Once the diagnosis of a diffuse glioma has been made, usually from tissue sampling, the treatment plan, including dose, type, and fractionation of radiotherapy as well as the sequence of chemotherapy, is mainly determined by the molecular characteristics of the tumor based on the WHO classification [
3]. Many groups have shown that these may also be derived non-invasively from imaging by application of machine learning [
22‐
25,
33‐
35,
38,
39,
42,
43]. Zhang and co-workers [
43] used conventional MRI as well as DWI and extracted a total of 2970 features from 120 patients with WHO grade III and IV gliomas. Finally, a subset of 386 features was used to build a model based on a random forest classification, resulting in an accuracy of 89% for IDH prediction in the validation dataset. In WHO grade II and III gliomas, Zhou and colleagues [
39] showed that a logistic regression model with only three features obtained from conventional MRI of 165 patients predicted the IDH genotype as well as the 1p/19q co-deletion status with AUC values of 0.86 and 0.96, respectively. In a subsequent study [
38], Zhou and colleagues extracted radiomics features from conventional MRI of a large multicentric dataset of more than 500 patients. The model was built using a random forest classifier and validated using another set of MR images from The Cancer Imaging Archive (TCIA). The IDH genotype could be predicted with an AUC of 0.92 in the test cohort. Lu and colleagues [
34] used conventional MRI data from 214 glioma patients from TCIA. Additionally, 70 patients with MRI scans from different institutions were used for model testing. A multi-level machine learning model based on support vector machine classifiers was used and allowed classification of IDH and 1p/19q status with accuracies of 88% and 96%, respectively. Lohmann and colleagues [
33] used FET PET radiomics for preoperative prediction of IDH genotype in a cohort of 84 glioma patients and identified a simple two-parameter logistic regression model that achieved a diagnostic accuracy of 80% after 10-fold cross validation. A subgroup analysis of 28 patients measured on a high-resolution BrainPET scanner showed the highest accuracy of 86% after 10-fold cross validation.
Again, deep learning-based radiomics approaches have also been investigated for prediction of molecular characteristics in gliomas. Chang and co-workers [
22] used a residual CNN for IDH genotype prediction on conventional MRI from a large cohort of 496 glioma patients from three different institutions and achieved an accuracy of 86% in an independent test set. The accuracy could be further increased to 89% by the incorporation of age at diagnosis. Eichinger and colleagues [
24] used a set of local binary pattern features extracted from T2-weighted MRI and DTI for training of a CNN with a single hidden layer. The prediction of IDH genotype yielded a diagnostic accuracy of 95% in the test dataset. These results suggest that radiomics has the potential to substitute neuropathological assessment of IDH mutation and 1p/19q LOH status in glioma patients in whom tissue sampling is not feasible.