IDH1 mutation prediction using MR-based radiomics in glioblastoma: comparison between manual and fully automated deep learning-based approach of tumor segmentation. Issue 128 (July 2020)

• Record Type: Journal Article
• Authors: Choi, Yangsean; Nam, Yoonho; Lee, Youn Soo; Kim, Jiwoong; Ahn, Kook-Jin; Jang, Jinhee; Shin, Na-Young; Kim, Bum-Soo; Jeon, Sin-Soo
• Is Part Of: European journal of radiology. Issue 128(2020)
• Publication Date: 2020-07
• Subjects: 3D three-dimensional -- ROC receiver operating characteristics -- AUC area under the curve -- CNN convolutional neural network -- DSC dice similarity coefficient -- IDH1 isocitrate dehydrogenase-1 -- TCIA The Cancer Imaging Archive -- TCGA The Cancer Genome Atlas -- T2WI T2-weighted image; glioblastoma -- isocitrate dehydrogenase -- magnetic resonance imaging -- machine learning -- sensitivity and specificity; Medical radiology -- Periodicals; Radiology -- Periodicals; Radiologie médicale -- Périodiques; Medical radiology; Periodicals; 616.075705
• DOI: 10.1016/j.ejrad.2020.109031
• Languages: English
• ISSNs: 0720-048X

Abstract

Highlights: For IDH1 status prediction, 9 radiomic features were selected via Boruta algorithm. Radiomics of glioblastoma predicted IDH1 status well in manual and V-Net methods. V-Net demonstrated robust segmentation of entire glioblastoma on T2-weighted MRI. Abstract: Purpose: This study aimed to determine whether MR-based radiomics of glioblastoma can predict the isocitrate dehydrogenase-1 (IDH1) mutation status and compare predictive performances between manual and fully automatic deep-learning segmentations. Method: Forty-five glioblastoma patients with pretreatment T2-weighted MRIs were retrospectively evaluated. Manual segmentations of glioblastoma and peri-tumoral edema were trained via a deep neural network (V-Net). An independent external cohort of 137 glioblastoma patients from the Cancer Imaging Archive was also included (test set 1, n = 46; test set 2, n = 91). Test set 1—without known IDH1 status—was used to calculate dice similarity coefficients (DSC) between the two segmentation methods (manual & V-Net). From test set 2, all-relevant radiomic features were selected via a random forest-based wrapper algorithm for IDH1 prediction. Receiver operating characteristics (ROC) curves with areas under the curve (AUC) were plotted as performance metrics for both methods. Results: Among 136 patients (45 and 91 patients from our institution and test set 2, respectively), 17 patients (11.2 %) had IDH1 mutations. The mean DSC of test set 1 was 0.78 ± 0.14 (range, 0.34−0.94). A subset of 9 all-relevant features (8.4 %, 9/107) was selected. V-Net segmentation of the test set 2 yielded similar performance in predicting IDH1 mutation as compared to manual segmentation (V-Net AUC = 0.86 vs. manual AUC = 0.90). The optimal cut-point threshold of AUC yielded 86.8 % accuracy for manual segmentation and 75.8 % for V-Net segmentation. Conclusions: V-Net showed robust segmentation capability of glioblastoma on T2-weighted MRI. All-relevant radiomics features from both segmentation methods yielded a similar performance in IDH1 prediction.