Bao et al.

Insights into Imaging (2022) 13:197

https://doi.org/10.1186/s13244-022-01338-w

ORIGINAL ARTICLE Open Access

Added value of histogram analysis of ADC

in predicting radiation‑induced temporal
lobe injury of patients with nasopharyngeal
carcinoma treated by intensity‑modulated
radiotherapy
Dan Bao1†, Yanfeng Zhao1†, Wenli Wu2, Hongxia Zhong1, Meng Yuan3, Lin Li1, Meng Lin1, Xinming Zhao1 and
Dehong Luo1,4*   

Abstract
Background: This study evaluated the predictive potential of histogram analysis derived from apparent diffusion
coefficient (ADC) maps in radiation-induced temporal lobe injury (RTLI) of nasopharyngeal carcinoma (NPC) after
intensity-modulated radiotherapy (IMRT).
Results: Pretreatment diffusion-weighted imaging (DWI) of the temporal lobes of 214 patients with NPC was retro-
spectively analyzed to obtain ADC histogram parameters. Of the 18 histogram parameters derived from ADC maps, 7
statistically significant variables in the univariate analysis were included in the multivariate logistic regression analysis.
The final best prediction model selected by backward stepwise elimination with Akaike information criteria as the
stopping rule included kurtosis, maximum energy, range, and total energy. A Rad-score was established by combining
the four variables, and it provided areas under the curve (AUCs) of 0.95 (95% confidence interval [CI] 0.91–0.98) and
0.89 (95% CI 0.81–0.97) in the training and validation cohorts, respectively. The combined model, integrating the Rad-
score with the T stage (p = 0.02), showed a favorable prediction performance in the training and validation cohorts
(AUC = 0.96 and 0.87, respectively). The calibration curves showed a good agreement between the predicted and
actual RTLI occurrences.
Conclusions: Pretreatment histogram analysis of ADC maps and their combination with the T stage showed a satis-
factory ability to predict RTLI in NPC after IMRT.

†
Dan Bao and Yanfeng Zhao contributed equally to this work.
*Correspondence: pumccancer@163.com
1
Department of Radiology, National Cancer Center/National Clinical Research
Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences
and Peking Union Medical College, 17 Panjiayuan Nanli, Chaoyang District,
Beijing 100021, China
Full list of author information is available at the end of the article

© The Author(s) 2022. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the
original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or
other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line
to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory
regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this
licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/.
Bao et al. Insights into Imaging (2022) 13:197 Page 2 of 11

Key points

• Histogram parameters from pretreatment ADC maps are associated with RTLI in NPC.
• The ADC and combined models show better performance than the T stage alone.
• The combined model shows excellent performance in predicting RTLI in different subgroups.

Keywords: Nasopharyngeal carcinoma, Diffusion-weighted MRI, Temporal lobe, Radiation therapy, Radiomics

Background in patients with head and neck squamous cell carci-

Radiotherapy remains the primary treatment for naso- noma treated with (chemo)radiation [17]. Diffusion ten-
pharyngeal carcinoma (NPC) [1]. However, radiation- sor imaging (DTI) and DWI can be used to differentiate
induced temporal lobe injury (RTLI) can be a serious benign and malignant head and neck lesions [18]. DWI
complication that severely affects the quality of life and and ADC images can also be used for segmentation [19].
long-term prognosis [2, 3]. Although radiotherapy tech- Histogram analysis is a mathematical method that pro-
niques with better conformance, such as intensity-mod- vides information about the distribution of data in the
ulated radiotherapy (IMRT), provide better long-term selected region of interest (ROI), providing more infor-
disease control and are less toxic [4], RTLI is still reported mation that is often ignored by the human eye [20]. To
in 4.6–8.5% of patients [5, 6]. Symptom-based diagnosis the best of our knowledge, only a few studies have inves-
of RTLI is problematic in clinical practice because most tigated the association between MRI features and RTLI
patients are asymptomatic even at a very late stage or are occurrence in NPC patients [21–25]. Moreover, histo-
already in a stage of non-reversible deterioration when gram analysis has not been extensively explored in this
noticeable symptoms start to appear, during which treat- field.
ment has limited effect [7]. In contrast, if RTLI could This study aimed to investigate the value of pretreat-
be identified early or even predicted before the onset of ment histogram analysis of ADC in the prediction of
symptoms, personalized intervention could be provided RTLI in patients with NPC.
in advance to reverse the unfavorable situation. There-
fore, it is particularly important to predict RTLI noninva- Methods
sively after IMRT. Patients
Several recent studies have focused on predicting RTLI This retrospective single-center study was approved
in patients with NPC [8–10]. Studies based on radia- by the local institutional review board, which waived
tion dosimetry-related factors have shown some predic- the requirement for informed consent. Our radiologi-
tive potential [9–11]; however, the optimal dose/volume cal database was queried between January 2017 and
predictors of RTLI still vary among different studies, and December 2021. The inclusion criteria were as follows:
clinical applications are limited [12]. The imaging diagno- (a) histopathologically confirmed NPC; (b) head–neck
sis of RTLI mainly depends on magnetic resonance imag- MRI including DWI performed in our institution within
ing (MRI) findings. MRI is a suitable tool with multiple 2 weeks before any treatment; (c) receiving IMRT; and
techniques available, which not only enables structure (d) RTLI after IMRT was found during follow-up. The
depiction, but also function quantification. exclusion criteria were as follows: (a) history of any
Diffusion-weighted imaging (DWI) is a functional tech- prior local–regional therapies, (b) poor image quality
nique that provides information about the tissue micro- due to severe artifacts, (c) temporal lobe invasion, and
environment depending on the microscopic mobility (d) recurrent NPC. In total, 107 patients with RTLI were
of water [13]. Because of the Brownian motion of water included according to the inclusion and exclusion crite-
molecules, DWI can be quantified using the apparent dif- ria. Propensity score matching was performed for this
fusion coefficient (ADC) derived from the Gaussian dif- cohort of patients. The control group included patients
fusion model. DWI and dynamic contrast-enhanced MRI without RTLI after IMRT who were matched 1:1 to each
showed potential for predicting the response to radiation case by sex (Fig. 1). Thus, 214 patients were included in
therapy for head and neck paragangliomas [14], promise this study, who were randomly allocated to a training set
in differentiating head and neck schwannomas and para- (135 patients) and a validation set (79 patients) at a ratio
gangliomas [15], detecting occult primary head and neck of 6:4.
squamous cell carcinoma [16], and survival prediction
Bao et al. Insights into Imaging (2022) 13:197 Page 3 of 11

Fig. 1 Diagram for inclusion of patients into the study. IMRT = intensity-modulated radiotherapy, NPC = nasopharyngeal carcinoma,
RTLI = radiation-induced temporal lobe injury

Clinical data every 6 months in years 3–5, and annually thereafter. The
Clinical information was analyzed in this study, includ- endpoint of this study was the development of RTLI or
ing sex, age, neutrophil-to-lymphocyte ratios, TNM the last follow-up for non-RTLI (> 36 months).
stage, pathologic subtype, treatment regimen, date
of pretreatment MRI scan, dosimetric parameters
including maximum dose for each temporal lobe, and Diagnostic criteria of RTLI
planning gross tumor volume including the primary The diagnostic criteria for temporal lobe injury (TLI)
nasopharyngeal tumor or enlarged retropharyngeal were as follows [28]: (a) white matter lesions with homo-
nodes. geneous high signal intensity on T2-weighted images and
low signal intensity on T1-weighted images without con-
trast enhancement, (b) contrast-enhanced lesions with or
Treatment regimen and follow‑up without necrosis on post-contrast T1-weighted images
All patients underwent a standard treatment regimen with heterogeneous signal abnormalities on T2-weighted
consisting of IMRT and concurrent or adjuvant chem- images, and (c) cysts or round or oval well-defined lesions
otherapy, with or without induction chemotherapy, with very high signal intensity on T2-weighted images
based on the National Comprehensive Cancer Network with a thin or imperceptible wall.
guidelines [26]. All patients were treated with IMRT
using the HiArt TomoTherapy system (Accuray, Sunny- Image acquisition
vale, CA) or a Varian-600CD linear accelerator (Varian All patients were examined using a 3.0-T MR scan-
Medical Systems, Palo Alto, CA) with a prescribed dose ner (GE Discovery MR 750; GE Healthcare, Chicago,
of 70–74 Gy in 30–33 fractions [27]. IL) with an 8-channel head and neck phased array coil.
After radiation therapy, patients routinely underwent DWI-MRI examinations were acquired axially using
follow-up MRI every 1–3 months during the first 2 years, a single-shot echo-planar imaging technique with a
Bao et al. Insights into Imaging (2022) 13:197 Page 4 of 11

spectral pre-saturation attenuated inversion-recovery fat- characteristic curve (ROC) analyses of significant find-
suppressed pulse sequence (repetition time/echo time, ings and combined analyses were performed to evaluate
4000/51 ms; bandwidth, 250 kHz; field of view, 24 cm; the predictive performance. Sensitivity, specificity, nega-
slice thickness, 5 mm; slice gap, 1 mm; number of excita- tive predictive value, and positive predictive value with
tions, 6.0). Diffusion gradients were applied with b values 95% confidence intervals (CIs) were calculated. The areas
of 0 and 800 s/mm2. under the curve (AUC) were compared using the DeLong
method.
Temporal lobe segmentation According to the results of the multivariate analysis, the
ADC maps were automatically calculated from b0 and predictive model was visualized as a nomogram to strat-
b800 images using the MRI console. MRI images were ify the individual risk of RTLI. A calibration curve was
reviewed by two radiologists (with 18 and 5 years of expe- used to describe the agreement between predicted and
rience in head and neck imaging, respectively). A tempo- observed RTLI occurrence probabilities. The Hosmer–
ral lobe ROI was drawn on the b = 800 s/mm2 DWI of the Lemeshow test was performed to explain the goodness-
pretreatment MRI using ITK-SNAP (version 3.6.0-RCI; of-fit of the multivariate logistic model. Decision curve
http://​www.​itk-​snap.​org). The ROI was manually deline- analysis (DCA) was used to evaluate the clinical useful-
ated along the boundaries of the middle and lower por- ness by quantifying the net benefits of the predictive
tions of the bilateral temporal lobes from the top level of model in the validation set. The optimum cutoff value of
the cerebral peduncle to the bottom of the temporal lobe the signature was identified using ROC analysis based on
(Additional file 1: Figure S1). One junior radiologist (Dan its association with the RTLI outcome. Accordingly, the
Bao) manually delineated and a senior neuroradiologist patients were divided into low- and high-risk groups, for
(Yanfeng Zhao) verified that both were blinded to clini- which the RTLI predictive outcomes were compared by
cal outcomes. The ROIs were then propagated to ADC ROC analysis in subgroups within clinical–pathologic
maps. Inter-observer segmentation variability was evalu- factors from the entire dataset.
ated using the Dice similarity coefficient (DSC) [29] in 50
randomly selected patients. Statistical analysis
Baseline characteristics were compared using the inde-
Histogram analysis pendent t test or the Mann–Whitney U test (for continu-
Quantitative analysis was performed by a radiologist with ous variables) and Pearson’s chi-square test or Fisher’s
five years of experience in head and neck MRI. For quan- exact test (for categorical variables). Statistical analyses
titative analysis, all ROIs were merged into the volume were conducted using SPSS (version 26.0; IBM, Armonk,
of interest in the ADC maps. Histogram features were NY) and R software (version 3.4.4; R Foundation, Vienna,
extracted using the non-open source software Analysis Austria). A two-sided p value less than 0.05 indicated a
Kit (Version v3.0.1. A, GE Healthcare) with the following significant difference.
parameters: skewness, kurtosis, entropy, energy, range,
uniformity, mean, median, minimum, maximum, vari- Results
ance, 10th percentile, 90th percentile, interquartile range Patient characteristics
(IQR), mean absolute deviation, robust mean absolute A total of 214 patients with pathologically proven NPC
deviation, root mean square, and total energy. and IMRT treatment (median age 47.50 years; IQR 37.8–
56 years; 69 females) were included, including 135 in the
Development and validation groups training set and 79 in the validation set. During follow-
The training cohort used 60% of the dataset and the vali- up, 107 patients were confirmed with RTLI (bilateral,
dation cohort used the remaining 40%. Univariate and 23; left, 39; right, 45). The median duration of follow-up
multivariate logistic regression analyses were performed from the pretreatment MRI was 33.4 months (IQR 26.2–
using the training data to determine the predictive fac- 41.9 months) in the RTLI group and 61.4 months (IQR
tors for RTLI. The backward stepwise was used to select 53.5–68.5 months) in the non-RTLI group. The baseline
variables included in the best models, and the Akaike’s clinical characteristics are given in Table 1. No significant
information as the stopping criterion [30, 31]. A function differences were observed between the training and vali-
based on the variance inflation factor was used to check dation groups (all p > 0.05). The rates of RTLI occurrence
the collinearity of the variables included in the regression were 55.56% (75/135) and 40.50% (32/79) in the training
equation, with a variance inflation factor greater than and validation cohorts, respectively.
10 indicating multicollinearity [32]. Receiver operating
Bao et al. Insights into Imaging (2022) 13:197 Page 5 of 11

Table 1 Characteristics of patients in the training and validation cohorts

Characteristic Training cohort Validation cohort p value
(n = 135) (n = 79)

Age (y)* 46.67 ± 12.81 (9–73) 44.43 ± 13.59 (11–69) 0.23

Sex
Male 88 57 0.29
Female 47 22
NLRs (mean ± SD) (range) * 3.22 ± 4.38 (0.43–48.63) 3.43 ± 2.31 (0.90–14.18) 0.70
T stage 0.34
T1 8 4
T2 8 6
T3 62 45
T4 57 24
N stage 0.53
N0 16 12
N1 47 28
N2 51 32
N3 21 7
TNM stage 0.22
I 1 1
II 6 3
III 56 44
IV 72 31
Pathology 0.98
Differentiated 43 25
Undifferentiated 92 54
Treatment 0.51
A 17 16
B 51 24
C 31 14
D 19 13
E 8 4
F 9 8
PGTVNX (mean ± SD) (range) (Gy) * 73.51 ± 1.25 (67.72–74.20) 73.72 ± 0.87 (69.96–73.92) 0.20
LDmax (mean ± SD) (range) (Gy) * 68.41 ± 5.68 (53.72–78.16) 68.11 ± 5.94 (52.86–86.07) 0.72
RDmax (mean ± SD) (range) (Gy) * 68.78 ± 5.28 (55.16–78.57) 68.62 ± 5.67 (55.93–86.07) 0.84
*
Data are mean ± standard deviation; data in parentheses are range. Treatment: A-Radiotherapy, B-Concurrent Chemoradiotherapy, C-Concurrent
Chemoradiotherapy + Targeted Therapy, D-Induction chemotherapy + Concurrent Chemoradiotherapy, E-Induction chemotherapy + Concurrent
Chemoradiotherapy + Targeted therapy, F-Radiotherapy + Targeted therapy. p > 0.05 suggests no significant difference between the subjects in the two cohorts.
­LDmax = maximum dose of left temporal lobe, NLRs = neutrophil-to-lymphocyte ratios, ­PGTVNX = planning gross tumor volume included the primary nasopharyngeal
tumor or enlarged retropharyngeal nodes, ­RDmax = maximum dose of right temporal lobe, TNM = tumor-node-metastasis

Temporal lobe segmentation analysis for predicting RTLI occurrence in the training
In assessing the reliability of segmentation, the intra- cohort (Additional file 1: Table S1).
reader Dice value was 0.981 ± 0.002 (range 0.979–0.982).
Multivariate analysis of histogram parameters
Univariate analysis of histogram parameters Statistically significant variables in the univariate analy-
Of the 18 histogram parameters derived from ADC maps, sis were included in the multivariate logistic regression
energy, kurtosis, maximum, minimum, range, skew- analysis. The final best prediction model selected by
ness, and total energy were significant in the univariate backward stepwise elimination with Akaike information
Bao et al. Insights into Imaging (2022) 13:197 Page 6 of 11

Table 2 Results of multivariate logistic regression histogram parameters in the training set
Variable β SE Wald p OR 95% CI
Lower Upper

Skewness 0.82 1.48 0.31 0.58 2.27 0.13 46.15

Kurtosis − 0.39 0.21 3.62 0.06* 0.67 0.43 0.99
Energy 6.42E−11 1.10E−10 0.34 0.56 1 1 1
Range − 0.01 0.01 3.49 0.06* 0.99 0.98 1.00
Minimum − 0.01 0.01 1.51 0.22 0.99 0.97 1.01
Maximum 0.01 0.01 3.75 0.05* 1.01 1.00 1.02
Total energy − 1.02E−10 2.81E−11 13.20 < 0.001* 1 1 1
CI = confidence interval, OR = odds ratio, SE = standard error. * indicates significant difference

Table 3 Clinical predictive factors according to univariate and multivariate logistic regression in the training set
Univariate analysis Multivariate analysis
Coefficient OR (95% CI) p Coefficient OR (95% CI) p

Age 0.01 1.01 (0.986–1.039) 0.37

Sex − 0.15 0.86 (0.152–1.575) 0.69
NLRs 0.13 1.14 (0.932–1.356) 0.21
T stage 1.12 3.07 (2.516–3.618) < 0.001 0.88 2.41 (1.204–5.724) 0.02*
N stage − 0.18 0.83 (0.450–1.218) 0.35
TNM stage 0.96 2.60 (1.996–3.211) 0.002 0.18 1.19 (0.450–2.883) 0.70
Pathology − 0.02 0.98 (0.256–1.714) 0.97
Treatment − 0.18 0.83 (0.577–1.085) 0.15
PGTVNX 0.35 1.42 (1.105–1.735) 0.03 0.19 1.21 (0.878–1.771) 0.26
LDmax 0.07 1.07 (1.006–1.131) 0.04 − 0.02 0.98 (0.898–1.073) 0.70
RDmax 0.10 1.11 (1.037–1.176) 0.004 0.08 1.08 (0.984–1.194) 0.11
CI = confidence interval, ­LDmax = maximum dose of left temporal lobe, NLRs = neutrophil-to-lymphocyte ratios, OR = odds ratio, P
­ GTVNX = planning gross tumor
volume included the primary nasopharyngeal tumor or enlarged retropharyngeal nodes, R ­ Dmax = maximum dose of right temporal lobe, TNM = tumor-node-
metastasis. * indicates significant difference

criteria as the stopping rule included kurtosis (p = 0.06), log (Rad − scrore) =10.34 ± 0.28 × Kurtosis + 0.005
maximum energy (p = 0.05), range (p = 0.06), and total × Maximum ± 0.004 × Range
energy (p < 0.001) (Table 2). ± 8.28E − 11 × Total Energy

Clinical feature selection A difference in Rad-score was present between the

The results of the univariate and multivariate logistic RTLI and non-RTLI groups in the training set (median
analyses for clinical and dosimetric features are pre- [IQR], 0.97 [0.82–0.99] vs. 0.11 [0.03–0.31]; p < 0.001)
sented in Table 3. In the multivariate regression analysis, and confirmed in the validation cohort (median [IQR],
the T stage was a significant clinical predictor of RTLI. 0.95 [0.56–0.99] vs. 0.08 [0.02–0.25]; p < 0.001) (Addi-
tional file 1: Figure S2). The Rad-score yielded an AUC of
Development and validation of predictive models 0.95 (95% CI 0.91–0.98) in the training cohort and 0.89
Predictive model derived from ADC map (95% CI 0.81–0.97) in the validation cohort.
Based on the results of the multivariate logistic analysis, four
histogram parameters were integrated into a Rad-score. Combination of clinical and histogram findings
The Rad-score was calculated using a linear combination of The variance inflation factors of the five potential pre-
these histogram parameters based on their respective coef- dictors ranged from 1.004 to 1.278, indicating no mul-
ficients. The calculation formula is as follows: ticollinearity. A combined model incorporating two
Bao et al. Insights into Imaging (2022) 13:197 Page 7 of 11

Fig. 2 Nomogram and calibration curves. a A nomogram was developed in training cohort, with Rad-score and T stage incorporated. Calibration
curves of the nomogram in (b) training and (c) validation cohorts

independent predictors (Rad-score and T stage) was 0.89) was slightly higher than that of the combined model
developed and presented as a nomogram (Fig. 2A (AUC, 0.87) in the validation cohort, but the difference
and Additional file 1: Table S2). The calibration plots was not significant (p = 0.47).
showed that the predicted RTLI probabilities of the After obtaining the risk scores based on the combined
combined model were in excellent agreement with model, an optimal threshold of 0.55 was determined
actual observations (Fig. 2B and C). The Hosmer– according to the maximized Youden index from the train-
Lemeshow test of model calibration showed no depar- ing cohort. Accordingly, all patients were classified into
ture from a good fit, with no statistical significance high- (Rad-score ≥ 0.55) and low-risk (Rad-score < 0.55)
(p = 0.23). RTLI groups (Additional file 1: Figure S3). Accord-
ing to the proposed risk classifier, the combined model
Performance and validation of predictive models achieved a sensitivity of 81.3% and a specificity of 82.0%
The ROC curves of the Rad-score and the combined for predicting RTLI in the validation cohort, whereas the
model are shown in Fig. 3 and Table 4. Compared with positive and negative predictive values were 81.3% and
the T stage alone (AUC, 0.63 [95% CI 0.52–0.74]), both 87.2%, respectively. Moreover, when the patients were
the Rad-score (p < 0.001) and the combined model stratified based on clinicopathological factors, the overall
(P < 0.001) exhibited better predictive performance for diagnostic accuracy of the risk classifier was excellent in
RTLI after IMRT. The AUC value of the Rad-score (AUC, all subgroups (AUC, 0.79–0.98). The performance of the
Bao et al. Insights into Imaging (2022) 13:197 Page 8 of 11

Fig. 3 Performances of two models in training cohort and validation cohort, respectively. a, b Rad-score, including four histogram parameters. c, d
Combined model, integrated T stage and four histogram parameters

constructed combined model in patients within different Discussion

clinicopathological subgroups is presented in Additional In this study, we assessed the capability of pretreatment
file 1: Figure S4. histogram parameters in predicting RTLI in patients with
Additionally, DCA indicated that the Rad-score or NPC after IMRT. Our results showed that the Rad-score
combined model achieved moderately better net ben- integrating the four histogram parameters was an inde-
efits than clinical factors alone (Additional file 1: Fig- pendent predictive factor of RTLI and showed a favorable
ure S5). The positive values of integrated discrimination predictive performance. A nomogram combining T stage
improvement (53.0% [95% CI 0.45–0.61], p < 0.001) and and Rad-score as a quantitative tool could facilitate RTLI
net reclassification index (69.0% [95% CI 0.55–0.83], risk stratification and clinical decision-making in NPC
p < 0.001) are shown. patients treated with IMRT.
Bao et al. Insights into Imaging (2022) 13:197 Page 9 of 11

Table 4 Predictive performances of two models in predicting the radiation-induced temporal lobe injury in the training and
validation cohort
Model AUC​ 95% CI Sensitivity Specificity PPV NPV
Lower Upper

Model—ADC
Training cohort 0.95 0.91 0.98 0.85 (64/75) 0.93 (56/60) 0.94 (64/68) 0.83
(56/67)
Validation cohort 0.89 0.81 0.97 0.75 (24/32) 0.89 (42/47) 0.83 (24/29) 0.89
(42/47)
Model—combined
Training cohort 0.96 0.92 0.99 0.88 (66/75) 0.93 (56/60) 0.94 (66/70) 0.86
(56/65)
Validation cohort 0.87 0.78 0.96 0.81 (26/32) 0.82 (41/50) 0.81 (26/32) 0.87
(41/47)
AUC = area under the receiver operating characteristic curve, CI = confidence interval, NPV = negative predictive value, PPV = positive predictive value. *Features used
for the Model-ADC are kurtosis, maximum, range, and total energy

While histogram analysis has been successfully dem- showing the maximum ADC may reflect the lowest cellu-
onstrated in various organs [33–35] and the predictive lar area within the temporal lobe. However, the results of
potential of radiomics features has been explored in RTLI our study indicate that the maximum value was positively
in NPC patients [21, 22, 24, 25], the utility of histogram correlated with RTLI occurrence, which contradicts our
parameters in predicting RTLI still needs to be further previous hypothesis. It is difficult to provide a reasonable
investigated. As previously suggested, image heteroge- explanation based on the current research, and further
neity is correlated with physiological heterogeneity [20]. research on the histological proof of temporal lobe het-
We found that some of the histogram parameters derived erogeneity and RTLI occurrence is required.
from ADC mapping of the temporal lobes were associ- Compared with other studies that established predic-
ated with RTLI occurrence. Kurtosis yielded the highest tion models for predicting RTLI based on pretreatment
(negative) coefficient in selected histogram parameters, MRI parameters, the AUC of the model in this study was
which is a measure of the “peakedness” of the distribution lower than that of the prediction model based on radi-
of values in the image ROI [36]. A lower kurtosis implies omics features extracted from contrast-enhanced T1- or
that the mass of the distribution is concentrated toward fat-suppressed T2-weighted MRI (AUC, 0.89 vs. 0.92)
a spike near the mean value, implying that the temporal [24], and higher than that of the proposed model based
lobes were more functionally homogeneous. The range on features extracted from T1- and T2-weighted MRI
represents the range of gray values in the voxel of inter- (AUC, 0.82) [25]. Therefore, the prediction model based
est (VOI), whereas total energy refers to the value of the on ADC histogram parameters showed persuasive per-
energy feature scaled by the volume of the voxel in m ­ m3 formance in predicting RTLI in NPC, and the feasibility
[36]. Higher values of range and total energy may indicate of a multiparametric MRI model to predict RTLI should
the complexity of the tissue components. In this study, be explored in future studies.
lower values of kurtosis, range, and total energy, which Concerning clinical predictors, the T stage was identi-
led to a higher Rad-score, were associated with patients fied as a clinical predictor for RTLI in our study, which
more prone to developing RTLI. As no histological was consistent with the findings of Wen et al. [9] and
proof of the precise mechanism that leads to RTLI and Guan et al. [8]. This study demonstrated that the nomo-
its association with the heterogeneity of temporal lobes gram incorporating histogram parameters and T stage
was available by this point, we can hypothesize that less yielded satisfactory predictive performance, with favora-
heterogeneous image textures corresponded to the abun- ble calibration and positive net reclassification improve-
dance of cells in the VOI of the temporal lobe, with the ment. DCA also illustrated that both the combined model
cells arranged tightly and regularly [37, 38]. Furthermore, and the Rad-score outperformed the T stage alone in
the abundant blood supply and high oxygen demand of predicting RTLI occurrence, but interestingly, the com-
the corresponding temporal lobe, which means greater bined model did not significantly improve the predictive
sensitivity to radiotherapy [39, 40], were more prone to performance compared to the Rad-score; a similar lack of
developing RTLI. It is well known that a high cell den- improvement in the extended model compared with indi-
sity is associated with a low ADC [41, 42], and the region vidual components has been previously observed in brain
Bao et al. Insights into Imaging (2022) 13:197 Page 10 of 11

tumors [43] and is attributed to the high intra-correlation Abbreviations

ADC: Apparent diffusion coefficient; AUC​: Area under the curve; CI: Confi-
of ADC histogram parameters. dence interval; DCA: Decision curve analysis; DTI: Diffusion tensor imaging;
This study had several limitations. First, selection bias DWI: Diffusion-weighted imaging; IMRT: Intensity-modulated radiotherapy;
exists because of the retrospective design of the study. IQR: Interquartile range; MRI: Magnetic resonance imaging; NPC: Nasopharyn-
geal carcinoma; ROC: Receiver operating characteristic; ROI: Region of interest;
Removing a significant portion of patients for a variety of RTLI: Radiation-induced temporal lobe injury; TLI: Temporal lobe injury; VOI:
reasons may have generated bias. Second, patients with- Voxel of interest.
out RTLI after IMRT were included by a propensity score
matching at 1:1 to each RTLI case by gender in this study. Supplementary Information
Third, patients without RTLI after IMRT were included The online version contains supplementary material available at https://​doi.​
by propensity score matching at 1:1 for each RTLI case by org/​10.​1186/​s13244-​022-​01338-w.
sex in this study. The preferred design should include all
Additional file 1. Supplementary tables and figures.
patients to ensure that no bias is introduced for all relevant
risk factors and outcomes; however, the low incidence of
RTLI in the clinic and the long follow-up time needed Author contributions
DB, YFZ, and DHL contributed to conception and design. DHL and YFZ con-
for RTLI outcomes in NPC may make the research diffi- tributed to administrative support. DB, YFZ, WLW, HXZ, MY, LL, ML, XMZ, and
cult to implement. Fourth, we performed DWI using only DHL, contributed to provision of study materials or patients. DB, YFZ, and WLW
two b values of 0 and 800 s/mm2 on a 3.0-T MRI machine contributed to collection and assembly of data. DB, YFZ, HXZ, MY, ML, and
LL contributed to data analysis and interpretation. DB performed writing—
from a single manufacturer; further studies on DWI with original draft. DB, YFZ, HXZ, XMZ, and DHL contributed to draft editing and
multiple b values with various MRI scanners and tech- visualization. All authors read and approved the final manuscript.
niques may contribute to the generalizability of the results.
Funding
Finally, the dosimetric parameters included in this study This work was supported by the Non-profit Central Research Institute Fund of
were limited and not independent predictors of RTLI in Chinese Academy of Medical Sciences (2019XK320073).
the training set; therefore, we did not include them in the
Availability of data and materials
final prediction model. Although patients with NPC who The datasets used and/or analyzed during the current study are available from
received radiotherapy were one of the causative factors for the corresponding author upon reasonable request.
possible RTLI, patients included in this study were treated
with IMRT and standardized treatment according to their Declarations
conditions. Therefore, in this case, the predictive model
Ethics approval and consent to participate
still had predictive potential for RTLI in patients with The study was approved by the Ethics Committee of the National Cancer
NPC who received IMRT. The prediction model used in Center/Cancer Hospital, Chinese Academy of Medical Sciences, and Peking
this study was based on MRI obtained before treatment. Union Medical College, and individual consent for this retrospective analysis
was waived (institutional ethics approval number 21/278–2949).
Patients receiving radiotherapy may have subtle changes
that are invisible to the naked eye and can be detected Consent for publication
early using radiomics. Prediction models based on MRI Not applicable.
obtained immediately after IMRT may yield different Competing interests
results [21, 22, 24]. In general, the feasibility of histograms, The authors declare that they have no competing interests.
clinical and dosimetric parameters, white and gray matter,
Author details
and their associated variables were considered separately, 1
Department of Radiology, National Cancer Center/National Clinical Research
and MRI after IMRT to predict RTLI should be explored Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences
in future studies, especially prospective studies with larger and Peking Union Medical College, 17 Panjiayuan Nanli, Chaoyang District,
Beijing 100021, China. 2 Medical Imaging Center, Liaocheng Tumor Hospital,
sample sizes at multicenter institutions. Shandong 252000, China. 3 Department of Radiation Oncology, National
Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital,
Conclusions Chinese Academy of Medical Sciences and Peking Union Medical College,
17 Panjiayuan Nanli, Chaoyang District, Beijing 100021, China. 4 Department
In summary, our study revealed that histogram parame- of Radiology, National Cancer Center/National Clinical Research Center
ters of ADC mapping based on temporal lobes are related for Cancer/Cancer Hospital & Shenzhen Hospital, Chinese Academy of Medical
to RTLI occurrence in patients with NPC after IMRT. Sciences and Peking Union Medical College, Shenzhen 518116, China.
The combined model achieved satisfactory pretreatment Received: 29 June 2022 Accepted: 20 November 2022
risk prediction of RTLI in patients with NPC, which may
help stratify high-risk patients who require intensive fol-
low-up and effective treatment guidance.
Bao et al. Insights into Imaging (2022) 13:197 Page 11 of 11

References 23. Wang HZ, Qiu SJ, Lv XF et al (2012) Diffusion tensor imaging and 1H-MRS
1. Chen YP, Chan ATC, Le QT, Blanchard P, Sun Y, Ma J (2019) Nasopharyngeal study on radiation-induced brain injury after nasopharyngeal carcinoma
carcinoma. Lancet 394:64–80 radiotherapy. Clin Radiol 67:340–345
2. Greene-Schloesser D, Robbins ME, Peiffer AM, Shaw EG, Wheeler KT, Chan 24. Bao D, Zhao Y, Li L et al (2022) A MRI-based radiomics model predicting
MD (2012) Radiation-induced brain injury: a review. Front Oncol 2:73 radiation-induced temporal lobe injury in nasopharyngeal carcinoma. Eur
3. Tang Y, Luo D, Rong X, Shi X, Peng Y (2012) Psychological disorders, cogni- Radiol. https://​doi.​org/​10.​1007/​s00330-​022-​08853-w
tive dysfunction and quality of life in nasopharyngeal carcinoma patients 25. Bin X, Zhu C, Tang Y et al (2022) Nomogram based on clinical and radiom-
with radiation-induced brain injury. PLoS One 7:e36529 ics data for predicting radiation-induced temporal lobe injury in patients
4. Lee N, Harris J, Garden AS et al (2009) Intensity-modulated radiation with non-metastatic stage T4 nasopharyngeal carcinoma. Clin Oncol (R
therapy with or without chemotherapy for nasopharyngeal carcinoma: Coll Radiol). https://​doi.​org/​10.​1016/j.​clon.​2022.​07.​007
radiation therapy oncology group phase II trial 0225. J Clin Oncol 26. Adelstein D, Gillison ML, Pfister DG et al (2017) NCCN guidelines insights:
27:3684–3690 head and neck cancers, version 2.2017. J Natl Compr Canc Netw
5. Zhou GQ, Yu XL, Chen M et al (2013) Radiation-induced temporal lobe 15:761–770
injury for nasopharyngeal carcinoma: a comparison of intensity-modu- 27. Ren W, Liang B, Sun C et al (2021) Dosiomics-based prediction of radi-
lated radiotherapy and conventional two-dimensional radiotherapy. PLoS ation-induced hypothyroidism in nasopharyngeal carcinoma patients.
One 8:e67488 Phys Med 89:219–225
6. Liang SB, Wang Y, Hu XF et al (2017) Survival and toxicities of IMRT based 28. Wang YX, King AD, Zhou H et al (2010) Evolution of radiation-induced
on the RTOG protocols in patients with nasopharyngeal carcinoma from brain injury: MR imaging-based study. Radiology 254:210–218
the endemic regions of China. J Cancer 8:3718–3724 29. Duane F, Aznar MC, Bartlett F et al (2017) A cardiac contouring atlas for
7. Abayomi OK (2002) Pathogenesis of cognitive decline following thera- radiotherapy. Radiother Oncol 122:416–422
peutic irradiation for head and neck tumors. Acta Oncol 41:346–351 30. Moons KG, Altman DG, Reitsma JB et al (2015) Transparent Reporting of
8. Guan W, Xie K, Fan Y et al (2020) Development and validation of a a multivariable prediction model for individual prognosis or diagnosis
nomogram for predicting radiation-induced temporal lobe injury in (TRIPOD): explanation and elaboration. Ann Intern Med 162:W1-73
nasopharyngeal carcinoma. Front Oncol 10:594494 31. Sauerbrei W, Boulesteix AL, Binder H (2011) Stability investigations of
9. Wen DW, Lin L, Mao YP et al (2021) Normal tissue complication probabil- multivariable regression models derived from low- and high-dimensional
ity (NTCP) models for predicting temporal lobe injury after intensity-mod- data. J Biopharm Stat 21:1206–1231
ulated radiotherapy in nasopharyngeal carcinoma: a large registry-based 32. Kim JH (2019) Multicollinearity and misleading statistical results. Korean J
retrospective study from China. Radiother Oncol 157:99–105 Anesthesiol 72:558–569
10. Huang J, Kong FF, Oei RW, Zhai RP, Hu CS, Ying HM (2019) Dosimetric 33. Perucho JAU, Wang M, Tse KY et al (2021) Association between MRI
predictors of temporal lobe injury after intensity-modulated radiotherapy histogram features and treatment response in locally advanced cervical
for T4 nasopharyngeal carcinoma: a competing risk study. Radiat Oncol cancer treated by chemoradiotherapy. Eur Radiol 31:1727–1735
14:31 34. Wu LF, Rao SX, Xu PJ et al (2019) Pre-TACE kurtosis of ADCtotal derived
11. Feng M, Huang Y, Fan X, Xu P, Lang J, Wang D (2018) Prognostic variables from histogram analysis for diffusion-weighted imaging is the best inde-
for temporal lobe injury after intensity modulated-radiotherapy of naso- pendent predictor of prognosis in hepatocellular carcinoma. Eur Radiol
pharyngeal carcinoma. Cancer Med 7:557–564 29:213–223
12. Lee AW, Cheng LO, Ng SH et al (1990) Magnetic resonance imaging in the 35. Kang Y, Choi SH, Kim YJ et al (2011) Gliomas: Histogram analysis of
clinical diagnosis of late temporal lobe necrosis following radiotherapy apparent diffusion coefficient maps with standard- or high-b-value
for nasopharyngeal carcinoma. Clin Radiol 42:24–31 diffusion-weighted MR imaging–correlation with tumor grade. Radiology
13. Padhani AR, Liu G, Koh DM et al (2009) Diffusion-weighted magnetic 261:882–890
resonance imaging as a cancer biomarker: consensus and recommenda- 36. van Griethuysen JJM, Fedorov A, Parmar C et al (2017) Computational
tions. Neoplasia 11:102–125 radiomics system to decode the radiographic phenotype. Cancer Res
14. Ota Y, Liao E, Kurokawa R et al (2021) Diffusion-weighted and dynamic 77:e104–e107
contrast-enhanced MRI to assess radiation therapy response for head and 37. Yang J, Xu Z, Gao J et al (2018) Evaluation of early acute radiation-induced
neck paragangliomas. J Neuroimaging 31:1035–1043 brain injury: hybrid multifunctional MRI-based study. Magn Reson Imag-
15. Ota Y, Liao E, Capizzano AA et al (2021) Diagnostic role of diffusion- ing 54:101–108
weighted and dynamic contrast-enhanced perfusion MR imaging in 38. Bashir U, Foot O, Wise O et al (2018) Investigating the histopathologic cor-
paragangliomas and schwannomas in the head and neck. AJNR Am J relates of 18F-FDG PET heterogeneity in non-small-cell lung cancer. Nucl
Neuroradiol 42:1839–1846 Med Commun 39:1197–1206
16. Martens RM, Stappen RV, Koopman T et al (2020) The additional value of 39. Jia Y, Weng Z, Wang C et al (2017) Increased chemosensitivity and
ultrafast DCE-MRI to DWI-MRI and 18F-FDG-PET to detect occult primary radiosensitivity of human breast cancer cell lines treated with novel
head and neck squamous cell carcinoma. Cancers (Basel) 12 functionalized single-walled carbon nanotubes. Oncol Lett 13:206–214
17. Martens RM, Koopman T, Lavini C et al (2021) Multiparametric functional 40. Xie Y, Huang H, Guo J, Zhou D (2018) Relative cerebral blood volume is
MRI and (18)F-FDG-PET for survival prediction in patients with head and a potential biomarker in late delayed radiation-induced brain injury. J
neck squamous cell carcinoma treated with (chemo)radiation. Eur Radiol Magn Reson Imaging 47:1112–1118
31:616–628 41. Sugahara T, Korogi Y, Kochi M et al (1999) Usefulness of diffusion-
18. Koontz NA, Wiggins RH 3rd (2017) Differentiation of benign and malig- weighted MRI with echo-planar technique in the evaluation of cellularity
nant head and neck lesions with diffusion tensor imaging and DWI. AJR in gliomas. J Magn Reson Imaging 9:53–60
Am J Roentgenol 208:1110–1115 42. Chen J, Xia J, Zhou YC et al (2005) Correlation between magnetic
19. Liu X, Han C, Wang H et al (2021) Fully automated pelvic bone segmenta- resonance diffusion weighted imaging and cell density in astrocytoma.
tion in multiparameteric MRI using a 3D convolutional neural network. Zhonghua Zhong Liu Za Zhi 27:309–311
Insights Imaging 12:93 43. Kyriazi S, Collins DJ, Messiou C et al (2011) Metastatic ovarian and primary
20. Just N (2014) Improving tumour heterogeneity MRI assessment with peritoneal cancer: assessing chemotherapy response with diffusion-
histograms. Br J Cancer 111:2205–2213 weighted MR imaging–value of histogram analysis of apparent diffusion
21. Zhang B, Lian Z, Zhong L et al (2020) Machine-learning based MRI coefficients. Radiology 261:182–192
radiomics models for early detection of radiation-induced brain injury in
nasopharyngeal carcinoma. BMC Cancer 20:502
22. Hou J, Li H, Zeng B et al (2022) MRI-based radiomics nomogram for Publisher’s Note
predicting temporal lobe injury after radiotherapy in nasopharyngeal Springer Nature remains neutral with regard to jurisdictional claims in pub-
carcinoma. Eur Radiol 32:1106–1114 lished maps and institutional affiliations.