Synopsis

MGMT promoter methylation is associated with longer survival and better treatment response of GBM patients. Intratumor heterogeneity is partly responsible for inaccurate detection of MGMT status. Therefore, assessing the effect of different heterogenous subregion of GBM on MGMT status would be critical. In this study, a radiomics approach integrated optimal features of heterogenous subregions in multimodal MRI and machine learning model was proposed for effectively predicting MGMT methylation, and meanwhile assessing the prediction efficiency of subregions or subregion combinations. The proposed approach achieved a promising MGMT methylation detection performance and indicated that rNEC may play a role in this issue.

Introduction

Glioblastomas (GBMs) are the most common primary brain tumors with a dismal outcome^{(1, 2)} and O6-methylguanine methyltransferase (MGMT) promoter methylation is associated with longer survival and better response to temozolomide⁽³⁾. However, MGMT methylation detection would be influenced by intratumor heterogeneity of GBM. Noninvasively assessing the influence of heterogeneity on MGMT methylation detection is critical for GBM patients.

Currently, the magnetic resonance imaging (MRI) has served as a noninvasive tool for intratumor heterogeneity analysis of GBMs because it occurred at the molecular level and can be macroscopically reflected by different phenotypes in MRI⁽⁵⁾. The phenotypes were commonly identified as the region of contrast-enhancing tumor (rCET), necrosis (rNec) and edema/non contrast-enhancing (rE/nCET). However, few studies focused on the influence of different subregions on MGMT status assessment. Therefore, it is necessary to establish a novel approach to accurately predict MGMT methylation status and evaluate the influence of different subregions on this issue.

In this study, we aimed to use high throughput radiomics features from each subregion in multimodal MRI and machine learning methods to stratify MGMT methylated and unmethylated groups, and meanwhile, to evaluate the influence of each subregion or subregion combination.

Methods

Patient selection

As depicted in Fig. 1, a total of 62 GBM patients retrospectively enrolled. Among them, 32 were MGMT methylated, and 30 were unmethylated.

Image data preprocessing and radiomics feature extraction

Each patient’s T1-weighted images (T1WI) before and after contrast-enhanced (T1C), T2-weighted (T2WI) and fluid-attenuation inversion recovery (FLAIR) images were collected. Modality registration was performed using the rigid transformation and mutual information. All the images were normalized using Collewet algorithm⁽⁶⁾. To extract high throughput features, both original and corresponding filtered MRI images after wavelet transform were obtained.

Then, four heterogenous subregions were manually drawn: (1) the region of entire abnormality (rEA) on FLAIR image, (2) rCET and (3) rNec on T1C. In addition, (4) rE/nCET acquired by rEA subtracting the union of rCET and rNec, as Fig. 2 shown. A total of 4000 radiomics features were finally extracted (4 modalities×5 images×4 subregions×50 texture and histogram features), as previous studies applied ^(7-10).

Feature Selection and Classification

For each classification task, the SVM-RFE method⁽¹¹⁾ was used to find the optimal feature subsets. Based on these, MGMT methylation detection models were establishment using SVM classifiers⁽¹²⁾. The accuracy and area under the curve (AUC) were calculated to evaluate the performance.

Statistical analysis

The Mann-Whitney U test, Chi-square test was applied on clinical characteristics to verify whether significant inter-group differences exist.

Results

Among 62 enrolled patients, 32 and 30 belonged to MGMT methylated and unmethylated groups respectively. Besides the overall survival (P <0.05), there were no significant intragroup differences of clinical characteristics.

Fig. 3 shows the efficiency of MGMT status assessment using optimal radiomics features from single subregion or subregion combination. The efficiency of rNEC was outstanding for MGMT methylated detection, with accuracy=91.94% and AUC=0.9531. However, there was an extremely small difference between the four-subregion (91.94% and 0.950) and the three-subregion combination (91.94% and 0.9521). No performance improvement was found when adding other single subregions besides of rNEC.

Fig. 4 shows the efficiency of MGMT status assessment using optimal radiomics features from different MRI modalities. The best accuracy and AUC were achieved when using T1 images (90.32%, and 0.9323).

Discussion

In this study, the proposed approach integrated optimal features and machine learning methods showed promising performance on MGMT methylation detection of GBM patients. Meanwhile, different subregions in multimodal MRI had different efficiency for MGMT status assessment.

Theoretically, each heterogenous subregion may contain specific microenvironment information. However, they did not complement each other for MGMT status assessment in this study. There was no improvement when adding other subregions besides of rNEC. The detection model using only rNEC reached the best performance. Therefore, rNEC was verified to be a key component of GBM for MGMT status assessment in this study.

For intratumor heterogeneity of GBMs, it was critical to determine which regions/volumes were extracted for further analysis. In most of previous GBM-related MRI studies, researchers used to focus on the rCET to investigate its potential to grading or survival prediction. Others indicated that the edema or non-enhancing region may better reflect the characteristics of GBMs. When delineating regions/volumes for further analysis, radiologists preferred to draw the active tumor region and avoid the necrosis and cystic sometimes. It now appears that rNEC may play a role in MGMT methylation assessment.

Conclusion

Acknowledgements

References