Pretreatment MR imaging radiomics signatures for response prediction to induction chemotherapy in patients with nasopharyngeal carcinoma

European Journal of Radiology 98 (2018) 100–106

Contents lists available at ScienceDirect

European Journal of Radiology journal homepage: www.elsevier.com/locate/ejrad

Research article

Pretreatment MR imaging radiomics signatures for response prediction to induction chemotherapy in patients with nasopharyngeal carcinoma Guangyi Wanga,1, Lan Hea,1, Cai Yuanb, Yanqi Huanga, Zaiyi Liua, a b

⁎,2

, Changhong Lianga,

T

⁎,2

Department of Radiology, Guangdong General Hospital, Guangdong Academy of Medical Sciences, Guangzhou, 510080, China Internal Medicine Residency, Florida Hospital, Orlando, FL, 32804, USA

A R T I C L E I N F O

A B S T R A C T

Keywords: Nasopharyngeal carcinoma Radiomics signature Predictor MRI Induction chemotherapy

Purpose: This study aimed to investigate the capability of magnetic resonance (MR) imaging radiomics signatures for pretreatment prediction of early response to induction chemotherapy in patients with nasopharyngeal carcinoma (NPC). Materials and methods: This was a retrospective study consisting of 120 patients with biopsy-proven NPC (stage II–IV). Texture features were extracted from the pretreatment morphological MR images for each case. Radiomics signatures were obtained with the least absolute shrinkage and selection operator method (LASSO) logistic regression model. The association between the radiomics signatures and the early response to induction chemotherapy was explored. Results: From the contrast-enhanced T1-weighted MR imaging (CE T1WI), 5 features were selected by the LASSO model. The radiomics signature categorised patients with NPC into response and nonresponse groups (P < 0.001). The area under the receiver operating characteristic curve values (AUC), sensitivity, specificity, positive predictive value (PPV) and negative predictive value(NPV) were 0.715(95% CI 0.699–0.731), 0.940, 0.500, 0.568 and 0.897 respectively, where non-responders are true-positives. The AUC of 1000 bootstrap internal validation was 0.715. Furthermore, when the features of T1-weighted MR imaging (T1WI), T2-weighted MR imaging (T2WI), T2-weighted fat-suppressed MR imaging (T2WI FS) and CE T1WI were analysed together, 15 features were selected to develop the radiomics signature. The performance of this radiomics signature was better than that developed only from CE T1WI (P < 0.05). The AUC value was 0.822(95% CI 0.809–0.835) with sensitivity of 0.980, specificity of 0.529, PPV of 0.593 and NPV of 0.949. The AUC of 1000 bootstrap analysis was 0.821. From T1WI, T2WI, and T2WI FS images separately, no valuable features were selected. Conclusions: Pretreatment morphological MR imaging radiomics signatures can predict early response to induction chemotherapy in patients with NPC.

1. Introduction Nasopharyngeal carcinoma (NPC) prevails in Southeast Asia. There were about 86,700 new cases of NPC and 50,800 deaths in 2012 [1]. Radiotherapy is the primary treatment regimen for all stages NPC [2]. Patients with advanced stages can benefit from additional chemotherapy, which can significantly improve overall survival (OS) [3–11]. Therefore, concurrent chemoradiotherapy (CCRT) was recommended as the standard treatment for patients with NPC with advanced stages by the guidelines of the National Comprehensive Cancer Network (NCCN). Recently, it had been proven that addition of induction

chemotherapy (IC) followed by CCRT can significantly improve failurefree survival in advanced NPC compared with CCRT alone [12]. However, in clinical practice, not all patients with NPC respond well to IC [13,14]. Predicting the response to IC may play an important role in individualising the therapeutic strategy for NPC patients. Furthermore, the prediction of response to IC may help avoid unnecessary side effects caused by ineffective IC. Moreover, some studies have revealed that different responses to IC in patients with NPC had prognostic values for survival outcomes [14,15]. Therefore, there would be an advantage in finding reliable and practical predictive markers that can predict the response to IC for patients with NPC before treatment. To date, at least 3 imaging modalities have been tried to predict the

⁎

Corresponding authors. E-mail addresses: [email protected] (Z. Liu), [email protected] (C. Liang). These authors contributed equally to this work and share first authorship. 2 These authors contributed equally to this work and share last authorship. 1

https://doi.org/10.1016/j.ejrad.2017.11.007 Received 23 May 2017; Received in revised form 19 October 2017; Accepted 13 November 2017 0720-048X/ © 2017 Published by Elsevier Ireland Ltd.

European Journal of Radiology 98 (2018) 100–106

G. Wang et al.

the disappearance of all target lesions. Partial response (PR) was defined as at least a 30% decrease in the sum of the diameters of the target lesions, taking as reference the baseline sum diameters. Progressive disease (PD) was defined as at least a 20% increase in the sum of the diameters of the target lesions, taking as reference the smallest sum on study. Stable disease (SD) was defined as neither sufficient shrinkage to qualify for PR nor sufficient increase to qualify for PD, taking as reference the smallest sum diameters. Both CR and PR patients were defined as response patients; SD and PD were defined as nonresponse patients.

response to IC in patients with NPC. These include diffusion-weighted MR imaging (DWI) with an apparent diffusion coefficient (ADC) map [16,17], dynamic contrast-enhanced MR imaging (DCE-MRI) [18,19], and intravoxel incoherent motion diffusion-weighted MR imaging (IVIM-DWI) [20,21]. According to the results of these studies, the efficacy of DWI and DCE-MRI for predicting response to IC was controversial. Although the D (pure diffusion coefficient) values developed from IVIM-DWI may potentially predict the response to IC in patients with NPC before treatment, with sensitivity and specificity ranging from 0.647 to 0.658 and from 0.722 to 0.818, respectively [20,21], the specificity of these studies was relatively low. Recently, some studies have indicated that quantitative MR imaging radiomics shows promise in tumour prognosis. Through MR imaging radiomics analysis, Nie et al. developed models to predict the pathological response after chemoradiotherapy for rectal cancer [22]. Breast cancer recurrence risk was predicted successfully by analysing MR imaging radiomics signatures [23]. Jia et al. tried to predict a treatment response to chemoradiotherapy in NPC by texture analysis based on MR images [24]. There was a study dealing with MR imaging radiomics signature and prediction of survival in NPC [25]. Their results are encouraging. The aim of our study was to evaluate the ability of MR imaging radiomics for pretreatment prediction of the response to IC in patients with NPC.

2.5. MR images acquisition protocol All patients underwent nasopharynx and cervical region contrastenhanced MR examination using head and neck coils with 1.5T MR scanners (Magnetom Espree, Siemens Medical Solutions, Erlangen, Germany, or Achieva, Philips Healthcare, Best, the Netherlands). T1weighted fast spin-echo images in the axial plane (repetition time [TR] = 450 ms; echo time [TE] = 8.8 ms, flip angle = 90°; matrix = 256 × 168, slice thickness = 4 mm; spacing between slices = 5 mm), T2-weighted fast spin-echo MR images in the axial plane (TR = 6000 ms; TE = 95 ms; flip angle = 90°; matrix = 256 × 168; slice thickness = 4 mm; spacing between slices = 5 mm), and T2weighted fat-suppressed spin-echo sequence (TR = 6360 ms; TE = 95 ms; flip angle = 90°; matrix = 256 × 168 ; slice thickness = 4 mm; spacing between slices = 5 mm) were obtained before contrast was administrated. After bolus injection of contract (0.1 mmol/kg body weight; Magnevist, Schering, Berlin, Germany), axial T1-weighted fast spin-echo sequences were performed (with the same parameters as before contrast).

2. Materials and methods 2.1. Patient selection and clinical characteristics Data were obtained from our institute. The research protocol was approved by the institutional review board. Since all the data were deidentified, informed patient consent was waived. Between August 2009 to May 2016, 155 consecutive patients who were biopsy-proven NPC at our institute were recruited retrospectively. The exclusion criteria for our study were: (1) unmeasurable (< 5 mm) nasopharynx lesion on pre-treatment MR(n = 7). (2) chosed to get treatment elsewhere (n = 12). (3) without pre- or post-IC MR examination (n = 7). (4) no IC or had inadequate IC (n = 9). A total of 120 (mean age 46.81 ± 10.89, range 22–70, male = 95, female = 25, stage II = 10, stage III = 70, stage IV = 40) were studied in this research.

2.6. Images retrieval procedure The DICOM format images of T1WI, T2WI, T2WI FS, and CE T1WI for each case were retrieved from the PACS (Carestream, Ontario, Canada). 2.7. Image intensity normalisation Because signal intensity in morphological MR images is a relative number and not an absolute number, normalisation would have been necessary even using the same protocol. All image intensities were normalised according to the following equation:

2.2. Assessment of tumour stage Patients were staged according to the latest seventh edition of the International Union Against Cancer/American Joint Committee on Cancer (UICC/AJCC) staging manual [26]. All medical images and clinical records were reviewed by two radiologists separately, and disagreements were resolved by discussion.

Inew = (I − IMin )

InewMax − InewMix + InewMin IMax − IMin

(1)

Normalisation transforms an MR image I with intensity values in the range (IMin, IMax) into a new image Inew with intensity values in the range (InewMin, InewMax). Here we InewMin = 0, InewMax = 1000 defined

2.3. Induction chemotherapy All the patients accepted IC every 3 weeks for 2 cycles (n = 16) or 3 cycles (n = 104). Each cycle IC protocol consisted of cisplatin (60 mg/ m2) on days 1–3, 5-fluorouracil (600 mg/m2) on days 1–5, and docetaxel (60 mg/m2) on day 1.

2.8. Texture analysis Using ImageJ 1.50i (Wayne Rasband, National Institutes of Health, Bethesda, MD, USA), a region of interest (ROI) was drawn around the entire tumour outline on the largest primary tumour cross-sectional area from the CE T1WI images by a radiologist who was blinded to the clinical outcome and had 10 years of experience in head and neck MR interpretation. Only tumour components within nasopharynx were included in the ROI. The ROI was then copied and applied on the same level slices of T1WI, T2WI, and T2WI FS (Fig. 1). Textures were then extracted from T1WI, T2WI, T2WI FS, and CE T1WI images, respectively, using an in-house texture analysis algorithm applied in Matlab 2010a (MathWorks, Natick, MA, USA).

2.4. Criteria for tumour response Tumour response was evaluated by two radiologists separately using MR images obtained before treatment and one week after the second cycle of IC respectively. Tumour response was defined according to RECIST 1.1 criteria [27]. Both of the radiologists blinded to the patients’ clinical data. Disagreements were resolved by their discussion. Maximum diameters of target lesions were measured on CE T1WI with tools of the picture archiving and communication system (PACS), (Carestream, Ontario, Canada). Complete response (CR) was defined as 101

European Journal of Radiology 98 (2018) 100–106

G. Wang et al.

Fig. 1. Axial pretreatment morphological MR images of a 33-year-old man. Features of radiomics were extracted from the primary tumour area (yellow overlay). (a-d): T1WI, T2WI, T2WI FS, and CE T1WI, respectively.

2.9. Image filtering

weighted by their coefficients respectively, were combined linearly to form a radiomics score (Rad-score).

A process was applied to selectively extract features of diverse sizes and intensity variations. A Laplacian of Gaussian spatial band-pass filter (∇ 2G) was used, by turning the filter parameter between 1.0 and 1.5. The filter values of 0 indicated no filtration, 1.0 indicated degrees of fine texture, 1.5 and 2.0 indicated medium textures, while 2.5 indicated coarse texture. The corresponding Matlab code is attached as supplementary material S1. The Laplacian of Gaussian filter (∇ 2G) distribution is given by

∇2 G(x, y) =

x2 + y 2 ⎞ −⎛ −1 ⎛ e ⎝ 1− 2σ 2 ⎠ πσ 4 ⎝ ⎜

⎜

⎟

2.11.2. Predictive performance The potential association of the radiomics signature with response to IC was assessed in the dataset using a Mann-Whitney U test. Furthermore, to evaluate the discrimination ability of the radiomics signature, the receiver operating characteristics (ROC) curves were then developed. The cut-off values were set at the point located on the ROC curve where the positive likelihood ratio was maximal, followed by the derivation of the sensitivity and specificity [29]. The AUC and diagnostic accuracy to distinguish the response to IC was derived. Internal validation was done with bootstrapping (1000samples). The AUCs of CE T1WI and multi-image radiomics signature were compared using a deLong test.

x2+ y 2 ⎞ ⎟ 2σ 2 ⎠ ,

where x and y denote the spatial coordinates of the pixel and σ is the value of the filter parameter.

2.11.3. Stratified analysis for the radiomics signature Considering that the derived results may be confounding [30], we designed a stratified analysis by gender, age, histological type, and stage. LASSO logistic regression was performed using the “glmnet” package running in R software, version 3.0.1 (http://www.Rproject. org). Other statistical analysis was performed with SPSS 20.0 for MAC (IBM, Armonk, NY, USA). A two-tailed P value was always computed. P < 0.05 was considered statistically significant.

2.10. Feature generation Overall, 591 texture features were extracted from the ROI area of the MRI images. Radiomics features included gray-level histogram, gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRL), Gabor and Wavelet features (supplementary material S2). 2.11. Statistical analyses

3. Results 2.11.1. Feature selection and radiomics signature building The LASSO [28]; method was adopted in selecting the most valuable predictive features from the T1WI, T2WI, T2WI FS, and CE T1WI images features individually and collectively. The selected features,

3.1. Patient characteristics Among the 120 included patients, 70 were identified respond to IC, 102

European Journal of Radiology 98 (2018) 100–106

G. Wang et al.

Fig. 2. ROC graph of the radiomics signature. Fig. 2a–b: A ROC graph of the radiomics signature of CE T1WI (a) and multiple images (b), respectively.

significant difference (P < 0.05) between the response and nonresponse dataset in their baseline characteristics.

and 50 were nonresponders. The baseline characteristics were comparable between the two groups. The mean age of the patients was 46 ± 11 years in the response group with 20 (29%) females and 47 ± 11 years in the nonresponse group with 7 (14%) females. In the response group, 67 (96%) were non-keratinising horologic type while 3(4%) were keratinising squamous; In the nonresponse group, 43 (86%) were non-keratinising and 7(14%) were keratinsing squamous type. Stage II was diagnosed for 6 (9%) patients, and stage III to IV for 64 (91%) patients in the response group. Among the nonresponse group, 4 (8%) were at stage II, and 46 (92%) were at stage III to IV. There was no

3.2. Radiomics signature development From T1WI, T2WI, and T2WI FS images separately, no valuable feature were selected by the LASSO model. Five features were selected from the CE T1WI images. When features of the T1WI, T2WI, T2WI FS, and CE T1WI images were analysed together, 16 valuable features were selected. The CE T1WI-based Rad-score was calculated for every patient 103

European Journal of Radiology 98 (2018) 100–106

G. Wang et al.

4. Discussion

Table 1 The number of valuable features and Rad-score of MR images. Images

Number of features

T1WI T2WI T2WI FS CE T1WI

0 0 0 5

T1WI + T2WI + T2WI FS + CE T1WI

15

Rad-score (Median [IQR]) Response

Nonresponse

– – – 0.411 (1.781–0.100) 0.590 (2.845–0.240)

– – – 0.104 (0.833–0.241) 0.001 (0.721–0.793)

In this study, an MR imaging-based radiomics signature was developed to be a predictor for the pretreatment individualised discrimination of responders and non-responders to IC in patients with NPC. Non-responders may avoid ineffective IC cause of this predictor. But further research may be needed before this radiomics signature can be clinically useful, because the specificity and PPV for non-responders were 0.529 and 0.593 respectively. The values of them may be not high enough to satisfy clinical application. Recent studies demonstrated that the response to IC correlates with the clinical outcomes of NPC. Hao et al. [14] reported that the tumour response to IC was an independent factor for 4-year disease-free survival (DFS), overall survival (OS) and locoregional relapse-free survival (LRRFS). Liu et al. [15] found that an unsatisfactory tumour response after IC could serve as a predictor for patients with advanced-stage NPC. The stratified analysis showed that the MR-based radiomics signatures were still independent predictors for the discrimination of response and nonresponse to IC in patients with NPC even after adjusting for age and gender (P < 0.05). However, for the patients whose histologic type were keratinising squamous and the stage II patients, radiomics signatures cannot differentiate between responders and nonresponders (P > 0.05), which may be due to the small number of these patients (10 with keratinising squamous and 10 with stage II). In terms of the limitations in our study, RECIST criteria [27] was adopted to delineate response or non-response to IC. That may not be the best surrogate indicator for clinical outcomes such as local control, DFS or OS. Indeed, other criteria for designating response may be more discriminatory. Determining which criteria are useful would require larger studies with extended outcomes. MR images of this study were obtained from two different vendors. That may affect image texture. Further research may be needed to quantitatively evaluate the effect. In our study, only 2-dimensional (2D) analysis was applied on the largest tumour area in an axial slice. Considering that the selected crosssectional slice may not completely represent the heterogeneous characteristics of NPC, a 3-dimensional (3D) analysis that may perform better is well worth to be tried in future research. However, there are some studies indicating that the 3D analysis tends to be more representative of tumour tissue heterogeneity [31]. On the other hand, the 3D analysis is more complex and time-consuming. Moreover, a study by Lubner et al. illustrated that single-slice 2D tissue texture analysis was adequate, because there was no significant difference between the tissue texture results of 2D and 3D analysis [32]. Hence, we decided to adopt a single-slice 2D analysis in our initial research. Further research exploring the potential value of 3D radiomics analysis for the predicting response to IC in patients with NPC is of interest. In conclusion, in this study, a radiomics signature was developed and validated to be a significant predictor of the response to IC in patients with NPC. As a non-invasive MR-based imaging biomarker, the radiomics signature may provide a valuable and practical method to identify individual characteristics to guide the individual treatment.

P value

– – – < 0.001 < 0.001

Note: IQR, interquartile range; a P value < 0.05 indicates a significant difference in the median Rad-score between the response and nonresponse NPC patients.

based on the features as follows: CE T1WI-Rad-score = 0.1813599 + 2.150588 × CE_T1WI_correlation _135_3_0 − 0.0561872 × CE_T1WI _his_10_mean_2.0 − 0.002037721 × CE_T1WI_his_10_mean_2.5 − 0.0545887 × CE_T1WI_gabor_1_3 − 9.8073 × 10−16 × CE_T1WI _gabor_3_1 The multiple images (T1WI, T2WI, T2WI FS, and CE T1WI)-based Rad-score was calculated for every patient based on the features as follows: multiple images Rad-score = 1.2467201451 + 3.3610715766 × CE_T1WI_correlation_135_3_0 − 0.053144013 × CE_T1WI_his_10_ mean_2.0 − 0.0266184043 × CE_T1WI_his_10_mean_2.5 − 0.031548 064 × CE_T1WI_gabor_1_3 − 0.0063304727 × CE_T1WI_gabor_3_1 − 0.9212419868 × T1WI_correlation_45_3_0 − 0.0211424045 × T1WI _his_10_mean_0 + 0.163070193 × T2WI_correlation_135_2_0 − 2.9 34023 × 10−4 × T2WI_his_10_SD_0 + 0.006666803 × T2WI_his_10 _SD_1.5 + 1.855622 × 10−4 × T2WI_FS_his_10_mean_0 − 0.135469 0517 × T2WI_FS_kurtosis_2.0 − 0.0014010095 × T2WI_FS_wavelet_1 − 0.0608122693 × T2WI_FS_gabor_1_3 − 0.0955222569 × T2WI_FS _gabor_2_4

3.3. Predictive performance of the radiomics signature The median of Rad-score of CE T1WI and multiple images in the response and nonresponse dataset is listed in Table 1. The Rad-score of CE T1WI had a significant difference between the response and nonresponse patients (P < 0.001). The CE T1WI-based radiomics signature illustrated good performance in discriminating the nonresponse (positive) and response (negative) to IC in patients with NPC, which yielded a sensitivity of 0.940, specificity of 0.500, AUC of 0.715 (Fig. 2a), PPV of 0.568 and NPV of 0.897. The AUC of 1000 bootstrap internal validation was 0.715. The multiple image-based radiomics signature demonstrated better performance, which yielded a sensitivity of 0.980, specificity of 0.529, AUC of 0.822 (Fig. 2b), PPV of 0.593 and NPV of 0.949. The AUC of 1000 bootstrap internal validation was 0.821. The AUC value of multiple image-based radiomics signature was significant higher than that developed only from CE T1WI (P < 0.05).

Conflict of interest All authors have no conflict of interest to disclose.

3.4. Stratified analysis for the radiomics signature in IC response prediction The stratified analysis demonstrated that the CE T1W and multiple images-based radiomics signature were still independent predictors for the discrimination of response and nonresponse to IC in patients with NPC even after adjusting for age and gender (P < 0.05; Table 2). For the patients whose histologic types were keratinising squamous and the stage II patients, radiomics signatures cannot differentiate between the response and nonresponse patients (P > 0.05).

Funding This work was supported by the National Natural Scientific Foundation of China (Nos. 81271569, 81271654), and the Medical Science Research Foundation of Guangdong Province (No. A2014052)

104

European Journal of Radiology 98 (2018) 100–106

G. Wang et al.

Table 2 The Rad-score of the CE T1WI and multiple images for the response and nonresponse groups. CE T1WI

P value

Rad-score (Median [IQR])

T1WI + T2WI + T2WI FS + CE T1WI

P value

Rad-score (Median [IQR])

Response

Nonresponse

Response

Nonresponse

Gender Male Female

0.318 (1.781–0.100) 0.547 (1.729–0.052)

0.103 (0.548–0.241) 0.104 (0.833–0.107)

0.002 0.036

0.383 (2.845–0.240) 0.691 (2.311–0.033)

0.012 (0.721–0.793 0.001 (0.206–0.390)

0.000 0.000

Age Age < = 45 yr Age > 45 yr

0.593 (1.781–0.100) 0.337 (1.619–0.052)

0.103 (0.833–0.031) 0.127 (0.411–0.241)

0.014 0.002

0.652 (2.845–0.199) 0.412 (2.0264–0.240)

0.013 (0.721–0.715) −0.028 (0.457–0.793)

0.000 0.000

Histologic type Non-keratinising Keratinising squamous

0.400 (1.781–0.100) 0.460 (0.473–0.015)

0.103 (0.833–0.241) 0.105 (0.345,0.019)

0.000 0.517

0.617 (2.845–0.240) 0.345 (0.395–0.081)

−0.002 (0.721–0.793) 0.105 (0.453–0.289)

0.000 0.383

Stage II III ∼ IV

0.554 (1.781–0.132) 0.398 (1.729–0.100)

0.408 (0.411–0.345) 0.084 (0.833–0.241)

0.476 0.000

0.781 (2.845–0.239) 0.517 (2.311–0.240)

0.373 (0.457–0.276) −0.027 (0.721–0.793

0.114 0.000

Note: IQR, interquartile range; a P value < 0.05 indicates a significant difference in the median Rad-score between the response and nonresponse groups.

Appendix A. Supplementary data Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/j.ejrad.2017.11.007.

[12]

References [1] L.A. Torre, F. Bray, R.L. Siegel, J. Ferlay, J. Lortet-Tieulent, A. Jemal, Global cancer statistics 2012, CA. Cancer J. Clin. 65 (2015) 87–108. [2] A.W.M. Lee, B.B.Y. Ma, W.T. Ng, A.T.C. Chan, Management of nasopharyngeal carcinoma: current practice and future perspective, J. Clin. Oncol. 33 (2015) 3356–3364. [3] A.T.C. Chan, S.F. Leung, R.K.C. Ngan, P.M.L. Teo, W.H. Lau, W.H. Kwan, E.P. Hui, H.Y. Yiu, W. Yeo, F.Y. Cheung, K.H. Yu, K.W. Chiu, D.T. Chan, T.S.K. Mok, S. Yau, K.T. Yuen, F.K.F. Mo, M.M.P. Lai, B.B.Y. Ma, M.K.M. Kam, T.W.T. Leung, P.J. Johnson, P.H.K. Choi, B.C.Y. Zee, Overall survival after concurrent cisplatinradiotherapy compared with radiotherapy alone in locoregionally advanced nasopharyngeal carcinoma, J. Natl. Cancer Inst. 97 (2005) 536–539. [4] X. Wu, P.Y. Huang, P.J. Peng, L.X. Lu, F. Han, S.X. Wu, X. Hou, H.Y. Zhao, Y. Huang, W.F. Fang, Y.Y. Zhao, C. Xue, Z.H. Hu, J. Zhang, J.W. Zhang, Y.X. Ma, W.H. Liang, C. Zhao, L. Zhang, Long-term follow-up of a phase III study comparing radiotherapy with or without weekly oxaliplatin for locoregionally advanced nasopharyngeal carcinoma, Ann. Oncol. 24 (2013) 2131–2136. [5] J. Wee, E.H. Tan, B.C. Tai, H.B. Wong, Randomized trial of radiotherapy versus Concurrent Chemoradiotherapy followed by adjuvant chemotherapy in patients with american joint committee on Cancer/International union against Cancer Stage III and IV nasopharyngeal cancer of the Endemic variet, J. Clin. Oncol. 23 (2005) 6730–6738. [6] M. Al-Sarraf, M. LeBlanc, P.G. Giri, K.K. Fu, J. Cooper, T. Vuong, A.A. Forastiere, G. Adams, W.A. Sakr, D.E. Schuller, J.F. Ensley, Chemoradiotherapy versus radiotherapy in patients with advanced nasopharyngeal cancer: phase III randomized Intergroup study 0099, J. Clin Oncol. 16 (1998) 1310–1317. [7] Y. Chen, Y. Sun, S.-B. Liang, J.-F. Zong, W.-F. Li, M. Chen, L. Chen, Y.-P. Mao, L.L. Tang, Y. Guo, A.-H. Lin, M.-Z. Liu, J. Ma, Progress report of a randomized trial comparing long-term survival and late toxicity of concurrent chemoradiotherapy with adjuvant chemotherapy versus radiotherapy alone in patients with stage III to IVB nasopharyngeal carcinoma from endemic regions of China, Cancer 119 (2013) 2230–2238. [8] A.W.M. Lee, S.Y. Tung, R.K.C. Ngan, R. Chappell, D.T.T. Chua, T.X. Lu, L. Siu, T. Tan, L.K. Chan, W.T. Ng, T.W. Leung, Y.T. Fu, G.K.H. Au, C. Zhao, B. O’Sullivan, E.H. Tan, W.H. Lau, Factors contributing to the efficacy of concurrent-adjuvant chemotherapy for locoregionally advanced nasopharyngeal carcinoma: combined analyses of NPC-9901 and NPC-9902 Trials, Eur. J. Cancer. 47 (2011) 656–666. [9] A.W.M. Lee, S.Y. Tung, D.T.T. Chua, R.K.C. Ngan, R. Chappell, R. Tung, L. Siu, W.T. Ng, W.K. Sze, G.K.H. Au, S.C.K. Law, B. O’Sullivan, T.K. Yau, T.W. Leung, J.S.K. Au, W.M. Sze, C.W. Choi, K.K. Fung, J.T. Lau, W.H. Lau, Randomized trial of radiotherapy plus concurrent-adjuvant chemotherapy vs radiotherapy alone for regionally advanced nasopharyngeal carcinoma, J. Natl. Cancer Inst. 102 (2010) 1188–1198. [10] Q.-Y. Chen, Y.-F. Wen, L. Guo, H. Liu, P.-Y. Huang, H.-Y. Mo, N.-W. Li, Y.-Q. Xiang, D.-H. Luo, F. Qiu, R. Sun, M.-Q. Deng, M.-Y. Chen, Y.-J. Hua, X. Guo, K.-J. Cao, M.H. Hong, C.-N. Qian, H.-Q. Mai, Concurrent chemoradiotherapy vs radiotherapy alone in stage II nasopharyngeal carcinoma: phase III randomized trial, J. Natl. Cancer Inst. 103 (2011) 1761–1770. [11] P. Blanchard, A. Lee, S. Marguet, J. Leclercq, W.T. Ng, J. Ma, A.T.C. Chan, P.Y. Huang, E. Benhamou, G. Zhu, D.T.T. Chua, Y. Chen, H.Q. Mai, D.L.W. Kwong,

[13]

[14]

[15]

[16]

[17]

[18]

[19]

[20]

[21]

[22]

[23]

[24]

105

S.L. Cheah, J. Moon, Y. Tung, K.H. Chi, G. Fountzilas, L. Zhang, E.P. Hui, T.X. Lu, J. Bourhis, J.P. Pignon, Chemotherapy and radiotherapy in nasopharyngeal carcinoma: an update of the MAC-NPC meta-analysis, Lancet Oncol. 16 (2015) 645–655. Y. Sun, W.-F. Li, N.-Y. Chen, N. Zhang, G.-Q. Hu, F.-Y. Xie, Y. Sun, X.-Z. Chen, J.G. Li, X.-D. Zhu, C.-S. Hu, X.-Y. Xu, Y.-Y. Chen, W.-H. Hu, L. Guo, H.-Y. Mo, L. Chen, Y.-P. Mao, R. Sun, P. Ai, S.-B. Liang, G.-X. Long, B.-M. Zheng, X.-L. Feng, X.-C. Gong, L. Li, C.-Y. Shen, J.-Y. Xu, Y. Guo, Y.-M. Chen, F. Zhang, L. Lin, L.-L. Tang, M.-Z. Liu, J. Ma, Induction chemotherapy plus concurrent chemoradiotherapy versus concurrent chemoradiotherapy alone in locoregionally advanced nasopharyngeal carcinoma: a phase 3 multicentre, randomised controlled trial, Lancet Oncol. 17 (2016) 1509–1520. R.-F. Yen, T.H.-H. Chen, L.-L. Ting, K.-Y. Tzen, M.-H. Pan, R.-L. Hong, Early restaging whole-body (18)F-FDG PET during induction chemotherapy predicts clinical outcome in patients with locoregionally advanced nasopharyngeal carcinoma, Eur. J. Nucl. Med. Mol. Imaging 32 (2005) 1152–1159. H. Peng, L. Chen, Y. Zhang, W.-F. Li, Y.-P. Mao, X. Liu, F. Zhang, R. Guo, L.-Z. Liu, L. Tian, A.-H. Lin, Y. Sun, J. Ma, The tumour response to induction chemotherapy has prognostic value for long-term survival outcomes after intensity-modulated radiation therapy in nasopharyngeal carcinoma, Sci. Rep. 6 (2016) 24835. L.-T. Liu, L.-Q. Tang, Q.-Y. Chen, L. Zhang, S.-S. Guo, L. Guo, H.-Y. Mo, C. Zhao, X. Guo, K.-J. Cao, C.-N. Qian, M.-S. Zeng, J.-X. Bei, M.-H. Hong, J.-Y. Shao, Y. Sun, J. Ma, H.-Q. Mai, The prognostic value of plasma epstein-barr viral DNA and tumor response to neoadjuvant chemotherapy in advanced-stage nasopharyngeal carcinoma, Int. J. Radiat. Oncol. Biol. Phys. 93 (2015) 862–869. D. Zheng, Y. Chen, Y. Chen, L. Xu, F. Lin, J. Lin, C. Huang, J. Pan, Early assessment of induction chemotherapy response of nasopharyngeal carcinoma by pretreatment diffusion-weighted magnetic resonance imaging, J. Comput. Assist. Tomogr. 37 (2013) 673–680. Y. Chen, X. Liu, D. Zheng, L. Xu, L. Hong, Y. Xu, J. Pan, Diffusion-weighted magnetic resonance imaging for early response assessment of chemoradiotherapy in patients with nasopharyngeal carcinoma, Magn. Reson. Imaging 32 (2014) 630–637. D. Zheng, Y. Chen, X. Liu, Y. Chen, L. Xu, W. Ren, W. Chen, Q. Chan, Early response to chemoradiotherapy for nasopharyngeal carcinoma treatment: value of dynamic contrast-enhanced 3.0 T MRI, J. Magn. Reson. Imaging 0 (2014) 1528–1540. D. Zheng, Q. Yue, W. Ren, M. Liu, X. Zhang, H. Lin, G. Lai, W. Chen, Q. Chan, Y. Chen, Early responses assessment of neoadjuvant chemotherapy in nasopharyngeal carcinoma by serial dynamic contrast-enhanced MR imaging, Magn. Reson. Imaging 35 (2017) 125–131. Y. Xiao-Ping, H. Jing, L. Fei-Ping, H. Yin, W. Qiang, W. Wei, Intravoxel incoherent motion MRI for predicting early response to induction chemotherapy and chemoradiotherapy in patients with nasopharyngeal carcinoma, J. Magn. Reson. Imaging 24 (2015) 1179–1190. Y. Xiao, J. Pan, Y. Chen, Y. Chen, Z. He, X. Zheng, Intravoxel incoherent motionmagnetic resonance imaging as an early predictor of treatment response to neoadjuvant chemotherapy in locoregionally advanced nasopharyngeal carcinoma, Medicine (Baltimore) 94 (2015) e973. K. Nie, L. Shi, Q. Chen, X. Hu, S.K. Jabbour, N. Yue, T. Niu, X. Sun, Rectal cancer assessment of neoadjuvant chemoradiation outcome based on radiomics of multiparametric MRI, Clin. Cancer Res. 22 (2016) 5256–5264. H. Li, Y. Zhu, E.S. Burnside, K. Drukker, K.A. Hoadley, C. Fan, S.D. Conzen, G.J. Whitman, E.J. Sutton, J.M. Net, M. Ganott, E. Huang, E.A. Morris, C.M. Perou, Y. Ji, M.L. Giger, MR imaging radiomics signatures for predicting the risk of Breast cancer recurrence as given by research versions of MammaPrint, oncotype DX, and PAM50 gene assays, Radiology 281 (2016) 382–391. J. Liu, Y. Mao, Z. Li, D. Zhang, Z. Zhang, S. Hao, B. Li, Use of texture analysis based on contrast-enhanced MRI to predict treatment response to chemoradiotherapy in nasopharyngeal carcinoma, J. Magn. Reson. Imaging 44 (2016) 445–455.

European Journal of Radiology 98 (2018) 100–106

G. Wang et al.

[29] J.A. Hanley, B.J. McNeil, The meaning and use of the area under a receiver operating characteristic (ROC) curve, Radiology 143 (1982) 29–36. [30] M.H. Katz, Multivariable analysis: a primer for readers of medical research, Ann. Intern. Med. 138 (2003) 644–650. [31] F. Ng, R. Kozarski, B. Ganeshan, V. Goh, Assessment of tumor heterogeneity by CT texture analysis: can the largest cross-sectional area be used as an alternative to whole tumor analysis? Eur. J. Radiol. 82 (2013) 342–348. [32] M.G. Lubner, N. Stabo, S.J. Lubner, A.M. del Rio, C. Song, R.B. Halberg, P.J. Pickhardt, CT textural analysis of hepatic metastatic colorectal cancer: pretreatment tumor heterogeneity correlates with pathology and clinical outcomes, Abdom. Imaging 40 (2015) 2331–2337.

[25] B. Zhang, J. Tian, D. Dong, D. Gu, Y. Dong, L. Zhang, Z. Lian, J. Liu, X. Luo, S. Pei, X. Mo, W. Huang, F. Ouyang, B. Guo, L. Liang, W. Chen, C. Liang, S. Zhang, Radiomics features of multiparametric MRI as novel prognostic factors in advanced nasopharyngeal carcinoma, Clin. Cancer Res. 23 (2017) 4259–4269. [26] A. Edge, S. Byrd, D.R. Compton, C.C. Fritz, A.G. Greene, F.L. Trotti, AJCC Cancer Staging Manual, 7th edition, Springer, 2009. [27] E.A. Eisenhauer, P. Therasse, J. Bogaerts, L.H. Schwartz, D. Sargent, R. Ford, J. Dancey, S. Arbuck, S. Gwyther, M. Mooney, L. Rubinstein, L. Shankar, L. Dodd, R. Kaplan, D. Lacombe, J. Verweij, New response evaluation criteria in solid tumours: revised RECIST guideline (version 1.1), Eur. J. Cancer. 45 (2009) 228–247. [28] R. Tibshirani, The lasso method for variable selection in the Cox model, Stat. Med. 16 (1997) 385–395.

106