Effect of segmentation dimension on radiomics analysis for MGMT promoter methylation status in gliomas

Abstract

Introduction and objective: We investigated the impact of 2D (2D_seg) and 3D (3D_seg) segmentation on the accuracy of prediction models in the radiomics analysis to determine the presence or absence of methylation in the O6-methylguanine DNA methyltransferase (MGMT) gene promoter region of gliomas. Materials and methods: Magnetic resonance imaging images of gliomas were obtained from the Cancer Imaging Archive for 50 methylated and 50 unmethylated cases respectively. For each case, 2D_seg and 3D_seg were performed, and 788 radiomics features, including wavelet transform, were obtained. Ten features were selected by LASSO regression. The coefficients of determination (R2) and root mean squared error (RMSE) were calculated by multiple regression analysis. Discriminant boundaries to discriminate methylation were created by linear discriminant analysis, and the sensitivity and specificity of each method were calculated. The discriminant accuracy of both methods was evaluated by receiver operating characteristics (ROC) analysis. Results: The R2 value and RMSE were 0.72/0.28 and 0.73/0.33 for 2D_seg and 3D_seg, respectively. Similarly, sensitivity and specificity were 82.5/67.5% and 85/62.5%, respectively. The area under the curve determined by ROC analysis was 0.80 and 0.79, respectively, i.e. slightly larger for 2D_seg. The p-value by the DeLong method was 0.73. Conclusions: In the radiomics analysis using 2D_seg and 3D_seg, no difference in discriminant accuracy was observed between them. Therefore, 2D segmentation should be chosen because it is easier to segment.