Prediction of IDH-mutation and 1p19q-codeletion status in adult-type diffuse gliomas using three-class radiomics models.

PURPOSE: To predict the genetic subtypes of adult-type diffuse gliomas with three-class MRI radiomics. MATERIAL AND METHODS: Four hundred and eighty patients with pathologically proved diffuse gliomas from our institution were randomly divided into training set (n = 336) and internal validation set (n = 144) according to a ratio of 7 to 3, and 105 patients from the cancer imaging archive (TCIA) were utilized as an external validation set. All patients were classified into IDH wild-type (IDHwt), IDH-mutant with 1p19q-noncodeleted (IDHmut-intact) and IDH-mutant with 1p19q-codeleted (IDHmut-codel) subtypes, and all the patients underwent conventional MR imaging and diffusion weighted imaging (DWI) scans. Two regions of interest (ROI) segmentation schemes (tumor ROI and tumor with peritumoral edema ROI) and MRI sequences with and without apparent diffusion coefficient (ADC) maps were tested with 6 machine learning classifiers to filter the optimal models. RESULTS: Support vector machine (SVM) classifier combined with conventional MR sequences based on tumor ROI was proved to be the best diagnostic model, with the area under the curve (AUC) of 0.963, 0.964, and 0.929, and accuracy of 0.902, 0.929, and 0.890 for IDHwt, IDHmut-intact, and IDHmut-codel prediction, respectively, and overall accuracy (micro average accuracy) of 0.860. The model was validated in both internal and external validation sets. CONCLUSIONS: Three-class MRI radiomics can be used preoperatively in predicting the molecular subtype of adult diffuse gliomas with satisfactory performance and show potential value for the diagnosis and risk stratification in glioma patients. This retrospective study received ethics approval with a waiver of informed consent.