Introduction
Nasopharyngeal carcinoma (NPC) is a cancer arising from the nasopharynx epithelium with a very unique geographic distribution [
1]. It is one of the most common malignant tumours in South-Eastern China, South-Eastern Asia and Northern Africa [
2]. Despite advances in radiotherapy and chemotherapy, locoregional recurrence and distant metastasis can occur in nearly 10–15% patients during the first 2 years after the start of treatment and only 72.9% patients have a 2-year progression-free survival (PFS) [
3]. TNM staging system is insufficient to predict the prognosis of NPC, as NPC patients with the same TNM stage often show different clinical outcomes [
4]. Several molecular biomarkers have been correlated with survival in NPC patients [
5,
6]. Nonetheless, these biomarkers are obtained through randomly sampled biopsy that evaluates a small fraction of the tumour. As such, they have inherent limitations including the evaluation of invasiveness and misrepresentation of the entire tumour due to heterogeneity [
7].
High-throughput extraction of quantitative features from images is an attractive strategy for objective assessment of tumour heterogeneity [
8]. Texture features that analyse the distribution and relationship of pixel or voxel grey levels within the images are most widely used [
9]. In previous studies, texture features extracted from images of PET, CT or MRI have been associated with clinical prognosis in various types of cancers [
10‐
13]. However available data for NPC are scarce. Only a recent study reported that MRI images radiomics features could be used to predict PFS in patients with advanced NPC [
14]. However, this study only investigated NPC patients with advanced disease (stages III–IV). In addition, radiomics-based nomograms are difficult to interpret and time-consuming to be used in daily practice. Furthermore, tumour volume has been shown to be a very important prognostic factor for NPC patients [
15,
16]. Whether texture features combined with tumour volume and TNM stage can provide a better prognostic ability for NPC patients remains unknown.
The purpose of our study was to determine the predictive value of pretreatment MRI texture analysis for PFS in patients with primary NPC. To achieve this aim, we performed texture analysis of pretreatment T2-weighted images (T2WIs) and contrast-enhanced T1-weighted images (CE-T1WIs), and combined it with tumour volume and TNM stage.
Discussion
Our study demonstrated that tumour volume and CE-T1WI-based uniformity were independent predictors for PFS in patients with NPC. Specifically, higher CE-T1WI-based uniformity and smaller tumour volume were prognostic factors for favourable PFS. A single texture parameter, CE-T1WI-based uniformity, when combined with tumour volume and the overall stage, showed higher predictive ability (AUC, 0.825) than the tumour volume (AUC, 0.659) or the overall stage alone (AUC, 0.636). Texture analysis can improve PFS prediction when combined with clinical indexes, such as tumour volume or overall stage.
Identification of high-risk patients would be beneficial, like inviting to more intensive observation and/or more aggressive treatment [
22]. High tumour heterogeneity is usually associated with poor prognosis [
23]. Texture analysis allows for objective assessment of heterogeneity beyond visual interpretation [
24]. Statistical texture analysis techniques have been the most widely used method, which can yield three orders of texture parameters [
9]. The first-order texture parameters are obtained from the histogram of pixel intensity values, including uniformity (measure of homogeneity of the distribution of grey levels), skewness (measure of asymmetry of the pixel histogram) and kurtosis (measure of peakness of the pixel histogram), which describe the image grey-level heterogeneity [
25]. The grey-level co-occurrence matrix (GLCM) measurement is a well-known second-order statistics method, which is calculated using spatial grey-level dependence matrices, and measures local heterogeneity related only to the neighbouring pixels, yielding texture parameters such as GLCM entropy (measure of randomness of the GLCM) and angular second moment (measure of homogeneity of the GLCM) [
26]. The third-order statistics reveals the spatial relationship among three or more pixels [26]. A recent study showed that among 177 radiomics features including intensity, shape and texture features, many radiomics features were redundant [
27]. In our study, five first-order parameters together with four GLCM parameters were selected for simplicity.
Previously, texture features have been useful in predicting prognosis of many types of cancers, such as oesophageal, head-and-neck, colorectal, breast and non-small cell lung cancer [
10‐
13]. For NPC, intratumour heterogeneity can be assessed by texture features measured on the PET component of
18F-FDG PET/CT, and uniformity and skewness were found to be superior to traditional PET parameters in predicting clinical outcomes in patients with primary NPC [
28]. Compared with PET/CT, MRI is more widely used in clinical workup to diagnose and stage NPC before treatment due to its excellent spatial resolution, absence of radiation and lower cost. In a recent study of 118 NPC patients, radiomics-based nomograms from pretreatment MRI were found to be useful prognostic predictors [
14]. In this study, a total of 970 radiomics features were derived from T2WIs and CE-T1WIs in advanced NPC patients (stages III–IV). The prognostic ability of this MRI-based radiomics nomograms remains to be evaluated in low-stage NPC patients (stages I–II). In addition, only eight features of 970 radiomics features were found to be prognostic [
14]. In the present study, 79 patients with stages I–IV NPC were enrolled. We assessed texture features derived from T2WIs and CE-T1WIs together with tumour volume. Our results demonstrated that a single texture parameter, CE-T1WI-based uniformity, was an independent factor for PFS in NPC patients. Moreover, both tumour volume and CE-T1WI-based uniformity could predict PFS in NPC patients with a comparable prognostic ability. Comparatively, a single texture parameter would be more favourable for clinical application than radiomics-based nomograms. Notably, when combined either with tumour volume or with the overall stage, CE-T1WI-based uniformity can yield a higher C-index for predicting PFS (C-index, 0.754 and 0.756, respectively), which is very close to the C-index attained by previously reported radiomics-based nomograms (C-index, 0.756) [
14]. Moreover, when combined with both tumour volume and overall stage, the prognostic ability was further improved (AUC, 0.825; C-index, 0.794). These results suggest that CE-T1WI-based uniformity alone has a high prognostic performance. It can be used as a complementary index to improve the prognostic ability of frequent clinical indexes, such as tumour volume and overall stage.
In our study, higher CE-T1WI-based uniformity was a predictive factor for favourable PFS in patients with NPC; a more homogeneous enhancement with a tumour on CE-T1WI was associated with better prognosis. Similar results were also found in other cancers [
29,
30]. For example, lung cancer with less heterogeneity on CE-T1WI (less entropy) is associated with improved 2-year PFS [
31]. Higher uniformity based on contrast-enhanced CT is associated with improved survival in oesophageal cancer treated with chemotherapy and radiation therapy [
32]. More homogeneity on contrast-enhanced radiographic images suggests more homogeneous angiogenesis within the tumour, associated with better prognosis in head and neck squamous cell carcinoma [
33,
34]. Taken together, texture analysis of postcontrast images offers a non-invasive and low-cost new insight into the relationship between angiogenesis and patient survival.
In our study, larger tumour volume was shown to be an adverse predictor for PFS in NPC patients. It has been reported that larger volume is associated with poorer prognosis in many solid tumours, such as NPC, cholangiocarcinoma and tongue carcinoma [
16,
35,
36]. Larger solid tumours are more likely to release more cells, exposing to metastases and poor prognosis [
37].
Our study still has some limitations. First, the sample size, in particular for the recurrent cases, is relatively small because patients were enrolled from a single centre. The prognostic value of uniformity remains to be validated in a large cohort study. Second, the follow-up period was relatively short. Only 2 years PFS was adopted as the endpoint of survival outcome. Long-term overall survival was not attempted in our study because most recurrences or metastasis occur within 2 years after primary treatment in NPC patients [
38]. Third, NPC patients were treated by concurrent radiotherapy and chemotherapy or radiotherapy alone; and varying doses of radiation were adopted for radiotherapy regimen. These different strategies might be confounding factors for the evaluation of PFS.
In conclusion, our study showed that tumour volume and pretreatment CE-T1WI-based uniformity are prognostic predictors in NPC patients. A small size and homogeneous contrast-enhancement with a primary lesion are predictive of better PFS. The addition of CE-T1WI-based texture analysis to tumour volume and overall stage can improve the prediction of PFS in NPC patients.