Development and validation of an MRI-based radiomic model for predicting overall survival in nasopharyngeal carcinoma patients with local residual tumors after intensity-modulated radiotherapy

The Medical Ethics Committee of First Affiliated Hospital of Guangxi Medical University approved the study (NO. 2022-KY-E-249), and the informed consent of the patients was waived for the nature of the retrospective study.

The study cohort was from an endemic region where nonkeratinizing pathological subtype accounts for more than 95% of the cases [18]. We retrospectively reviewed 925 nonmetastatic patients with NPC who received IMRT in our hospital from May 2015 to December 2017. Participants were selected according to the inclusion and exclusion criteria below. Inclusion criteria included the following: (1) NPC was pathologically confirmed; (2) the initial MRI scans were completed prior to treatment, and there was no more than 1 week between receiving treatment and the initial MRI scans; (3) local residual tumors were diagnosed immediately after the completion of IMRT. The exclusion criteria were as follows: (1) the patients had been treated in other hospitals prior to their admission to our hospital; (2) those who were unable to complete the radiotherapy course; (3) those whose MRI images were marred by artifacts caused by patient movement, oral metal materials, or other factors; (4) patients with other malignant tumors or NPC with distant metastasis; and (5) those lost to follow-up (with a loss to follow-up rate of 4.4%). Finally, 218 eligible NPC patients with local residual disease after IMRT were enrolled in this study, and the patients were randomly divided into training and validation cohorts according to the ratio of 8:2. Demographic characteristics of the study population, the American Joint Committee on Cancer (AJCC) staging of NPC, pretreatment hemoglobin level, neutrophil to lymphocyte ratio (NLR), albumin level, as well as chemotherapy strategies (induction chemotherapy, concurrent chemotherapy, adjuvant chemotherapy) were collected. The tumors were restaged in accordance with the eighth edition of the AJCC staging manual.

The treatment of NPC is carried out under the relevant guidelines: definitive radiotherapy was applied to patients with T1N0M0; concurrent chemoradiotherapy followed or not followed by adjuvant chemotherapy or induction chemotherapy followed by chemoradiotherapy for patients with T1, N1-3, M0, and T2–T4, any N, M0 [19, 20]. All patients completed the course of IMRT. Details on the radiotherapy protocol are included in Additional file 1. Platinum-based regimens were used for chemotherapy. During the first 3 years following IMRT, patients were followed-up every 3 months, every 6 months during the fourth and fifth years, and once a year after the fifth year. OS, the endpoint of our study, was defined as the time from the completion of IMRT to the date of death or last follow-up. The patient's survival status was ascertained through telephone consultation with the patient or his family. Living patients were censored between July and August 2021.

All MRI scanning of the patients was performed on a GE Signa EXCITE 1.5 T MRI Scanner (GE Medical Systems, USA). Scanning sequences included axial fast relaxation fast spin echo (FRFSE) T2 weighted imaging (T2WI) sequence, axial/sagittal/coronal unenhanced T1WI Flair sequences, and axial/sagittal/coronal CE-T1WI Flair sequences with fat suppression. Gadopentetate Dimeglumine Injection (Bayer Healthcare Pharmaceuticals Inc.) was used as the contrast agent at a dose of 0.2 ml/kg body weight (up to 10 ml) and an injection rate of 2 ml/s. Radiomic features were extracted from pretreatment axial T2WI and CE-T1WI sequences. The main acquisition parameters of axial FRFSE T2WI sequence included: repetition time (TR) = 2620 ms, echo time (TE) = 101 ms, field of view (FOV) = 220 mm × 220 mm, matrix = 320 × 160, flip angle = 90°, slice thickness = 6 mm, intersection gap = 1 mm. The main acquisition parameters of axial CE-T1WI sequence included: TR = 2162 ms, TE = 10 ms, FOV = 220 mm × 220 mm, matrix = 320 × 160, flip angle = 90°, slice thickness = 6 mm, intersection gap = 1 mm.

Post-treatment MRI scanning was performed immediately after the completion of the course of IMRT to assess tumor regression. Blind to the clinical outcomes of the patients, a head and neck oncology radiologist (Investigator A) with 9-year experience and a head and neck radiation oncologist with 14-year experience (Investigator B) evaluated the MR images, respectively. The divergence between the two investigators regarding the evaluation of the MR images will be adjudicated through a third head and neck oncology radiologist (Investigator C) with 30-year experience to achieve agreement. The diagnostic criteria of local residual tumors on MRI were as follows: (1) residual tumors in the nasopharynx, parapharyngeal soft tissues, or intracranial spaces exhibit low signal intensity on T1WI, intermediate signal intensity over muscle on T2WI, and moderate signal enhancement on CE-T1WI [21]; (2) lesions of the skull base were thought to be tumor residuals if the bone of the skull base was destroyed by soft tissues and the degree, and/or extent of bone strengthening was not diminished or increased compared with pretreatment images [11, 22]. Figure 1 shows MRI images of the local residual tumor in a patient following IMRT.

The workflow of radiomics analysis is shown in Fig. 2. The delineation of the regions of interest (ROI) and features extraction were carried out on the RadCloud Radiomics Cloud Platform (Huiying Medical Technology Co., Ltd) using pretreatment T2WI and CE-T1WI sequences of the nasopharynx. Radiomic features extraction was based on pyradiomics 2.2.0.

We randomly selected 30 patients from the study population to assess the inter-and intra-class correlation coefficients (ICCs). Firstly, each of the 218 patients was assigned a number between 1 and 218; after that, 30 of those numbers were chosen at random. Blind to the clinical outcomes of the patients, Investigator A and Investigator B delineated the ROI of the entire tumor layer by layer in the patient's nasopharyngeal T2WI and CE-T1WI images independently. Investigator A repeated the process 2 weeks later. Before feature extraction, images were resampled using the B-spline interpolator to a pixel spacing of 1.0 mm × 1.0 mm × 1.0 mm and normalized based on the mean and standard deviation of gray values.

1409 features for each patient and each sequence were extracted from the original images (107 features) and derived images using filters including logarithm (93 features), exponential (93 features), gradient (93 features), square (93 features), squareroot (93 features), localBinaryPattern2D (lbp-2D, 93 features), and wavelet (744 features). These 1409 features included first order statistics (270 features), Shape-based (14 features), Neighbouring Gray Tone Difference Matrix (NGTDM, 75 features), Gray Level Co-occurrence Matrix (GLCM, 360 features), Gray Level Dependence Matrix (GLDM, 210 features), Gray Level Run Length Matrix (GLRLM, 240 features), and Gray Level Size Zone Matrix (GLSZM, 240 features). A total of 2818 features were extracted from each patient's images. Most of the features extracted conform to IBSI definitions of features. The pyradiomics documentation provides explanations of these radiomic features (https://​pyradiomics.​readthedocs.​io/​en/​latest/​index.​html). Z-score normalization of the features was performed to eliminate the effect of extreme values and facilitate the comparison of features of different magnitudes. Features with ICCs > 0.75 reveal good reproducibility and will be used for further analysis. Subsequently, Investigator A conducted lesion segmentation and feature extraction on the remaining samples.

Firstly, the features with ICCs > 0.75 were tested for the proportional hazards assumption in the training cohort, and then the features that met the proportional hazards assumption were included in the univariate Cox regression analysis. The least absolute shrinkage and selection operator (LASSO)-Cox regression [23] based on statistically significant features with a p-value < 0.05 of the univariate Cox regression analysis was conducted. The optimal lambda was selected based on tenfold cross-validation. Features with non-zero coefficients were selected to establish the radiomic signature. Radscore is calculated based on the value of the features and their coefficients multiplied together.

Continuous variables are expressed as the means and standard deviations if they are normally distributed; otherwise, they are expressed as the medians and interquartile ranges. Categorical data are expressed as frequencies and percentages. Independent sample t-tests or Mann–Whitney U tests were used to compare continuous variables between the training and validation cohorts. χ² or Fisher exact tests were used for the comparison of categorical variables between the two groups. The clinical model for predicting OS was established using the independent clinical predictors identified by multivariate Cox regression in the training cohort. The radiomic model was established using radscore-based univariate Cox regression. The combined model, which integrated significant clinical factors and radscore, was constructed using stepwise back multivariate Cox regressions based on Akaike's Information Criterion (AIC). The Harrell’s concordance index (C-index) and 2-year, 3-year, and 5-year receiver operating characteristics (tROC) curves were used to evaluate the models’ discrimination. To prevent the influence of a single random training-validation set split leading to unreliable results, the discrimination of the different models was determined by fivefold cross-validation with 200 repetitions. The agreement between the model-predicted OS probability and actual observed survival rate was evaluated using calibration curves, and the clinical usefulness was evaluated based on decision curve analyses. Likelihood ratio tests for comparing nested models or the AIC were used to determine which model fit the data better in the training and validation cohorts. Kaplan–Meier method with log-rank tests was used to compare the difference in survival between high-risk and low-risk groups. All statistical analysis were conducted using SPSS 22.0 (SPSS Inc., Chicago, IL) or R 4.1 (R Project for Statistical Computing). Two-tailed p < 0.05 is considered statistically significant.

The clinical characteristics of the patients are shown in Table 1. A total of 218 eligible patients were included in this study, including 164 males (75.2%) and 54 females (24.8%), aged 46.1 ± 11.4 years (range 16.0–81.0 years). The median follow-up time in this study was 50.4 months (range of 4.0–72.0 months). Fifty-two patients died during follow-up of the entire cohort, including 41 deaths in the training cohort and 11 deaths in the validation cohort. The 2-year, 3-year, and 5-year OS rates were approximately 91.3%, 83.5%, and 74.0%.

Age, gender, and albumin level were independent predictors of OS in the training cohort (Table 2). The clinical nomogram was constructed using these three independent predictors. The C-index of the clinical nomogram predicting OS was 0.672 (95% CI 0.599–0.745) in the training group. Cumulative/dynamic time-dependent ROC (tROC) revealed that the area under curve (tAUC) of the clinical model predicting 2-year, 3-year, and 5-year OS in the training group was 0.659 (95% CI 0.528–0.790), 0.667 (95% CI 0.572–0.763), and 0.685 (95% CI 0.572–0.798), respectively.

The C-index of the clinical nomogram predicting OS was 0.634 (95% CI 0.593–0.675) in the validation group. The tAUCs of the clinical model predicting 2-year, 3-year, and 5-year OS in the validation group were 0.690 (95% CI 0.491–0.890), 0.646 (95% CI 0.488–0.804), and 0.620 (95% CI 0.398–0.841), respectively. The 2-, 3-, and 5-year mean tAUCs of the clinical model determined by fivefold cross-validation with 200 repetitions were 0.657 ± 0.126, 0.642 ± 0.094, and 0.647 ± 0.115, respectively (Fig. 3A).

Of the 2818 features extracted from T2WI and CE-T1WI sequences, 2150 features with ICCs > 0.75 were tested for the proportional hazards assumption in the training cohort. Of the features eligible for the proportional hazards assumption tests, 195 features were significant in univariate Cox regression analysis. The 195 features were then included in the LASSO-Cox regression analysis, and an optimal Lambda (λ) = 0.084 with Log (λ) = − 2.47 was selected using tenfold cross-validation via one standard error rule (Fig. 4A, B). Finally, five features with non-zero coefficients were selected. The selected features and their coefficients are shown in Table 3. The equation to calculate the radscore for each patient was as follows: Radscore = 0.081381 × T2WI_wavelet.LLL_Glrlm_ShortRunHighGrayLevelEmphasis + 0.061505 × CE-T1WI_original_Gldm_DependenceVariance + 0.142812 × CE-T1WI_original_ Ngtdm_Busyness + 0.016691 × CE-T1WI_squareroot_Firstorder_Variance—0.423306 × CE-T1WI_wavelet.LLH_Gldm_LargeDependenceLowGrayLevelEmphasis. The radscore for each patient is shown in Fig. 4C, D, and Additional file 2. The radscore of the training and validation cohorts was not significantly different (P = 0.903).

The C-index of the radiomic model (Fig. 5) predicting OS was 0.788 (95% CI 0.724–0.852) and 0.753 (95% CI 0.604–0.902) in the training and validation cohorts, respectively. The tAUCs of the radiomic model predicting 2-year, 3-year, and 5-year OS in the training group were 0.826 (95% CI 0.727–0.925), 0.848 (95% CI 0.772–0.924), and 0.746 (95% CI 0.628–0.863), respectively, and the corresponding tAUCs in the validation group were 0.774 (95% CI 0.540–1.000), 0.781 (95% CI 0.617–0.944), and 0.719 (95% CI 0.495–0.944). The 2-, 3-, and 5-year mean tAUCs of the radiomic model determined by fivefold cross-validation with 200 repetitions were 0.802 ± 0.120, 0.817 ± 0.083, and 0.743 ± 0.113, respectively (Fig. 3B).

Calibration curves showed good agreement between the radiomic model-predicted probability of 2- and 3-year OS and the actual observed probability in the training and validation groups (Fig. 5C, E). Decision curve analysis showed that the radiomic model had higher clinical usefulness than the clinical model (Fig. 6).

Age, gender, albumin level, and radscore were ultimately included in the combined model (Additional file 3). The C-index of the combined model predicting OS was 0.834 (95% CI 0.777–0.891) and 0.734 (95% CI 0.688–0.780) in the training and validation groups, respectively. In the training group, tAUCs for predicting 2-year, 3-year, and 5-year OS were 0.848 (95% CI 0.745–0.951), 0.869 (95% CI 0.794–0.943), and 0.815 (95% CI 0.712–0.919), respectively, and the corresponding tAUCs in the validation group were 0.756 (95% CI 0.516–0.997), 0.759 (95% CI 0.589–0.930), and 0.661 (95% CI 0.430–0.892). The 2-, 3-, and 5-year mean tAUCs of the combined model determined by fivefold cross-validation with 200 repetitions were 0.813 ± 0.123, 0.821 ± 0.085, and 0.770 ± 0.107, respectively (Fig. 3C).

The likelihood-ratio test demonstrated that the combined model had better goodness of fit than the clinical and radiomic models in the training cohort but not in the validation cohort (Table 4). The AIC revealed better goodness of fit of the radiomic models than the clinical models in the training and validation cohorts, with an AIC of 356.65 versus 385.53 and an AIC of 76.34 versus 79.24, respectively.

Patients were divided into high- and low-risk groups according to the radscore cutoff (0.062261) obtained from the training cohort with maximally selected test statistics. Patients in the high-risk group had a significantly lower OS rate than those in the low-risk group in the training and validation cohorts (p < 0.01) (Fig. 7).