The utility of texture analysis based on quantitative synthetic magnetic resonance imaging in nasopharyngeal carcinoma: a preliminary study

Nasopharyngeal carcinoma (NPC) is an aggressive head and neck cancer with high incidence rates in several provinces in south-eastern China [1]. Due to the deep location of the nasopharynx and occult symptoms of NPC, early diagnosis is to some extent difficult. With excellent soft-tissue resolution, MRI is the preferred imaging modality for NPC diagnosis, staging, and treatment monitoring. However, NPC and benign nasopharyngeal mucosal overlap to some extent [2‐4] and asymmetric hyperplasia causes up to 14% misdiagnosis rate [5], which poses a challenge to the detection of NPC. Moreover, occipital clivus (OC) invasion, which is a poor survival factor [6], occurs in 45.3% NPC patients at the time of diagnosis [7, 8]. Notably, invaded OC typically shows low-intensity on T1WI, but normal OC exhibits age-related changes from uniformly low to uniformly high intensity [9, 10], leading to increased difficulty in discriminating OC invasion. Accordingly, accurate diagnosis of NPC and OC invasion is vital of importance, by developing advanced imaging techniques.

Recently, synthetic magnetic resonance imaging (SyMRI), based on a multiple-delay multiple-echo (MDME) sequence, could generate multiple contrast images in a single scan without contrast agents. Besides, SyMRI could simultaneously provide the longitudinal and transverse relaxation times (T1 and T2) and proton density (PD) of tissues, which are the basic intrinsic properties of MRI physics and independent from the MRI scanners or scanning parameters at a given field strength [11]. Microstructural differences and histopathological information within lesions could be accurately reflected by these three quantitative values [11, 12]. Recently, many studies have demonstrated the value of SyMRI in nasopharynx, breast, bladder and rectal cancer [13‐18]. However, it should be noted that most studies used only the Mean value parameter, and additional information such as the tumour heterogeneity, which could be reflected by first-order features, may have been ignored.

First-order features and analysis, which are relatively easier to understand and implement compared with radiomics, reflect the distribution of voxel intensity with high reproducibility [19, 20]. Previous studies [21‐24] showed that whole-tumour first-order features, such as Skewness, Variance and so on, can potentially be used for diagnosis and survival prediction in NPC. The research in SyMRI demonstrated that Mean value could help in discriminating NPC from NPH, but the test–retest repeatability and tumour heterogeneity were not evaluated [13]. To avoid the instability in the feature selection step of radiomics [25], we used first-order features as a complement to the Mean value [19]. Therefore, the purposes of our study are to explore whether first-order features derived from SyMRI could effectively discriminate NPC from NPH, and to explore the performance of SyMRI in evaluating OC invasion.

To our knowledge, this is the first prospective study to investigate the role of SyMRI combined with first-order features for the diagnosis of NPC. Our results demonstrated that the T1, T2 and PD relaxation times of NPC derived from SyMRI images were significantly different from those for NPH. Moreover, in OC invasion discrimination, the T1 map-derived model showed the highest diagnostic performance. These findings indicate that SyMRI could reflect the changes in the intrinsic characteristics of benign nasopharyngeal mucosa progressing to a tumour. These quantitative parameters provide auxiliary value for the morphology, which makes the evaluation more objective.

Endoscopic biopsy is the gold standard for diagnosis of NPC. However, this examination method is invasive and it may miss some lesions generally attributed to submucosal location, coexistent hyperplasia, and difficult pharyngeal recess structure [30, 31]. About 12% of NPC invisible under endoscopy can be successfully detected by MRI [32]. Conventional morphology on MRI, such as location site, symmetry, stripes and signal intensity could be valuable in differentiating NPC from NPH [30]. However, these image features are susceptible to the subjective influence of observers, and contrast agents are required. Recently, radiomics based on T2WI/FS provide a quantitative method to diagnose NPC, but instability in the feature selection step could reduce its reliability [25]. The quantitative parameters of various functional MRIs have also been applied to identify NPC from NPH, and higher K_trans and blood flow, and lower diffusion parameters (including ADC, D, f) were found in NPC [30, 33, 34]. Moreover, Meng et al. [13] found Mean of T1, T2 and PD maps derived from SyMRI could help in discriminating NPC from NPH, but the performance of PD_Mean was poor (AUC = 0.624) and percentile parameter has an advantage than Mean parameter [35, 36]. Conducted in a non-endemic area, our study increased the sample size and included tumour heterogeneity to further explore the performance of SyMRI in diagnosing NPC, as well as OC invasion.

T1, T2, and PD values reflect the intrinsic properties of tissues. The higher the cellularity is, the lower the extracellular space and free water content, thereby resulting in a reduction in the T1 and T2 values [37]. The 90th percentile, Median and Minimum represent the grey-level intensity within the VOIs. Our results found that NPC had significantly lower T1_Median and T2_Minimum values than NPH. A possible reason was the abundance of mucus in the benign nasopharyngeal mucosa, which correlates with higher T1 and T2 values. Similar results were found in previous studies on NPC vs. NPH, as well as prostate cancer vs. peripheral zone/benign prostatic hyperplasia and cervical cancer vs. normal mucosa [11, 13, 38]. However, recent study [15] found that the mean T1 value of breast malignancy was significantly higher than that of benign tumours, possibly because the ROI of benign tumours included more fat tissue. Thyroid research also showed that papillary thyroid carcinoma had a higher T2 value than contralateral normal tissue [39]. These controversial results stated that the different control groups led to different variations.

The PD value represents the apparent concentration of water protons (mobile hydrogen atoms) in each voxel [40]. When malignant cells invade normal tissue, the density of lesions increases, and the water content decreases, resulting in reduced PD values. The lower PD value also correlates with worse biological tumour behaviour [11, 16]. A study in breast cancer [41] found lower PD values in malignancies than in benign tumours. Our study obtained consistent results, given that several PD values (the 10th, 90th percentile, Mean, Median, Minimum) in NPC were lower than those in NPH, with the most representative being the PD_90th percentile. Due to statistically 90th percentile is less influenced by random fluctuations than Mean value [35], parameter 90th percentile has an advantage over the Mean, which may be one of reasons for the poor performance of PD_Mean in previous study [13].

OC invasion at initial diagnosis is a poor factor for survival of T3 stage NPC patients [6]. Previously studies found the signal intensity of OC is related to age and gender [9], and abnormal residual OC signal could persist several years after radiotherapy without recurrence [27], which make the identification of OC invasion and recurrence somewhat difficult. In addition, according to the criteria reported in studies [26‐28], contrast-enhanced sequences are an indispensable examination, so false-positive is more likely to occur in people who only performed non-enhanced scans. Our study demonstrated that SyMRI without contrast agents could objectively evaluate OC invasion. Compared to the OC non-invasion group, the OC invasion group had a higher T1_Mean and lower PD_Maximum and T2_10th percentile, while T1_Mean was the most representative parameter. Cancer cells infiltrate the trabecular architecture and replace the fatty and haematopoietic marrow in normal bone, resulting in increased T1 values and decreased PD and T2 values, which is supported by studies of bone metastasis [42‐44]. The MAD parameter may represent tumour heterogeneity, and a higher T2_MAD in the OC invasion group reflects a more heterogeneous tumour structure than that of the OC non-invasion group.

Energy and Total Energy, as larger values, imply a greater sum of the squares of these values and volume-confounded, are thus positively correlated with the magnitude of voxel value and the volume of delineation. In our study, NPC had lower T1, T2 and PD values but higher Total Energy, while OC invasion had a higher T1 value and higher T1_Total Energy, but lower T2 value and lower T2_Energy. We speculated that the volume may greatly impact this formal comparison, since there is a large difference in the parameter volume between NPC tumours and NPH (5945.12 mm³ vs. 1403.05 mm³) but relatively small between OC invasion and non-invasion group (1674.28 mm³ vs. 3281.64 mm³). Zhang et al. [45] found stage IA endometrial carcinoma had lower signal intensities (including Mean, Median and so on) and Total Energy on ADC map than benign endometrial lesions. Ghosh et al. [46] proved that neuroblastomas with MYCN amplification had higher signal intensities (Mean and Maximum) and Energy on ADC map than tumours without MYCN amplification. Zhao et al. [47] also revealed that rectal cancer with lymph node metastasis had both a higher signal intensity (Maximum of the T2 and PD maps) and Energy than cancer without lymph node metastasis. The mapping of values in those three studies was consistent with Energy/Total Energy, so the effect of volume is not mentioned. However, according to our results, we believe that both the mapping value and the volume of the object must be considered when evaluating Energy and Total Energy.

There were several limitations in this study. First, the study sample was relatively small and was not divided into training and validation groups before the model was constructed. We recruited volunteers as control group, who did not perform nasopharyngoscopy and may have nasopharyngeal hyperplasia or inflammation. Therefore, a minimum of 2 years follow-up was conducted on control group, and no tumour sign was found. Due to the spectral deviation of diseases in our hospital and the incidence of tumours, other malignancies (such as lymphoma and adenoid cystic carcinoma) were not included in our research. Second, this was a single-centre study using a certain type of MRI scanner with an identical protocol. Multicentre studies with different MRI scanners and protocols are warranted before our findings are introduced into routine clinical practice. Last, due to the limited number of samples, the tumour stage, histological types and grades of NPC could not be considered, and a comparison between early-stage NPC and NPH could not be conducted in this study. A larger sample size is needed to investigate the correlation between the quantitative parameters of SyMRI and those factors.

The first-order parameters from SyMRI were significantly different between NPC and NPH and between OC invasion and non-invasion. Consequently, SyMRI was shown to be a reliable and useful tool, not only to provide additional quantitative value based on morphology for the diagnosis of NPC but also for evaluating the development of tumours by quantitatively measuring the changes in intrinsic characteristics in tissues.