + N1	Reference		Reference
N2 + N3	2.958 (1.082, 8.085)	0.035	2.217 (0.919, 5.347)	0.076
ADC down
< 7.79	Reference		Reference
≥ 7.79	2.737 (1.178, 6.360)	0.019	2.720 (1.228, 6.022)	0.014

The TNM stage was evaluated based on the first MRI scan

DMFS, distant metastasis-free survival; PFS, progression-free survival; HR (95%CI), hazard ratio (95% confidence interval); ADC, apparent difference coefficient

Table 5.

Model performance in predicting long-term survival

Models	Training Cohort		Validation Cohort
	DMFS	PFS	DMFS	PFS
TNM stage model	0.669 (0.551, 0.787)	0.638 (0.540, 0.736)	0.649 (0.473, 0.825)	0.649 (0.473, 0.825)
5-fold CV (n = 1000)	0.649 (0.414, 0.884)	0.614 (0.379, 0.849)	-	-
Combined model	0.728 (0.650, 0.806)	0.717 (0.639, 0.795)	0.728 (0.571, 0.885)	0.711 (0.554, 0.868)
5-fold CV (n = 1000)	0.700 (0.484, 0.916)	0.675 (0.459, 0.891)	-	-
NRI*	0.348 (0.142)	0.343 (0.148)	0.331 (0.380)	0.283 (0.258)
IDI*	0.065 (0.134)	0.076 (0.094)	0.071 (0.354)	0.150 (0.306)

^*, the value in the parentheses means P value

DMFS, distant metastasis-free survival; PFS, progression-free survival; CV, cross validation; NRI, net reclassification index; IDI, integrated discrimination improvement

Fig. 4.

Nomogram (a), calibration curve (b), and K-M curves (c) of the combined model for predicting DMFS. Patients with advanced T and N stages and high ADC_down values were more likely to experience distant metastases during follow-up. DMFS, distant metastasis-free survival

Discussion

In this study, we found the median primary tumor DT time was 131.9 days for NPC, and the TSGR serves as an independent prognostic indicator, with rapidly progressing cases showing significantly inferior outcomes compared to slowly progressing ones. As patients may not have two MRIs prior to treatment for evaluation of the growth rate, we further evaluated the prognostic value of DWI during the first MRI scan combined with habitat analysis in predicting TSGR. Patients with high ADC_down values grew faster. The combination of ADC_down and TNM stage improved the prognostic prediction performance compared to the TNM stage only. Our study included 98 patients from two tertiary institutes in China and the USA. Owing to its retrospective design, this study did not increase the imaging burden. The diversity of patients ensured the generalizability of the findings. Understanding the growth process of NPC not only provides a basis for clinical research, but also helps to reduce patient stress.

RT is the main treatment method for head and neck cancers [14]. Previous studies using diagnostic and RT-planning CT scans showed that the median DT in oropharyngeal carcinoma was 93.7–112 days [4, 5], which was similar to our results. However, another preliminary study of NPC found that the median local tumor (primary tumor and retropharyngeal lymph nodes) DT was 279 days based on 15 patients [15], which was longer than our results. A plausible reason was that the present study was exclusively based on MRI. The anatomy of the neck is complex, and the CT setting with slice intervals of 5 mm in Ho et al. study [15] impeded the accuracy of the volume measurement to some extent. MRI enables more accurate separation of the primary tumor from the retropharyngeal lymph nodes, and further realistically reflects primary tumor growth. Patients in the RG group had significantly poorer DMFS and PFS in both the training and validation cohorts than those in the SG group. Patients in the RG group were more aggressive and had a higher likelihood of treatment failure. Previous studies conducted in head and neck cancer, lung cancer, and hepatic carcinoma also found that TSGR was correlated with a higher TNM stage, more aggressive imaging characteristics (solid vs. non-solid nodule; LI-RADS M vs. 3), and was an independent predictor of survival [4–6, 16–18]. Noteworthy, in the postponed treatment group (> median), patients in the RG group exhibited significantly poorer DMFS and PFS than those in the SG group, whereas there was no significant difference in patients who underwent early treatment. Therefore, timely identification of RG patients is essential to provide prompt treatment.

To ensure the accuracy of primary tumor volume delineation, MRI scan was required [19]. However, it is nearly impossible to perform two MRIs scans within a suitable time interval in the real world. Allowing tumors to grow without intervention not only poses ethical concerns, but also places patients at a significant risk of extensive locoregional invasion, distant metastasis, and poor clinical outcomes [3, 20]. Therefore, we further explored the robust factors that reflect the TSGR. Habitat analysis have facilitated the comprehensive characterization of tumor subregions [11, 21, 22]. A previous study found that an elevated RT dose in the subregion of lower than the individual ADC_Mean in patients with head and neck cancer significantly improved the survival [10, 23]. Therefore, our study divided whole tumor volumes into two subregions based on individual ADC_Mean values in the first MRI scan and found that ADC_down value could be a replaceable indicator of TSGR, and patients with high ADC_down had the potential to grow rapidly. We also found the ADC_down was an independent predictor of DMFS and PFS. In addition, we explored the additional value of ADC_down to TNM stage for long-term survival prediction. The combined model showed better performance than the TNM stage model, which was validated through 5-fold cross-validation as well as in the international validation cohort. The ADC value is inversely proportional to tumor cellularity, and a higher ADC value indicates higher stromal content and more invasive biological features, which may indirectly reflect the potential of RG and poorer survival [7, 8, 10].

MRI, with its excellent soft-tissue resolution, is the modality of choice for the evaluation of NPC. Without imposing additional clinical burden, our study offers a robust method for identifying patients with rapidly growing tumors who are at increased risk of poor survival before treatment. Tailored treatment strategies—such as intensified induction chemotherapy or dose-escalated radiotherapy—may further improve outcomes in this high-risk subset. However, there are some limitations. Firstly, the sample size is relatively small. Ethical constraints precluded prospective large-scale research. Additionally, the calculated sample size was 31 patients (5 patients in RG group and 26 patients in SG group, Figure S5), a minimum which the validation cohort in our study exceeded. However, multi-center studies are required in the future. Secondly, we did not evaluate the DT of the metastatic lymph nodes because manual segmentation was time-consuming. The selection of one cervical lymph node per region as a representative may increase feasibility. Thirdly, the time interval in the training cohort was significantly longer than that in the validation cohort. However, there was no significant intergroup difference in the DT. Lastly, while the heterogeneity in b-value settings (800 and 1000 s/mm²) could potentially compromise statistical power, employing habitat analysis derived from intrinsic ADC_Mean values may reduce the influence of the b-value setting to some extent. VOI delineation was performed on b = 0 s/mm² imaging to ensure accurate delineation of the primary tumor, albeit with the potential inclusion of edema or inflammation tissue. However, the improved performance of the combined model over the TNM stage model in the validation cohort indirectly supports the stability of the model, even under different b-value settings.

Conclusions

Our study demonstrated that the median DT of NPC was 131.9 days, based on the analysis of two MRI scans. The high tissue resolution of MRI and inclusion data from two centers enhanced the reliability and generalizability of our findings. Patients in the RG group had a poorer prognosis than those in the SG group. In addition, ADC_down emerged as a potential imaging biomarker of tumor growth rate. Furthermore, incorporating ADC_down into the TNM staging system enhanced prognostic prediction compared with using TNM staging alone.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Supplementary Material 1 ^{(4.1MB, docx)}

Abbreviations

ADC

Apparent diffusion coefficient

DMFS

Distant metastasis-free survival

DT

Doubling time

DWI

Diffusion-weighted imaging

IDI

Integrated discrimination improvement

LRFS

Loco-regional recurrence-free survival

MRI

Magnetic resonance imaging

NPC

Nasopharyngeal carcinoma

NRI

Net reclassification index

OS

Overall survival

PFS

Progression-free survival

RG

Rapid growth

SG

Slow growth

T1WI

T1-weighted imaging

T2WI

T2-weighted imaging

TSGR

Tumor special growth rate

Author contributions

F.Y.: Conceptualization; Data curation; Investigation; Formal analysis; Methodology; Writing - original draft. Y.Z.: Data curation; Investigation; Formal analysis; Methodology; Writing - original draft. M.I.: Data curation; Investigation; Formal analysis. H.R.W.: Data curation; Investigation; Formal analysis. S.W.: Data curation; Investigation. X.L.L.: Data curation; Investigation. X.D.Y.: Data curation; Investigation; Writing - review & editing. L.L.: Investigation; Methodology. Y.F.Z.: Investigation; Methodology. L.Z.X.: software; D.H.L.: Conceptualization; Formal analysis; Supervision; M.L.: Conceptualization; Formal analysis; Supervision; Validation; Writing - review & editing; H.M.Z.: Conceptualization; Formal analysis; Supervision; Validation; Writing - review & editing.

Funding

This work was supported by the National High Level Hospital Clinical Research Funding (80102022505); and the Beijing Hope Run Special Fund of Cancer Foundation of China (LC2022A23).

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the institutional ethics committee (No.23/242–3984).

Consent for publication

The requirement for written informed consent was waived because of the retrospective nature of the study.

Competing interests

The authors declare no competing interests.