Quantitative dynamic contrast-enhanced MR imaging can be used to predict the pathologic stages of lingual squamous cell carcinoma

To investigate whether quantitative DCE-MRI pharmacokinetic parameters can be used to predict the pathologic stages of lingual squamous cell carcinoma (LSCC). For this prospective study, DCE-MRI was performed in participants with LSCC 2016 2017. The pharmacokinetic parameters, including K trans , K ep , V e , and V p , were derived from DCE-MRI by utilizing a two-compartment extended Tofts model and a three-dimensional volume of interest. The postoperative pathologic stage was determined in each patient based on the 8th AJCC cancer staging manual. The quantitative DCE-MRI parameters were compared between stage I-II and stage III-IV lesions. Logistic regression analysis was used to determine independent predictors of tumor stages, followed by receiver operating characteristic (ROC) analysis to evaluate the predictive performance.


Abstract Background
To investigate whether quantitative DCE-MRI pharmacokinetic parameters can be used to predict the pathologic stages of lingual squamous cell carcinoma (LSCC).

Methods
For this prospective study, DCE-MRI was performed in participants with LSCC from May 2016 to June 2017. The pharmacokinetic parameters, including K trans , K ep , V e , and V p , were derived from DCE-MRI by utilizing a two-compartment extended Tofts model and a three-dimensional volume of interest. The postoperative pathologic stage was determined in each patient based on the 8th AJCC cancer staging manual. The quantitative DCE-MRI parameters were compared between stage I-II and stage III-IV lesions. Logistic regression analysis was used to determine independent predictors of tumor stages, followed by receiver operating characteristic (ROC) analysis to evaluate the predictive performance.

Results
The mean K trans , K ep and V p values were signi cantly lower in stage III-IV lesions compared with stage I-II lesions (P = 0.013, 0.005 and 0.011, respectively). K ep was an independent predictor for the advanced stages as determined by univariate and multivariate logistic analysis. ROC analysis showed that K ep had the highest predictive capability, with a sensitivity of 64.3%, a speci city of 82.6%, a positive predictive value of 81.8%, a negative predictive value of 65.5%, and an accuracy of 72.5%.

Conclusion
The quantitative DCE-MRI parameter K ep can be used as a biomarker for predicting pathologic stages of LSCC.

Background
Lingual squamous cell carcinoma (LSCC) is the most common malignancy of the oral cavity and comprises 25-40% of oral carcinomas. It has a more aggressive clinical behavior and a relatively poor prognosis compared with other oral cavity and head and neck cancers [1]. However, the prognosis of LSCC in early disease (TNM stage I-II) is better than that of advanced disease (TNM stage III-IV). For example, the ve-year survival rate in patients with stage I disease exceeds 80% [2], but it drops below 40% for those with advanced disease at the time of diagnosis [3]. Likewise, the treatment selection is highly dependent on the TNM classi cation of staging at diagnosis. Early stage tumors may be treated using a single modality (surgery or radiotherapy), while advanced tumors frequently bene t from multimodality therapy. Thus, accurate staging of LSCC prior to treatment is crucial for the treatment planning and prognosis prediction.
At present, the initial staging of LSCC relies on a panel of procedures, including physical examination, direct endoscopic examination, tumor tissue sampling, CT, and MRI, to assess tumor extension and in ltration as well as lymph node involvement. Among them, MRI is widely used to reveal the extent of soft tissue involvement and perivascular and perineural spread of LSCC [4,5]. However, there are limited quantitative imaging biomarkers with su cient sensitivity or speci city to predict the prognosis or stage of LSCC [5,6]. Quantitative dynamic contrast-enhanced MRI (DCE-MRI) can provide multiple pharmacokinetic parameters, such as model-free semiquantitative parameters and model-based quantitative parameters (parameters derived from pharmacokinetic model calculation). These parameters can fundamentally characterize the perfusion and vascularization of tissues and, indirectly, the state of the tumor [4][5][6]. Compared with semiquantitative parameters, quantitative parameters are less affected by wide variability in the MRI scanner, scanning sequence, temporal resolution, injection of contrast media, and image postprocessing calculation [4,7]. Previously, the clinical stages of oral squamous cell carcinoma, including LSCC, were found to be associated with quantitative parameters derived form DCE-MRI [8]. However, whether quantitative DCE-MRI (qDCE-MRI) pharmacokinetic parameters can be used to predict the pathologic stages of LSCC remains unknown so far.
In this study, DCE-MRI was prospectively performed in patients with LSCC. The pharmacokinetic parameters were derived from DCE-MRI data by using a two-compartment extended Tofts model and three-dimensional volume of interest (3-D VOI). The purpose of this study was to determine the role of quantitative DCE-MRI in predicting the pathologic stages of LSCC.

Patients
This prospective study was approved by the Ethics Committee of Sun Yat-Sen Memorial Hospital (Sun Yat-Sen University, Guangzhou, China), and written informed consent was obtained from all participants.
Between May 2016 and June 2017, consecutive patients with suspected LSCC on physical examination and/or CT were recruited. Patients were eligible for enrollment if they had a rst diagnosis of LSCC.
Exclusion criteria included biopsy of tongue lesion before MRI examination, no surgical resection, previous history of chemotherapy or radiation therapy in the head and neck region, a lesion smaller than 1 cm in the maximum diameter (to avoid large partial volume effect during the measurement of qDCE-MRI parameters), obvious motion artifacts in the MRI images, contraindications to either gadoliniumbased contrast material administration or MRI (e.g., metallic implant), or the inability to provide informed consent.
The total duration of the DCE acquisition was 5.5 min. After DCE imaging, conventional axial, sagittal and coronal contrast-enhanced T1WI were obtained with the same parameters as the unenhanced T1WI.

Imaging processing
The sequential DCE-MRI data were analyzed using a specialized quantitative analysis software (Omni Kinetics; GE Healthcare). A nonlinear registration framework utilizing the Free Form Deformation algorithm was rstly applied to correct the misalignment of body motion. Pharmacokinetic quantitative parameters were calculated from DCE images using a patient-speci c arterial input function (AIF) drawn on the common or external carotid artery ipsilateral to the tumor, the variable ip angle method and the two-compartment extended Tofts model [9]. To obtain the 3-D VOI, two experienced head and neck radiologists (N.G., with 4 years of experience in diagnostic imaging, and X.D. with 10 years of experience in diagnostic imaging), who were blinded for histologic results, independently drew the regions of interest (ROIs) slice by slice to encompass the entire lesion. Large feeding vessels and necrotic areas were excluded from the VOI. Quantitative parameters including K trans , K ep , V e , and V p were calculated. The pharmacokinetic parameters are described in Table 1.

Surgery and Histology
All patients were treated by surgery. Surgical resection was conducted within 7 days after MRI. The entire resected tongue specimens were processed for conventional histologic examination. The tumor invasion thickness, growth patterns (exophytic, ulcerated or endophytic), and pathologic TNM stages were recorded. The TNM staging was performed according to the 8th AJCC staging system [10].

Statistical analysis
Statistical analyses were operated using SPSS (Version 22.0, IBM SPSS Statistics, Armonk, NY, USA). All quantitative variables were showed as mean ± standard deviation. Interobserver agreement of the evaluation of qDCE-MRI parameters were evaluated using the intra-class correlation coe cient (ICC

Study population
Of the 56 patients enrolled, 5 were excluded because of the maximum diameter of the lesion < 1.0 cm (n = 2) and the presence of obvious motion or metallic artifacts on MRI (n = 3). Finally, 51 patients were included in this study, including 31 male and 20 female patients aged from 23-90 years, with a mean of 55.5 ± 14.6 years. There was a total of 51 tongue tumors in the 51 patients. The clinicopathologic characteristics of these patients are shown in Table 2. The 51 patients were divided into the early stage group (stage I-II, n = 23) and the advanced stage group (stage III-IV, n = 28). The tumor thickness in stage III-IV lesions was greater than that in stage I-II lesions (P < 0.001). Most of the stage I-II lesions were exophytic or endophytic types, while an exophytic growth pattern was predominant in the stage III-IV lesions.  Fig. 1. The diagnostic performances of these parameters are shown in   AUC, area under the curve; CI, con dential interval; PPV, positive predictive value; NPV, negative predictive value. qDCE-MRI with tracer pharmacokinetic modeling has emerged as a versatile technique for characterizing the microvasculature function of the tumor, including tissue perfusion, vessel permeability and extracellular leakage space by monitoring the delivery and distribution of intravascular contrast agent [4,5]; thus, it has been widely used for tumor detection and characterization, therapy monitoring and predicting prognosis in various clinical applications. Nevertheless, the relative low reliability of this technique restricts its adoption in routine clinical practice. There are many critical factors that in uence the reliability of qDCE-MRI, including baseline T1 mapping, temporal resolution, and AIF in data acquisition [4,5]. Baseline T1 mapping, which is used to compensate for the nonlinear relationship between MRI signal intensity and contrast agent concentration, is essential for accurate kinetic tting of acquired DCE-MRI data [4]. In our study, we used ve ip angles before injection of contrast agent to obtain the ideal baseline T1 mapping. Compared with other techniques of data acquisition for baseline T1 mapping (e.g., double ip angle technique, the inversion recovery technique, and the Look-Locker technique), the MFA method is now regarded as the technique of choice because it can provide more accurate, robust T1 mapping and kinetic parameter estimation with a short scan time but without sacri cing signal-to-noise ratio (SNR) [4,5]. In addition, the temporal resolution of DCE-MRI in our study was 3 sec, which was higher than what was found in most of the previous studies [8,11,12]. It has been suggested to use a temporal resolution from 1 to 5 sec, after which the errors of quantitative DCE-MRI parameters calculation grow rapidly with the decrease in temporal resolution [13]. The chosen temporal resolution of 3 sec in our study is an appropriate balance between the temporal resolution, SNR and spatial resolution, which allowed us to obtain high-quality DCE-MRI images and, in the meantime, capture the hemodynamic processes of contrast agents. AIF, which estimates the time course of the contrast agent concentration in the feeding arteries, is another crucial prerequisite for quantitative analysis of DCE-MRI. At present, the individual or population AIF can be used in DCE-MRI [4]. In our study, the AIF was extracted from individual patients rather than the population. Compared with the population AIF applied by previous studies [8,11,12], individual AIF could re ect the real AIF more closely, as it takes contrast agent injection rates and doses into account and presumes small intersubject variabilities [14]. In addition to the above key points in data acquisition, a 3-D VOI was used in our study. To date, most of previous studies have used a two-dimensional ROI (2-D ROI) derive the pharmacokinetic parameters from DCE-MRI for tumor assessment in head and neck cancer, while few studies have used a 3-D VOI for analysis [8,15]. Compared with 2-D ROI for tumor analysis, 3-D VOI can obtain the volumetric parameters and the heterogeneity data of the whole tumor, thus theoretically can more accurately describe the physiological characteristics of lesions [16].
Quantitative DCE-MRI has been determined as a useful tool for diagnosis and differential diagnosis, characterizing metastatic cervical lymph nodes, evaluating tumor cell proliferation and microvessel attenuation, predicting treatment response, and evaluating treatment outcome and prognosis in head and neck cancers [17][18][19][20]. Nevertheless, there have been a limited number of studies that have utilized quantitative DCE-MRI for predicting the staging of SCC in the head and neck, and discrepancies exist in the previous reports. For example, Chikui et al found that the clinical T stage of oral squamous cell carcinoma is negatively correlated with K trans and the N stage showed a negative correlation with K trans and V p [8]. In contrast, Leifels et al reported that the K ep was higher in HNSCC cancers with N2-3 stages; however, no differences were observed in DCE-MRI parameters between T1-2 and T3-4 tumors [15]. In our study, K trans , K ep and V p were found to be lower in pathologic stage III-IV lesions than in stage I-II lesions; these results are similar to that of Chikui et al [8] but different from that of Leifels et al [15]. This discrepancy might be related to the different protocol of DCE-MRI scanning as well as different method of pharmacokinetic analysis. In our study, the MFA method for T1 measurement, individual AIF, higher temporal resolution and 3-D VOI were applied. These data acquisition and DCE-MRI data analysis methods made our results more reliable than the previous studies, which used the dual ip angle method, population AIF, and temporal resolution of 3.5 ms [8] and 6 ms [15]. Additionally, the pathologic TNM stage was used in our study, which is different from the study by Chikui et al [8], which applied the clinical TNM stage. It is reasonable that our results are more favorable for reference in clinical practice.
Previously, the K trans and K ep values of breast cancer were found to increase with the degree of tumor malignancy [21]. Conversely, the advanced stage LSCCs in our study had lower K trans , K ep and V p values than early stage LSCCs. The different pathological features (i.e., tumor hypoxia) between LSCCs and breast cancers might contribute to this controversy. It has been shown that DCE-MRI parameters, such as K trans and K ep , are negatively correlated with tumor hypoxia [5]. In addition, the more invasive oral squamous cell carcinoma had more highly hypoxic areas but less vessel density because of the gradual destruction of microvessels during tumor growth [22][23][24]. K trans is positively coupled to blood ow, microvessel permeability and surface area, while K ep represents microvessel permeability [6,25].
Numerous studies have demonstrated that K trans was negatively correlated with the fraction of hypoxic cells in tumors [26,27], and there is a strong positive correlation between K ep and microvessel density in HNSCCs [16]. Therefore, the highly hypoxic areas but with low microvessel density might be an explanation why advanced stages of LSCC had lower K trans and K ep values in our study.
In our study, multivariate logistic analysis showed that K ep was an independent predictor for advanced stage LSCC. K ep had the highest predictive capability, with a sensitivity of 64.3%, a speci city of 82.6%, a PPV of 81.8%, a NPV of 65.5%, and an accuracy of 72.5%. K ep was more valuable for predicting the staging of LSCC compared with other DCE-MRI parameters such as K trans and V p . K ep , which represents the rate constant between the plasma and extracellular space, and is regarded as a marker that directly re ects microvessel permeability [16]. Previous studies have shown that the K ep positively correlated with the mean blood vessel count and mean vessel area fraction parameter [16]. The advanced stage LSCCs commonly had less vessel density because of the highly hypoxic areas [24], resulting in its lower K ep within tumors. Taken together, our results indicated that K ep can be used as a valuable predictive biomarker for tumor staging of LSCC.
There was no signi cant difference in V e between the stage III-IV lesions and the stage I-II lesions in our study. V e , which represents the volume of the extravascular extracellular leakage space, was mainly in uenced by cellular density and tumor interstitium [7]. Pathologically, the cellular density and tumor interstitium of LSCCs were variable among different stages and grades of tumor. Cell proliferation in advanced LSCCs may be more intensive than that of early stage LSCCs, which would cause high cellular density and contractible tumor interstitium resulting in low V e . Nonetheless, the rapidly growing late stage LSCCs may have large regions suffering from chronic or acute hypoxia in the central area, which may lead to a focal or extensive apoptotic response and then decrease the cellular density [28,29]. Therefore, V e can be in uenced by the varying cellular density between advanced and early stages of LSCC; thus, it is less robust for predicting the stages of LSCC compared with other qDCE-MRI parameters.
Our study had several limitations. First, the number of patients included was relatively small; a larger cohort is needed to con rm our results in a future investigation. Second, the enrolled patients did not receive follow-up. As a result, the correlation between qDCE-MRI parameters and survival outcomes remains unknown. Future follow-up investigation is needed to determine whether this method could be used to predict survival outcomes in patients with LSCC.

Conclusion
In summary, our study results showed that the mean K trans , K ep and V p values were higher in stage I-II LSCCs than in stage III-IV LSCCs. K ep was an independent predictor of stage III-IV LSCCs. The quantitative DCE-MRI-derived parameter K ep can be used as a predictive biomarker for pathologic stages of LSCC.

Consent for publication
Not applicable.

Availability of data and materials
The datasets used and analyzed during the current study are available from the corresponding author on reasonable request.

Competing interests
The authors declare that they have no competing interests.