Improving MGMT methylation status prediction of glioblastoma through optimizing radiomics features using genetic algorithm-based machine learning approach

Abstract

O6-Methylguanine-DNA-methyltransferase (MGMT) promoter methylation was shown in many studies to be an important predictive biomarker for temozolomide (TMZ) resistance and poor progression-free survival in glioblastoma multiforme (GBM) patients. However, identifying the MGMT methylation status using molecular techniques remains challenging due to technical limitations, such as the inability to obtain tumor specimens, high prices for detection, and the high complexity of intralesional heterogeneity. To overcome these difficulties, we aimed to test the feasibility of using a novel radiomics-based machine learning (ML) model to preoperatively and noninvasively predict the MGMT methylation status. In this study, radiomics features extracted from multimodal images of GBM patients with annotated MGMT methylation status were downloaded from The Cancer Imaging Archive (TCIA) public database for retrospective analysis. The radiomics features extracted from multimodal images from magnetic resonance imaging (MRI) had undergone a two-stage feature selection method, including an eXtreme Gradient Boosting (XGBoost) feature selection model followed by a genetic algorithm (GA)-based wrapper model for extracting the most meaningful radiomics features for predictive purposes. The cross-validation results suggested that the GA-based wrapper model achieved the high performance with a sensitivity of 0.894, specificity of 0.966, and accuracy of 0.925 for predicting the MGMT methylation status in GBM. Application of the extracted GBM radiomics features on a low-grade glioma (LGG) dataset also achieved a sensitivity 0.780, specificity 0.620, and accuracy 0.750, indicating the potential of the selected radiomics features to be applied more widely on both low- and high-grade gliomas. The performance indicated that our model may potentially confer significant improvements in prognosis and treatment responses in GBM patients.

Figure 1

The overall feature selection steps. The left part demonstrates the pre-processing and segmentation steps while the right part list the two-stage feature selection procedure. The extracted feature set is then evaluated for its efficacy.

Figure 2

The Genetic Algorithm workflow. The steps are: (1) Generation of the initial population of solutions; (2) Evaluation of fitness values of each solution within the population; (3) The “mating” process of the solution, in which the probability of a solution to be selected is proportional to the estimated fitness value; (4) The random designation of crossover points on each vector of solution during the “mating” process. SC and DC stand for Single- and Double-Crossover, respectively; (5) The introduction of random mutations on the crossover-ed solution vectors; (6) The replacement of the entire population by daughter solutions.

Figure 3

Performance evaluations and comparisons of different GA-incorporated models in predicting MGMT methylation statuses. Y-axis represents accuracy. Statistical significances evaluated by Kolmogorov–Smirnov test are represented by stars, in which three stars (***) indicate p < 0.001.

Figure 4

Receiver operating characteristic (ROC) curves of different feature sets as evaluated by the random forest (RF) algorithm. The feature sets are: (A) all 704 radiomics features; (B) 38 features selected by XGBoost; (C) the feature set selected by F-scores; and (D) the feature set selected by the genetic algorithm (GA)-RF algorithm.

Figure 5

Common radiomics features selected by different methods. Solid circles represent the presence of certain features in each feature set.