Intra- and peritumoral MRI radiomics assisted in predicting radiochemotherapy response in metastatic cervical lymph nodes of nasopharyngeal cancer

The retrospective study received approval from our ethics committee and institutional review board, and informed consent was not required. Between October 2016 and November 2019, we enrolled 185 patients with NPC who received induction chemotherapy (IC) and CCRT. All patients received an individualized treatment based on NCCN guidelines [23]. The details of treatment regimens are shown in Additional file 1 (Additional file 1: S1). The inclusion criteria were as follows: (1) Primary tumor confirmed by biopsy as NPC; (2) available pretreatment T2-weighted imaging (T2WI) and contrast-enhanced T1-weighted imaging (CE-T1WI) images after biopsy; (3) available post-treatment T2WI and CE-T1WI images for predicting treatment response; (4) no treatment before baseline biopsy; and (5) available clinical variables such as age, sex, T-stage, N-stage, clinical stage, lymph node involvement, and lymph node gross tumor volume (GTV-ln). The exclusion criteria were as follows: (1) the period between baseline MRI and initial treatment was further than 2 weeks (n = 14); (2) the existence of other malign tumors (n = 9); (3) missing clinicopathological information (n = 10); (4) poor image quality (n = 7). Finally, a total of 145 consecutive patients (mean age 47.2 ± 12.6, range 13–77, male = 117, female = 28, stage II = 2, stage III = 59, stage IV = 84) were enrolled and allocated to the training set (102 patients, October 2016 to December 2018) and validation set (43 patients, December 2018 to November 2019). [22] (Fig. 1). The time all patients who underwent MRI examination after the completion of treatment was 2–14 weeks. Patients were staged based on the 8th version American Joint Committee on Cancer (AJCC) Tumor-Node-Metastasis (TNM) staging system [24]. Demographic information and stages were extracted from the Hospital Information System of our institution.

All patients underwent nasopharyngeal and cervical region MRI examination using a 1.5-T MR scanner (Avanto, Siemens, Germany) or a 3.0-T MR imaging scanner (Skyra, Siemens, Germany) with head-neck combined coils. Overall, 94/145 (64.8%) patients underwent imaging in the 1.5-T MRI scanner, and 51/145 (35.2%) underwent imaging in the 3.0-T scanner. To keep away from magnetic exposure and motion artifacts, patients were instructed to eliminate all metal-containing items and lie supinely in the scanner before scanning. Before the MRI examination, all patients will be required to wear earplugs and headphones to decrease noise. The DICOM format images of axial fat-suppressed CE-T1WI (contrast agent = Gd-DTPA, Magnevist, Schering, Berlin, Germany; dose = 0.1 mmol/kg body weight) and T2WI for each case were retrieved from the picture archiving and communication system (Carestream, Ontario, Canada). The details of MRI scanning parameters are shown in Additional file 1 (Additional file 1: S2). MRI images were reconstructed using the inverse fourier transform with the linear filling algorithm.

Multiple radiologic criteria were used to determine whether metastatic cervical lymph nodes were involved [25‐28]. These included: (1) regions of central necrosis/cystic necrosis (T2WI with a focal high signal intensity or CE-T1WI with low signal intensity/ with or without an adjacent border of enhancement), (2) extracapsular distribution in any size lymph node, including ambiguous nodal boundaries, irregular nodal capsular improvement, and infiltration into neighboring muscle or fat, (3) the shortest diameter of the cervical or medial retropharyngeal lymph node is 10 mm; the lateral retropharyngeal lymph node is 5 mm.

In each patient, only the largest metastatic lymph node within the scanning range is considered. The largest lymph nodes were then assessed on T2WI for maximum axial diameter and minimum axial diameter long axis. The minimum axial diameter matched the node’s largest diameter in the axial plane perpendicular to its maximum axial diameter. The Response Evaluation Criteria in Solid Tumors 1.1 (RECIST) was then used to assess the largest lymph node’s response after treatment [29]. The responders were defined as a reduction of at least 30% in the largest lymph node’s maximum axial diameter. In contrast, non-responders had inadequate shrinking to qualify for the responders.

All pretreatment MRI images were transmitted to the Artificial Intelligence Kit software (A.K., version 3.2.0, GE Healthcare, China) for preprocessing. Firstly, all raw images were reconstructed using the trilinear interpolation algorithm to a final voxel size of 1 mm ×1 mm × 1 mm. The center was employed to align the interpolation grid’s location, and its measurements were rounded to the closest integer. Secondly, the Gaussian filter and bias-field correction were applied after the reconstruction. Lastly, to eliminate the influence of the different ranges of gray values, image was normalized using z-score normalization.

Three-dimensional manual segmentation of the largest lymph nodes and region of interest (ROIs) were delineated by two radiologists with 7 and 20 years of clinical diagnosing experience using ITK-SNAP software (version 3.8.0, http://​www.​itksnap.​org) to generate Intra ROIs. Both radiologists were blinded to the clinical information. The ROIs based on the lymph nodes were drawn to cover the whole tumor on each consecutive slice of the T2WI and CE-T1WI images separately, necrosis and extranodal extension regions were avoided when delineated ROIs. A distance of 2 mm from the tumor boundary was defined as the Peri region. The Peri ROI was then obtained by subtracting the Intra ROI from the dilated ROI. Figure 2 showed a typical MRI image and its associated Intra and Peri ROIs. The reproducibility of radiomic features was evaluated using the intraclass correlation coefficient (ICC). The senior radiologist (PZ) re-delineated 30 lymph nodes that were randomly selected from the T2WI and CE-T1WI images, respectively. Feature with an ICC > 0.75 was regarded as having good reproducibility and remained.

Radiomic features from the Intra and Peri ROIs were extracted using an open-source Python software (PyRadiomics, version 3.0, https://​pyradiomics.​readthedocs.​io) [30] that followed the image biomarker standardisation initiative (IBSI) standard [31]. These features belonged to three categories: shape, firstorder, and texture features. Image intensity was discretized using a fixed bin width of 25. Detailed descriptions of these features are provided in Table S1 and Table S2 in the Additional file 1. Radiomic features were standardized using z-score normalization. Five procedures were performed to select radiomic features. We used the Mann-Whitney U test to make the initial selection from the training set. For the remaining significant features, the P-value threshold was set at 0.05. Then, to eliminate redundant features, Spearman correlation assessment and maximum relevance-minimum redundancy (mRMR) were performed successively. Features with Spearman correlation coefficient values greater than 0.9 were eliminated. 15 features with high relevance and low redundancy were retained after mRMR. Finally, the least absolute shrinkage and selection operator (LASSO) algorithm and multivariate logistic regression using Akaike information criterion as the stopping rule were applied to select the most predictive features [32, 33].

Radiomic models (Intra, Peri, and Intra + Peri) were built using the radiomics score (Rad-score). The Rad-score was computed utilizing a linear combination of the chosen features weighted by their corresponding coefficients. To examine possible multicollinearity between the features included in the radiomic model, Spearman’s correlations were employed.

Univariate logistic regression analysis was utilized to recognize independent risk factors for differentiating responders from non-responders between radiomic models and clinical variables. Subsequently, a clinical-radiomic model combining the radiomic model and significant independent risk factors was built in the training set, using the multivariable logistic regression analysis.

The predictive performance of the radiomic model and clinical-radiomic model were assessed in the training set and then validated in the validation set using the area under the receiver operating characteristic curve (ROC-AUC). The threshold was established utilizing the maximum Youden index (sensitivity + specificity − 1) on the training set and this same threshold was used for the validation set. We also calculated the accuracy, sensitivity, and specificity. The 95% confidence intervals (CIs) were calculated using the exact binomial method. Calibration curves were plotted to evaluate the calibration performance of radiomic models. The Hosmer-Lemeshow (H-L) test was used to assess the goodness-of-fit of radiomic models [34].

SPSS (version 26.0, available at https://​www.​ibm.​com), MedCalc (version 18.2.1, available at https://​www.​medcalc.​org), and R software (version 4.0.0, http://​www.​r-project.​org) were utilized for all statistical analyses. The Shapiro-Wilk test was used to examine if quantitative data had a normal distribution. The clinical variables of the responders and non-responders were compared using an independent samples t-test, the Mann-Whitney U test, Fisher’s exact analysis, or the chi-squared (χ2) test, as applicable. The DeLong test was used to compare the AUCs among the radiomic models [35]. P < 0.05 was considered statistically significant (two-tailed).