Introduction

Lower-grade gliomas of World Health Organization (WHO) grades II and III are infiltrative neoplasms with variable clinical outcomes [1]. A mutation in isocitrate dehydrogenase (IDH) is one of the key genetic events which leads to the stratification of gliomas with significantly different survival rates in grade II/III gliomas as well as glioblastomas [1,2,3,4]. Among lower-grade gliomas of histologic grades II and III, IDH wild-type (IDHwt) tumors constitute less than 30% of cases and have been reported to follow worse prognosis than those with IDH mutations [2, 5, 6]. Moreover, a previous study suggested that the majority of IDHwt lower-grade gliomas with grade II/III were considered as the molecular equivalent of conventional glioblastoma [7].

Even though the presence of an IDH mutation is a powerful prognostic factor, heterogeneous clinical outcomes have been reported in IDHwt lower-grade gliomas according to the variable combination of genetic profiles [8,9,10]. Recently, the Consortium to Inform Molecular and Practical Approaches to CNS Tumor Taxonomy (cIMPACT-NOW) provided the description of a “diffuse astrocytic glioma, IDH-wildtype, with molecular features of glioblastoma, WHO grade IV,” which corresponds to grade II/III IDHwt gliomas with high-level EGFR amplification, the combination of a whole chromosome 7 gain and a whole chromosome 10 loss (+ 7/− 10), or TERT promoter mutations [11]. These specifications highlight that the specific subset of IDHwt lower-grade gliomas follow an aggressive clinical course more closely resembling that of an IDHwt glioblastoma. Thus, it would be beneficial if MRI could stratify IDHwt lower-grade glioma patients according to their risks and identify patients with worse prognosis.

Radiomics exploits MRI data and extracts high-dimensional quantitative imaging features such as intensity distributions, spatial relationships, textural heterogeneity, and shape descriptors [12]. Because radiomics models use high-throughput imaging features, hidden information, which may be visually imperceptible, could be revealed [13]. Radiomics has been applied in many previous studies to predict a specific mutation status, stratify risk, or predict treatment responses in patients with glioblastomas [14,15,16,17]. Several previous studies have also used radiomics in patients with lower-grade gliomas to predict survival [18, 19] or detect molecular subtypes, including IDH mutation status [20,21,22,23]. However, the prognostic significance of radiomics in patients with IDHwt lower-grade gliomas has not been well studied.

We hypothesized that MRI radiomics can independently predict the survival in patients with IDHwt lower-grade gliomas. Therefore, the purpose of this study was to evaluate whether radiomics from MRI would allow prediction of survival in patients with IDHwt lower-grade gliomas and to investigate the added prognostic value of radiomics over clinical features.

Materials and methods

This retrospective study was approved by the institutional review board of our hospital, and the requirement for obtaining informed patient consent was waived. The NIH/NCI-approved Cancer Genome Atlas (TCGA) and The Cancer Imaging Archive (TCIA) databases are publicly available datasets in which all data are anonymized [24]. Thus, individual institutional approval was not required for the external validation set.

Patients

Institutional cohort

From January 2007 to February 2018, 459 patients with pathologically confirmed lower-grade gliomas were identified. Patients with IDHwt lower-grade gliomas who underwent preoperative MRI were included. Patients were excluded if they presented any of the following: (1) IDH mutant tumors (n = 202); (2) unknown IDH mutation status (n = 91); (3) previous history of brain surgery or treatment (i.e., radiation therapy or chemotherapy) (n = 21); (4) age under 18 years old (n = 16); (5) no preoperative MRI performed (n = 12). Thus, 117 patients with IDHwt lower-grade gliomas were enrolled in this study (Fig. 1). These included 35 (29.9%) WHO grade II gliomas and 82 (70.1%) grade III gliomas.

Fig. 1
figure 1

Flow chart of the patient population

Full size image

The extent of tumor resection was determined by visually comparing the preoperative and postoperative lesion volume on both T2-weighted images (T2WI) and contrast-enhanced T1-weighted imaging (T1C), in conjunction with intraoperative impressions of the surgeons, classified as one of three categories: gross total, subtotal (< 100% and ≥ 75% of gross tumor removal)/partial (< 75% of gross tumor removal) resection, or biopsy only. Karnofsky Performance Status (KPS) scores which were determined preoperatively and postoperative treatment (i.e., chemotherapy, radiation therapy) were retrieved from their electronic medical records.

Overall survival (OS) was used as a clinical outcome. OS was defined as the number of days from the initial surgery, i.e. tumor resection or biopsy, to either patient death or the date of the last follow-up if the patient did not die.

External validation cohort

TCGA (http://cancergenome.nih.gov) provides 265 lower-grade gliomas of histologic grades II and III, within which 46 were IDHwt. Among them, 33 patients (10 grade II and 23 grade III) with preoperative MRI containing both T2WI and T1C images were retrieved and constituted the external validation cohort. Information including age, extent of resection, WHO grade, and death were available for these 33 patients. KPS was known for 22 patients.

MR image acquisition

A 3.0-T MRI scanner (Achieva, Philips Healthcare) and an 8-channel SENSE head coil were used for all preoperative MRI scans. The preoperative MRI protocol included T2WI (TR/TE, 2800–3000/80–100 ms; FOV, 240 mm; section thickness, 5 mm; matrix, 256 × 256) and three-dimensional T1C (TR/TE, 6.3–8.3/3.1–4 ms; FOV, 240 mm; section thickness, 1 mm; matrix, 192 × 192). T1C images were acquired after administering 0.1 mL/kg gadolinium-based contrast material (Gadovist, Bayer).

The preoperative T2WI and T1C of 33 patients from the TCGA datasets were downloaded from TCIA (http://www.cancerimagingarchive.net).

Image pre-processing and radiomic feature extraction

First, T2WI and T1C images were resampled to an identical spatial resolution of 1 × 1 × 1 mm using nilearn (https://nilearn.github.io). Next, those images were subjected to N4 bias correction to remove low-frequency intensity and non-uniformity from the images [25, 26]. After resampling and N4 bias correction, T1C images were registered to identical spatial coordinates using T2WI as a template using SimpleITK (http://www.simpleitk.org). Signal intensity was normalized using the WhiteStripe R package [27], implemented in R version 3.4.0 (R Foundation for Statistical Computing). Tumor segmentation was then performed by C.J.P., a neuroradiologist with 6 years of experience, with a semi-automatic method using 3D slicer (www.slicer.org) [28]. The software automatically selected the T2 hyperintense lesion where the signal is above the specific threshold, which corresponds to the infiltrative tumor and surrounding edema. Then, C.J.P. revised the automatically selected region of interest (ROI) by excluding erroneously detected cerebrospinal fluid or correcting the tumor border, as the periphery of the tumor could not be exactly determined by the software due to the infiltrative nature of the gliomas. T1C was not referred for the tumor segmentation as it is difficult to determine the tumor border in the T1C. Another neuroradiologist (S.S.A.) with 13 years of experience subsequently re-evaluated and confirmed the ROIs.

In addition, to evaluate the reproducibility of tumor segmentations, we randomly selected 50 patients in whom we obtained two different ROIs from two different readers (C.J.P., and S.S.A.). Then, we calculated the Dice coefficient (DICE), which is the most commonly used metric in validating medical volume segmentations [29, 30]. The DSC measure represents the relative overlap between two binary volume data and is expressed in the equation below.

$$ \mathrm{DICE}=\frac{2\ X\ V\ \left(R1\cap R2\right)}{V(R1)+V(R2)} $$

V denotes the volume of binary data; R1 and R2 represent the ROIs from reader 1 and reader 2, respectively. The DICE score is 1.0 when the two volumes match perfectly.

Radiomic features were extracted from the ROIs on T2WI and T1C images by using Pyradiomics 1.2.0 (http://www.radiomics.io/pyradiomics.html) [31].

In total, 13 shape, 18 first-order, 23 gray-level co-occurrence matrix (GLCM), 16 gray-level run length matrix (GLRLM), 16 gray-level size zone matrix (GLSZM), and 5 neighborhood gray tone difference matrix (NGTDM) were extracted from the ROIs on T2WI and T1C images, constituting a total of 182 radiomic features. A schematic workflow of image pre-processing and radiomic feature extraction is shown in Fig. 2.

Fig. 2
figure 2

The workflow for imaging data pre-processing and radiomic feature extraction

Full size image

Pathological evaluation and molecular subtyping

All surgical specimens were histopathologically diagnosed according to the 2016 WHO classification. Both peptide nucleic acid–mediated clamping polymerase chain reaction and immunohistochemical analyses were performed to detect the presence of a IDH1-R132H mutation [2]. Monoclonal antibody H09 was used for immunohistochemical analysis. The degree of IDH1-R132H staining was determined positive if there were any stained cells. Cases without IDH1-R132H staining were determined to be negative [32, 33]. In IDH1-negative cases, IDH1/2 status was confirmed by a peptide nucleic acid–mediated clamping polymerase chain reaction.

Statistical analysis

Statistical analysis was performed in R software (version 3.5.1; R Foundation for Statistical Computing) (www.R-project.org). Because of the relatively large number of imaging variables compared with the number of events, the least absolute shrinkage and selection operator (LASSO) and elastic net were used to select important features, to minimize the potential risk for overfitting, by shrinking the regression coefficients of irrelevant variables toward zero [34]. The performances of these methods were tested by 10-fold cross-validation with 100 replications to enhance the generalizability of the results. The “glmnet” with “coxnet” and the “glmnetUtils” R packages were used for the LASSO and elastic net, respectively [35, 36]. A radiomics risk score (RRS) was calculated for each patient using a linear combination of features frequently selected from the LASSO and elastic net, weighted according to their regression coefficients. Univariable analysis of RRS and clinical features—age, KPS, extent of resection, WHO grade, and postoperative treatment—for OS prediction was performed. Subsequently, the variables which were significantly associated with OS were included for multivariable Cox regression analyses to create prognostic models: (1) model 1, clinical features only; (2) model 2, clinical features and RRS. To assess the discriminative ability, we obtained time-dependent receiver operating characteristic (ROC) curves and calculated the integral of the area under the curves (iAUC) [37]. The difference in the iAUC of the two models in the institutional cohort was tested based on a 95% confidence interval (CI) from a bootstrap with 1000 resampling, and internal validation for the developed model was performed using a bootstrap on the same resampled data. The iAUC difference was considered statistically significant if the 95% CI of the iAUC difference did not contain a zero value. Furthermore, the Akaike information criterion (AIC) was used to compare competing prognostic models, with a smaller AIC indicating a better model fit [38]. Additionally, likelihood ratio chi-square tests were performed to compare model fitness in terms of the log likelihood (LL) between two nested models, which was presented as minus 2LL (− 2LL).

With regard to the external validation, the prognostic value of RRS was assessed by comparing the iAUCs of model 1 (clinical features only) and model 2 (clinical features and RRS), which was tested based on a 95% CI from a bootstrap with 1000 resampling. In this analysis, 11 patients who lacked the KPS in the TCGA dataset were excluded.

In addition, the optimal cutoff values of RRSs were defined by the log-rank test, using the Contal and O’Quigley’s method, which was performed using the “cutp” function of “survMisc” in R [39]. The patients in the institutional and external validation cohort were then classified into low-risk and high-risk groups according to fixed cutoff values which were derived from institutional cohort. Differences in clinical characteristics between the low-risk and high-risk group were evaluated using Student’s t test and chi-square test. The Kaplan-Meier curves of those low-risk and high-risk groups were compared to evaluate the relationship between the RRS and OS. A p value < 0.05 was considered to be statistically significant.

Results

Characteristics of the 150 enrolled patients are summarized in Table 1. In the institutional cohort and external validation set, the median OS was 693 days (interquartile range, 384–1168 days) and 269 days (interquartile range, 155–582 days), respectively. The extent of resection was significantly different between the two groups (p < 0.001), but other clinical features were not significantly different.

Table 1 Patients’ clinical characteristics
Full size table

The DICE scores which were calculated to assess the reproducibility of segmentations from two different readers ranged from 0.809 to 0.992 (mean ± standard deviation, 0.966 ± 0.035), which represents high reproducibility.

LASSO and elastic net selected 13 and 80 radiomic features for OS prediction, respectively, that were the most important for predicting tumor outcome. These radiomic features ranked according to their regression coefficients are listed in Supplementary Tables 1 and 2. RRS_LASSO and RRS_elasticnet were derived from the sum of features selected by LASSO and elastic net multiplied by each regression coefficient, respectively (Supplementary Tables 1 and 2). Thirteen features that were selected from both LASSO and elastic net included 12 texture features from GLSZM, NGTDM, and GLCM and one shape feature. The texture feature zone percentage from T1C was the parameter with the highest negative coefficient, associated with a favorable prognosis.

Univariable analysis of clinical features and RRS revealed that age, KPS, extent of resection, WHO grade, and RRS showed significant association with OS. Those selected features were included in the multivariable Cox models. Multivariable Cox models with and without RRS for OS prediction are summarized in Table 2. Both RRS_LASSO and RRS_elasticnet were independent predictors of OS with hazard ratios of 9.479 (95% CI, 3.220–27.847) and 6.148 (95% CI, 3.009–12.563), respectively. Time-dependent ROC curves from model 1 (clinical features only) and model 2 (clinical features and RRSs) are presented in Fig. 3. The iAUC of model 1 for OS prediction was 0.726 (95% CI, 0.678–0.814). When RRS_LASSO and RRS_elasticnet were added (model 2), the iAUC significantly increased to 0.780 (95% CI, 0.737–0.855) and 0.797 (95% CI, 0.752–0.872), respectively. On bootstrap testing, the increase in iAUC was statistically significant, with the differences in iAUC being 0.054 (95% CI, 0.013–0.104) and 0.040 (95% CI, 0.009–0.091), respectively. Smaller AIC values were noted in the models with RRS (392.2 and 381.0 for models with RRS_LASSO and RRS_elasticnet, respectively) compared with those without RRS (409.9), which indicate better prognostic models. In addition, smaller − 2LL values were also noted in the models with RRS (378.2 and 367.0 for models with RRS_LASSO and RRS_elasticnet, respectively) compared with those without RRS (397.9) with significant differences (p < 0.005 for all), which indicates better fitness of models.

Table 2 Multivariate Cox regression models with and without RRS for overall survival prediction
Full size table
Fig. 3
figure 3

Time-dependent receiver operating characteristic curves from models with and without radiomics risk scores (RRSs). Model 2 with both RRS_LASSO (a) and RRS_elasticnet (b) yielded significantly higher integrated area under the ROC curves (iAUC) of 0.758 (confidence interval [CI], 0.697–0.821) and 0.770 (CI, 0.706–0.833), respectively, compared with that of model 1 with only clinical features (0.701 [CI, 0.631–0.772])

Full size image

The performance of survival prediction models was tested through internal validation (Table 3). Models with RRS derived from both LASSO and elastic net demonstrated significantly superior performance, with higher iAUC values (LASSO: 0.758, 95% CI, 0.697–0.821; elastic net: 0.770, 95% CI, 0.706–0.833) than in the model without RRS (0.701, 95% CI, 0.631–0.772) for OS prediction. On bootstrap testing, the increase in iAUC was statistically significant, with the differences in iAUC being 0.058 (95% CI, 0.008–0.090) for LASSO and 0.069 (95% CI, 0.001–0.108) for elastic net.

Table 3 Internal validation of survival prediction models with and without RRS
Full size table

The performance of survival prediction models was also tested in the external validation cohort (Table 4). Models with RRS derived from both LASSO and elastic net yielded higher performance (LASSO: 0.754, 95% CI, 0.586–0.901; elastic net: 0.787, 95% CI, 0.619–0.931) than in the model without RRS (0.746, 95% CI, 0.543–0.921); however, the increase in iAUC was not statistically significant.

Table 4 External validation of survival prediction models with and without RRS
Full size table

The optimal cutoff values of RRS_LASSO and RRS_elasticnet were derived from the institutional cohort, which were − 11.53 and − 14.86, respectively. These RRSs stratified institutional cohort into low-risk and high-risk groups with significantly different OS (p < 0.001, both, Supplementary figure and Supplementary Table 3). There was significant difference in the extent of resection between the two groups and total resection of the tumor was more frequently performed in the low-risk group. Those optimal cutoff values of RRS also stratified external validation cohort into two groups with significantly different OS (p = 0.001 and 0.020, respectively, Fig. 4). The detailed clinical characteristics of the low-risk and high-risk groups in the external validation cohort are demonstrated in Supplementary Table 4.

Fig. 4
figure 4

Kaplan-Meier curves from the external validation cohort~ stratified on the basis of optimal cutoff values derived from the institutional cohort. Radiomics risk scores (RRSs) from both LASSO (a) and elastic net (b) stratified patients into two groups with significantly different overall survival

Full size image

Discussion

We identified a subset of radiomic features that was significantly associated with OS in patients with IDHwt lower-grade gliomas of histologic grades II and III. When the RRS which was derived from the selected radiomic features was integrated with the clinical features, a superior model performance was achieved in predicting prognosis. There was also a trend toward better performance when RRS was added to a model with clinical features in the external validation cohort. In addition, the optimal cutoff of the RRS derived from the institutional cohort divided the external validation cohort into two groups with significantly different survival outcomes, demonstrating the prognostic value of the RRS. Our study therefore suggests that the radiomic features may allow noninvasive risk stratification of patients with IDHwt lower-grade gliomas and can be used as a potential imaging biomarker.

Several previous studies have adapted radiomics to predict survival in patients with lower-grade gliomas. Zhou et al utilized various combinations of clinical features, a standardized feature set named Visually AccesSAble Rembrandt Images features, and radiomic features to predict tumor progression, revealing that the radiomic features alone achieved the highest performance compared with the other combinations tested in patients with grade II/III gliomas [18]. Another recent study identified a significant association between radiomic features and progression-free survival in patients with grade II/III gliomas, and the model integrating both clinicopathologic features and radiomic features demonstrated high accuracy for progression-free survival prediction [19]. These studies have pointed to the prognostic potential of radiomic features. Similar results were observed in our study, demonstrating that the radiomic features also have prognostic value in patients with the IDHwt subgroup of lower-grade gliomas. In this study, we identified that radiomics not only is an independent factor for survival but also has added prognostic value when integrated with clinical features. Even though the absolute increase of iAUC was not considerably big when the radiomics was added to a model with clinical features (iAUC increased to 0.780–0.797 from 0.726), still it was statistically significant. Furthermore, the prognostic value of radiomics was proved in the external validation set. In addition, after incorporating strong prognostic clinical features such as extent of resection or WHO grade, RRSs were revealed to have significant association with survival. Therefore, radiomics has the potential for accurate noninvasive risk stratification in patients with IDHwt lower-grade gliomas.

The majority of radiomic features that were selected by LASSO and elastic net for survival prediction were texture features. Other studies also reported that texture features among radiomic features showed the prognostic values in patients with gliomas [15, 17, 40, 41]. Texture features quantify the image pattern on the basis of the spatial relationship or co-occurrence of the pixel value [42], which provide information on intratumoral heterogeneity [12]. Specifically, the GLSZM feature named “zone percentage” derived from T1C, which was consistently selected from both LASSO and elastic net, presented the strongest association with overall survival. The feature is a measurement of coarseness of the texture [29], which was reported to have strong discriminative ability between lower- and high-grade gliomas according to a recent study [43]. Therefore, in our study, tumors with higher values of GLSZM zone percentage could follow more favorable outcomes based on more homogenous textures. In addition, intratumoral heterogeneity in gliomas has been reported to be associated with aggressive tumor behavior and drug resistance [44, 45]. Therefore, texture features may play a key role in predicting prognosis.

In our study, we extracted the radiomic features from T2WI and T1C images. FLAIR, T1WI, and advanced imaging such as DWI or perfusion images were not used because many patients from the external validation cohort lack these data. Extracting features from more sequences may increase the model performance for survival prediction but would inevitably decrease the generalizability of the model. Our study results showed that the radiomic features only obtained with conventional MR images (T2WI and T1C) had prognostic value, which was validated in the independent cohort, and therefore these findings could be applied to clinical situations in which advanced imaging cannot be routinely performed. In addition, a previous study had demonstrated that the T1C alone performed best for the accurate survival stratification of patients with glioblastoma, which was comparable to the performance of a combination of other conventional MR images (T1WI, T1C, T2WI, and FLAIR) [45]. Similarly, another study reported that texture features from T1C images most accurately predicted OS in patients with glioblastomas compared with those from T1WI, T2WI, or FLAIR images [46]. Therefore, our study results are in line with those previous studies which concluded that T1C conveys prognostic information. Furthermore, we extracted features not only from T1C but also from T2WI because the study cohort included lower-grade gliomas that may present with large proportion of nonenhancing tumor, and almost half of consistently selected texture features were derived from T2WI. A recent study suggested that progression-free survival in patients with grade II/III gliomas could be predicted accurately by radiomic features solely extracted from T2WI [19]. Thus, we believe that radiomics from conventional MRI, especially T2WI and T1C, may provide with useful information for predicting a patient’s prognosis. Fortunately, this is also more feasible in clinical practice with regard to its application and validation.

There was a significant difference in the extent of resection between the institutional cohort and the validation cohort. Gross total resection of the tumor was less frequently performed, but a biopsy was more frequently performed in our institution compared with the validation cohort. These differences may be due to the heterogeneous nature of the TCGA, which consisted of data from multiple institutions, which may have different standards for deciding on the extent of resection. In addition, the extent of resection was determined by comparing the pre- and postoperative T2WI and T1C in our institution by a neuroradiologist; however, it remains unclear how the extent of resection has been determined in the TCGA dataset.

There are several limitations in this study to be addressed. First, this is a retrospective study and the number of study population was small, especially in the validation cohort, because IDHwt tumors account for less than 30% of lower-grade gliomas. In addition, unfortunately, the postoperative treatment was not retrievable in all patients, while KPS was not available in 11 patients in the TCGA dataset. However, we managed to externally validate the prognostic value of RRS in patients with available clinical data and found that adding RRS to a model with clinical features increased the performance of model for survival prediction. The differences of performances between models with and without RRS were not statistically significant, which might be attributed to the small number of patients. Further studies using a larger cohort with more comprehensive clinical information may be needed to validate our results; however, our study results revealed a potential of RRS as an imaging biomarker for predicting survival. Second, patients were enrolled from relatively long time periods (from January 2007 to February 2018) in the institutional cohort; therefore, the specific MR parameters could be changed over time. In addition, in the external validation cohort, MRI protocols were heterogeneous as the TCGA dataset contains data from different institutions. Therefore, those different scan parameters and MRI protocols which lied in the institutional and validation cohort might affect the extracted radiomic features. We performed pre-processing such as resampling and intensity normalization to mitigate the effects which might occur due to those heterogeneities. We believe that our results showed the potential of radiomics for survival prediction even using MRI with heterogeneous protocols. Third, EGFR amplification or TERTp mutation has been reported to have prognostic implication in IDHwt lower-grade gliomas according to the cIMPACT-NOW recommendations. However, unfortunately, those genetic mutation results were available in only a small number of patients in the institutional cohort; therefore, we were not able to include genetic markers in the model. Future studies are required to validate the prognostic value of RRS in conjunction with EGFR amplification or TERTp mutation status.

Conclusion

In conclusion, radiomics enables noninvasive risk stratification and can improve survival prediction in patients with IDHwt lower-grade gliomas when integrated with clinical features.