Predicting chemoradiotherapy response of nasopharyngeal carcinoma using texture features based on intravoxel incoherent motion diffusion-weighted imaging

Abstract

The aim of the study was to investigative the utility of gray-level co-occurrence matrix (GLCM) texture analysis based on intravoxel incoherent motion diffusion-weighted imaging (IVIM-DWI) for predicting the early response to chemoradiotherapy for nasopharyngeal carcinoma (NPC).

Baseline IVIM-DWI was performed on 81 patients with NPC receiving chemoradiotherapy in a prospective nested case–control study. The patients were categorized into the residue (n = 11) and nonresidue (n = 70) groups, according to whether there was local residual lesion or not at the end of chemoradiotherapy. The pretreatment tumor volume and the values of IVIM-DWI parameters (apparent diffusion coefficient [ADC], D, D^∗, and f) and GLCM features based on IVIM-DWI were compared between the 2 groups. Receiver operating characteristic (ROC) curves in univariate and multivariate logistic regression analysis were generated to determine significant indicator of treatment response.

The nonresidue group had lower tumor volume, ADC, D, Correlat_ADC, Correlat_D, InvDfMom_ADC, InvDfMom_D and InvDfMom_D^∗ values, together with higher Contrast_D, Contrast_f, SumAverg_ADC, SumAverg_D, and SumAverg_D^∗ values, than the residue group (all P < .05). Based on ROC curve in univariate analysis, the area under the curve (AUC) values for individual GLCM features in the prediction of the treatment response ranged from 0.635 to 0.879, with sensitivities from 54.55% to 100.00% and specificities from 52.86% to 85.71%. Multivariate logistic regression analysis demonstrated D (P = .026), InvDfMom_ADC (P = .033) and SumAverg_D (P = .015) as the independent predictors for identifying NPC without residue, with an AUC value of 0.977, a sensitivity of 90.91% and a specificity of 95.71%.

Pretreatment GLCM features based on IVIM-DWI, especially on the diffusion-related maps, may have the potential to predict the early response to chemoradiotherapy for NPC.

1. Introduction

As a tumor with a noticeable geographic and racial distribution worldwide, nasopharyngeal carcinoma (NPC) is a common malignancy of head and neck in South China, such as Guangxi and Guangdong provinces.^[1,2] In most areas worldwide, annual incidence of NPC is <1 per 100,000 for both sexes, for example, 0.5 and 0.2 per 100,000 person-years for male and female in the United States, respectively.^[1] However, the newly reported incidence of NPC in Guangxi and Guangdong is about 11 per 100,000.^[2] Chemoradiotherapy is currently regarded as the preferred treatment for locally advanced NPC as it can decrease the local recurrence rate and subsequently achieve significant survival benefit.^[3,4] However, local failure (ie, residue or recurrence) remain one of the predominant patterns of treatment failure in patients with NPC.^[5] Predicting therapeutic response to chemoradiotherapy as early as possible has remarkable clinical benefit for patients with NPC, which could optimize treatment regimen, avoid unnecessary suffering from chemoradiotherapy toxicity, and reduce medical cost.

Nowadays, magnetic resonance imaging (MRI) is playing an increasingly important role in lesion detection, clinical staging, and therapeutic effect evaluation and prediction for NPC.^[6–9] Convention morphology-based MRI offers little benefit in the early prediction of therapeutic effect in NPC. Recently, functional MRI approaches including dynamic contrast-enhanced MRI and diffusion-weighted imaging (DWI) have been utilized to predict the treatment response for NPC.^[8,10–12] Nevertheless, previous studies have demonstrated variable efficacy of dynamic contrast-enhanced MRI and DWI in early predicting the response of NPC to chemoradiotherapy.^[8,10–15] These controversial results might result from that these MRI approaches were based on the analysis of mean value of MRI parameter, which cannot accurately reflect histologic heterogeneity. Heterogeneity is a common phenomenon in malignancies and it is strongly linked with curative effect of NPC.^[16,17] Texture features derived from MRI have the ability to quantitate tumor heterogeneity.^[16,18–20] Some published studies have demonstrated the advantage of texture features over the analysis of morphology-based MRI and mean value of MRI parameter in discriminating clinical stage or treatment response for a variety of tumors.^[21–24]

Up to now, most of the available MRI literatures on texture analysis were based on morphologic image, although several on dynamic contrast-enhanced MRI or the apparent diffusion coefficient (ADC) map derived from DWI.^[16,20,21] Morphologic image cannot provide functional information of tissue microstructure, which may limit its potential to early predict treatment response of tumors. In regard to the ADC value, it is obtained from a mono-exponential diffusion model and only reflects hybrid diffusion information resulting from both Brownian movement and microcirculation perfusion. Intravoxel incoherent motion DWI (IVIM-DWI), with the ability to separate pure diffusion movement and perfusion, was found to be more powerful than the ADC value in predicting the curative effect of NPC.^[11,12]

The utility of texture features derived from IVIM-DWI in predicting the response to chemoradiotherapy of NPC is still unknown until now. Gray-level co-occurrence matrix (GLCM) is a widely used texture analysis algorithm which generates texture features to quantify the spatial gray-level variation within local neighborhoods around each pixel in an image.^[21,25–27] GLCM features have shown promise in predicting the treatment response in many kinds of tumors.^{[20,26,28,29]} Therefore, we hypothesize that GLCM analysis based on IVIM-DWI has a capacity to predict the early response to chemoradiotherapy for NPC, which is still unclear until now according to the best of our knowledge. The aim of this study was to investigate this capacity of GLCM approach.

2. Materials and methods

2.1. Patient selection and treatment procedure

This prospective single-center study was approved by the Medical Ethics Committee in our institution, and all patients with NPC signed written informed consent. Inclusion criteria of this study were as follows: newly diagnosed and pathologically confirmed of nonkeratinizing NPC, above 18 years old, patients were scheduled for chemoradiotherapy, and Karnofsky score ≥80. Patients were excluded if they had prior antitumor treatment for NPC, did not sign the informed consent form, or had contraindications for MRI. Ninety patients with NPC were initially enrolled from September 2016 to April 2017 (Fig. 1). All patients’ clinical stages were determined with reference to the 7th edition of the International Union Against Cancer/American Joint Committee on Cancer (UICC/AJCC) staging system.^[30]

Figure 1.

Study flow diagram in this study.

All patients received a 2-cycle induction chemotherapy (21 days per cycle) that consisted of 135 to 175 mg/m² paclitaxel on day 1 accompanied by 80 mg/m² nedaplatin on days 1, 2, and 3. After induction chemotherapy, all patients received radiotherapy with a total dose of 70 to 76 Gy and 30 to 33 times to completion. After radiotherapy, 2 head and neck radiologists (LF and HJ with 15 and 6 years of experience in neck radiology, respectively) who were blinded to the texture analysis data read the MRIs individually and classified the patients with NPC into the residue and nonresidue groups, with a double-blind control, according to whether regional residue was identified on MRI and pharyngorhinoscopy which were performed at the end of chemoradiotherapy. If they have different views of residual tumor, the third neck radiologist (YX with more than 20 years of experience in neck radiology) would analysis these data and eventually determined whether there was residue or not.

2.2. MRI protocols

Before chemoradiotherapy, all patients received conventional MRI and IVIM-DWI examinations on a 1.5 Tesla MRI scanner (Optima MR360; GE Healthcare, Milwaukee, WI). The conventional MRI protocols included the following: axial T1-weighted spin-echo images (repetition time [TR], 580 milliseconds; echo time [TE],7.8 milliseconds; slice thickness, 5 mm; intersection space, 1 mm; and number of excitations [NEX], 2); and axial T2-weighted spin-echo images with fat suppression (TR, 6289 milliseconds; TE, 85 milliseconds; slice thickness, 5 mm; intersection space, 1 mm; and NEX, 2).

Single-shot diffusion-weighted spin-echo echo-planar (SE-DW-EPI) sequence was used to gather IVIM-DWI data, with 22-cm field of view, 5-mm slice thickness, 1-mm slice gap, 4225-millisecond TR, 106-millisecond TE, 128 × 130 matrix, and 4 NEX. A total of 12 axial slices covering the nasopharynx were obtained. Ten b values (0, 50, 80, 100, 150, 200, 400, 600, 800, and 1000 s/mm²) were applied in the IVIM-DWI data acquisition. A local shim box covering the nasopharyngeal region was applied to minimize susceptibility artifacts.

2.3. IVIM-DWI parametric maps acquisition

All IVIM-DWI data were transferred to the Advantage Workstation (version AW 4.6; GE Medical Systems, Milwaukee, WI) for postprocessing using the MADC kit, a software package for multiple ADC measurement. Four IVIM-DWI parametric (ADC; pure diffusion coefficient, D; pseudo-diffusion coefficient, D^∗; perfusion fraction, f) maps of each primary NPC lesion were generated on the basis of a pixel-by-pixel fitting according to the Levenberg–Marquardt algorithm.^[31]

2.4. IVIM-DWI parametric values measurement

To determine the scope of each tumor, one radiologist (WL with 10 years of experience in head and neck MRI) who was blinded to the clinical and pathologic results and the treatment response manually traced the outer edge of the lesion on each axial ADC map with reference to the T2-weighted image, and the corresponding 2-dimensional (2D) region of interest (ROI) for each map was acquired. Both the most superior and the most inferior slices for each tumor were excluded to avoid volume averaging. Based on all the ROIs of this tumor, the MADC kit automatically generated a 3-dimensional (3D) volume of interest (VOI) for this tumor and output the volume datum and the mean ADC value of this VOI. The same VOI was also automatically copied and pasted by this software onto all the other IVIM-DWI maps, and the corresponding parametric values were obtained. Subsequently, these axial maps with ROI were saved as BMP format images for texture analysis.

2.5. Texture analysis

All the BMP format images were transferred into the MaZda program (http://www.eletel.p.lodz.pl/programy/mazda/index.php?action=mazda) for texture analysis. Because the ROIs in the IVIM-DWI maps cannot be utilized automatically by MaZda, one radiologist (WL) carefully manually traced the border of these original ROIs to generate new ROIs for texture analysis. Subsequently, a VOI for each tumor was generated automatically based on these new ROIs. For each VOI, gray-level normalization was performed by using μ ± 3σ (μ, gray-level mean; σ, gray-level standard deviation), to minimize the influence of contrast and brightness variation. For each IVIM-DWI parametric map, 11 GLCM features were extracted automatically from the VOI by MaZda, including Angular Second Moment (AngScMom), Contrast, Correlat, Difference Entropy (DifEntrp), Difference Variance (Difvarnc), Entropy, Inverse Difference Moment (InvDfMom), Sum Average (SumAverg), Sum Entropy (SumEntrp), Sum of Squares (SumOfSqs), and Sum Variance (SumVarnc).

2.6. Sample size estimation

The sample size was calculated according to the formula n = (Uα/δ)/P (1 – P). The allowable error (δ) was set to 10%.The residual lesion rate (P) was supposed to 35%, according to our prior data.^[11] The significance level (α) was defined as 0.05, and the critical value of Uα as 1.96; therefore, n = 86. The eventual sample size was 90, allowing for loss of up to 5% of the patients.

2.7. Statistical analysis

All statistical analyses were done through SPSS version 22.0 (SPSS Inc, Chicago, IL) or MedCalc v15.0 software (MedCalc Software bvba, Ostend, Belgium). A 2-tailed test pattern was used in all statistical analyses with the level of statistical significance determined as P < .05. Categorical variables (gender, Tumor-Node-Metastasis [TNM] stage, and pathologic grade) were presented as frequency, and were compared using the Chi-squared test. Continuous variables (age, tumor volume, and the values of IVIM-DWI parameters and GLCM features) were expressed as mean ± standard deviation. The nonparametric Mann–Whitney U test was used in univariate analysis to explore the possible differences in continuous variables between the residue and nonresidue groups. Subsequently, multivariate logistic regression analysis (forward stepwise, LR; probability for stepwise entry, 0.05; removal, 0.1) was performed to identify the independent prognostic factors, using the indicators with statistical significance (P < .05) in univariate analysis as input variables. The discrimination power of the individual GLCM features and multivariate regression model for predicting the therapeutic response was determined with receiver operating characteristic (ROC) curve analysis.

3. Results

Of the initially enrolled 90 patients, 9 were eliminated because of serious image distortion (n = 3), dental artifacts (n = 3), withdrawal by patients (n = 1), or receiving other antitumor therapy after the end of chemoradiotherapy (n = 2) (Fig. 1). Thus, this study eventually included the remaining 81 patients whose clinical and pathologic characteristics are shown in Table 1. At the end of chemoradiotherapy, 11 patients were found with residual lesion and 70 were not.

Table 1.

Patient's clinical and pathologic characteristics.

There was no significant difference in age, gender, clinical stage, and pathologic grade between the residue and nonresidue groups (Table 1). Significantly larger baseline tumor volume was found for the residue group than for the nonresidue group (P = .035). In regard to the IVIM-DWI parametric values, the nonresidue group exhibited obviously lower ADC and D values (all P < .01) than the residue group, whereas D^∗ and f did not show significant difference between the 2 groups (Table 2). Among those 44 GLCM features extracted from the IVIM-DWI maps, 10 of them showed distinct difference between the 2 groups (Tables 3–6). The nonresidue group exhibited lower Correlat_ADC, Correlat_D, InvDfMom_ADC, InvDfMom_D, and InvDfMom_D^∗ values, together with higher Contrast_D, Contrast_f, SumAverg_ADC, SumAverg_D, and SumAverg_D^∗ values, than the residue group (all P < .05). Figure 2 shows 2 examples of MRIs for patients with NPC with different responses.

Table 2.

Differences in the values of IVIM-DWI parameters between the residue and nonresidue groups.

Table 3.

Differences in the GLCM features values from ADC map between the residue and nonresidue groups.

Table 6.

Differences in the GLCM features values from f map between the residue and nonresidue groups.

Figure 2.

Examples of baseline intravoxel incoherent motion diffusion-weighted imaging (IVIM-DWI) maps, VOI for texture features extraction, and pre- and posttherapy T2WI images for nasopharyngeal carcinoma. The upper 2 rows show images from a patient (case A) in the nonresidue group, whereas the lower 2 rows exhibit images from a patient (case B) in the residue group. For case A, the primary lesion disappeared at the end of chemoradiotherapy, its baseline Contrast_D, Contrast_f, Correlat_ADC, Correlat_D, DifVarnc_ADC, DifVarnc_D, InvDfMom_ADC, InvDfMom_D, InvDfMom_D^∗, SumAverg_ADC, SumAverg_D, and SumAverg_D^∗ values were 63.738, 56.015, 0.711, 0.698, 34.892, 34.361, 0.245, 0.257, 0.328, 62.340, 63.267, and 60.157, respectively. For case B, the nasopharyngeal tumor demonstrated residue at the end of chemoradiotherapy, its baseline Contrast_D, Contrast_f, Correlat_ADC, Correlat_D, DifVarnc_ADC, DifVarnc_D, InvDfMom_ADC, InvDfMom_D, InvDfMom_D^∗, SumAverg_ADC, SumAverg_D, and SumAverg_D^∗ values were 52.135, 45.632, 0.773, 0.784, 26.876, 25.178, 0.401, 0.423, 0.387, 45.867, 49.376, and 48.763, respectively. ADC = apparent diffusion coefficient, D^∗ = pseudo-diffusion coefficient, D = true-diffusion coefficient, f = perfusion fraction, IVIM-DWI = intravoxel incoherent motion diffusion-weighted imaging, post-T2WI = posttreatment T2-weighted imaging, pre-T2WI = pretreatment T2-weighted imaging, VOI = volume of interest.

Table 4.

Differences in the GLCM features values from D map between the residue and nonresidue groups.

Table 5.

Differences in the GLCM features values from D^∗ map between the residue and nonresidue groups.

Based on ROC curve in univariate analysis, the area under the curve (AUC) values for individual GLCM features in the prediction of the treatment response ranged from 0.635 to 0.879, with sensitivities from 54.55% to 100.00% and specificities from 52.86% to 85.71% (Table 7). Among the 10 individual predictors, InvDfMom_D had the highest specificity (100.00%), InvDfMom_ADC owned the best specificity (85.71%), whereas SumAverg_D exhibited the highest AUC value (0.879), followed by InvDfMom_ADC (AUC = 0.865) and SumAverg_ADC (AUC = 0.861). Multivariate logistic regression analysis demonstrated D (P = .026), InvDfMom_ADC (P = .033), and SumAverg_D (P = .015) as the independent predictors for identifying NPC without residue, with an AUC value of 0.977, a sensitivity of 90.91% and a specificity of 95.71%.

Table 7.

Performance of predictors on the differentiation between the residue and nonresidue groups in univariate analysis.

4. Discussion

This study focused on the performance of GLCM texture analysis based on IVIM-DWI on predicting the early response to chemoradiotherapy for patients with NPC. Our data showed that NPC lesions with residue after chemoradiotherapy differs from those without in the pretreatment GLCM features related to IVIM-DWI, which demonstrates the potential of GLCM analysis to predict the curative effect for NPC. Our observations also revealed that GLCM analysis might facilitate more IVIM-DWI parameters (ie, D^∗ and f) to obtain the capacity for predicting the treatment response of NPC. Moreover, the diffusion-related IVIM-DWI maps may be superior to the perfusion-related ones in terms of providing GLCM features for predicting the early response of NPC to chemoradiotherapy.

The present study demonstrated that Contrast, Correlat, InvDfMom, and SumAverg could serve as MRI biomarkers to differentiate between the residual and nonresidual NPC lesions before the start of chemoradiotherapy. Among the above GLCM features strongly correlated to treatment response, Contrast and InvDfMom are descriptors of tissue heterogeneity, while Correlation and SumAverg are not directly related to heterogeneity, according to the GLCM algorithm.^[29,32] Lower InvDfMom or higher Contrast value indicates more heterogeneity.^[25,27] In the present study, the nonresidue NPC group had higher Contrast_D and Contrast_f values, as well as lower InvDfMom_ADC, InvDfMom_D, and InvDfMom_D^∗ values, than the residue group, indicating that NPC with more heterogeneity on the IVIM-DWI maps before treatment might be associated with better response to chemoradiotherapy. Similar to our observation, several studies also reported that malignancies with higher heterogeneity on clinical imaging maps might have better therapeutic response in a variety of tumors.^[25,29,33] In a recent study,^[29] malignant lymphomas demonstrating complete metabolic responders (CMRs) to chemotherapy reportedly showed higher pretreatment Contrast value on the standard uptake value map derived from ¹⁸F-FDG PET-CT than those of non-CMR. A prior report^[33] revealed that hepatocellular carcinomas with complete response (CR) after transcatheter arterial chemoembolization exhibited higher Moments (a GLCM feature that is virtually the opposite of homogeneity, namely higher Moments value represents lower homogeneity) value on the pretherapeutic iohexol-enhanced CT images than those with non-CR. Teruel et al^[34] also found that breast cancer lesions achieving CR to chemotherapy had higher baseline Contrast and Entropy (Entropy is a texture feature representing tissue heterogeneity) values on the 1-, 2-, and 3-minute postgadolinium T1-weighted MRIs than those with a response of stable disease according to the Response Evaluation Criteria in Solid Tumors (RECIST) v1.1. Nevertheless, contrary finding was reported in another study on breast cancer in which the responders (ie, with a chemotherapy-induced decrease in tumor diameter of >50%) demonstrated lower baseline Contrast value on the T1-weighted MRIs after administration of gadolinium agent, compared with the nonresponders.^[27] These conflicting findings might result from the differences in the imaging protocols and treatment methods, especially from the great variations with regard to the classification of response across studies.^[27,34]

The IVIM-DWI can simultaneously extract diffusion and perfusion information from viable tissues.^[6,11,12] Among the 4 IVIM-DWI parameters, both D^∗ and f are related to microcirculation perfusion, while D is associated with pure diffusion.^[6,11,12] ADC represents the total diffusion in viable tissues and is mainly related to diffusion rather than perfusion under the condition of b values >200 s/mm².^[11,12,35] In regard to the four IVIM-DWI parameters, our data indicated that all of them could provide baseline GLCM features to predict the early outcome of chemoradiotherapy for NPC. Our data demonstrated that the pretreatment ADC and D values had the potential to predict the early-term treatment response for NPC, while D^∗ and f had not, which is in line with several prior studies.^[11,12,15] In the present and previous studies,^[11,12,15] the performance of D^∗ and f on differentiating curative effect was based on measure of mean IVIM-DWI parametric value instead of on texture analysis. Thus, it could be inferred that GLCM texture analysis may improve the performance of the perfusion-related IVIM-DWI parameters on the prediction of therapeutic effect for NPC, compared with the traditional average-based IVIM-DWI analysis.

In the univariate analysis in this study, the majority (7/10, 70.0%) of the GLCM features predictive of therapeutic response were generated from the diffusion-related IVIM-DWI parametric maps, whereas only 3 (3/10, 30.0%) from the perfusion-related maps. Moreover, the 10 predictors in this study belonged to 4 kinds of GLCM features, namely Contrast, Correlat, InvDfMom, and SumAverg. As for Correlat, all the 2 predictors were extracted from the ADC and D maps instead of from the D^∗ or f maps. In terms of Contrast, InvDfMom, and SumAverg, although both the diffusion- and perfusion-related IVIM-DWI maps could provide predictor for therapeutic effect, the AUC value of each individual predictor from the diffusion-related maps was higher than that from the perfusion-related maps, for example, InvDfMom_D (AUC = 0.844) versus InvDfMom_D^∗ (AUC = 0.699). Furthermore, multivariate analysis demonstrated that all the independent predictors (ie, D, InvDfMom_ADC, and SumAverg_D) were generated from the ADC or D map, rather than from the D^∗ and f maps. Similar to the findings of this study, previous reports based on an analysis of mean parametric value have revealed that the diffusion-related IVIM-DWI indices were more powerful than the perfusion-related ones in the pretreatment prediction of response to chemoradiotherapy for NPC.^[11,12] Taken together, these observations might signify that the pretreatment diffusion-related IVIM-DWI indices have an advantage over the perfusion-related ones in the early prediction of the curative effect in patients with NPC, whether based on texture analysis or average parametric value.

In this study, we explored the routine MRI findings (ie, tumor volume and TNM stage that indicates the degree of tumor progression), IVIM-DWI mean-parameter-value analysis and IVIM-DWI texture analysis for predicting NPC with lesion residue. The residue and nonresidue groups did not differ from each other in the TNM stage, suggesting that the clinical stage is of little benefit to predict the early response of NPC. In this study, NPC with larger baseline tumor volume was found more likely to appear residue at the end of chemoradiotherapy, which implied that tumor volume may serves as an indicator to predict the early chemoradiotherapy effect for NPC. Nevertheless, among all the 13 individual indicators in the prediction of tumor residue in the present study, tumor volume exhibited the lowest AUC value (0.635). Additionally, in the multivariate analysis in this study, it is the functional MRI indices (ie, D and texture features derived from IVIM-DWI maps), rather than tumor volume, that acted as the independent predictors to the treatment outcome of NPC. Moreover, this multivariate model had a significantly higher AUC value than tumor volume (AUC, 0.977 vs 0.635; P = .002). These observations strengthened the opinion that functional MRI approach may be more powerful than morphologic MRI methods in predicting chemoradiotherapy efficacy of malignancies including NPC.

There are several limitations in our study. First, we did not analysis the posttreatment GLCM features from the IVIM-DWI maps for NPC, which may result in insufficient discuss on the correlation between the therapeutic response and IVIM-DWI-based texture features for NPC. Second, we did not perform an analysis of a subgroup of patients stratified by TNM stages because of the relatively small sample size of certain stage subgroups, for example, the T1, N0, N1, and M1 subsets. In addition, the long-term outcome of NPC receiving chemoradiotherapy had not been evaluated in the present study. Therefore, further studies with larger patient population, long-term follow-up, stratification by clinical stage, and analysis of posttreatment IVIM-DWI-based texture features are warranted to comprehensively understand the relevance between the texture features and therapeutic effect for NPC.

In conclusion, the observations in this study indicated that GLCM features based on IVIM-DWI, especially on the diffusion-related map, may be a potential tool for predicting the early response of NPC before starting chemoradiotherapy.

Acknowledgment

The authors thank Mr Zhang Zhong-ping (GE Healthcare China, Beijing, China) for his assistance in MRI analysis.

Author contributions

Conceptualization: Xiaoping Yu.

Data curation: Yi Fu.

Investigation: Jing Hou, Ying Hu, Feiping Li, Lu Wen, Qiang Lu.

Methodology: Jing Hou, Feiping Li, Lu Wen, Qiang Lu.

Supervision: Ying Hu.

Validation: Siye Liu.

Writing – original draft: Yuhui Qin.

Writing – review & editing: Xiaoping Yu.