LAUSR

PURPOSE H3 K27M mutation in gliomas has prognostic implications. Previous MRI studies have reported variable rates of tumoral enhancement, necrotic changes and peritumoral edema in H3 K27M-mutant gliomas, with no distinguishing imaging features compared to wild-type gliomas. Herein, we aimed to construct MRI a machine-learning (ML)- based radiomic model to predict H3 K27M mutation in midline gliomas. METHODS A total of 109 cases from three academic centers were included in this study. 50 cases had H3 K27M mutation and 59 were wild-type cases. Conventional MRI sequences (T1-W, T2-W, T2-FLAIR, post-contrast T1-W and ADC maps) were used for feature extraction. A total of 651 radiomics features per each sequence were extracted. Patients were randomly selected with a 7:3 ratio to create training (n = 76) and test (n= 33) data sets. An extreme gradient boosting algorithm, namely XGBoost, was used in ML-based model development. Performance of the model was assessed by area under the receiver operating characteristic curve (AUC). RESULTS Pediatric patients accounted for a larger proportion of the study cohort (60 pediatric (55%) vs 49 adult (45%) patients). XGBoost with additional feature selection had an AUC of 0.791 and 0.737 in the training and test data sets, respectively. The model achieved accuracy, precision (positive predictive value), recall (sensitivity) and F1(harmonic mean of precision and recall) measures of 72.7%, 76.5%, 72.2% and 74.3%, respectively, in the test set. CONCLUSIONS Our multi-institutional study suggests that ML-based radiomic analysis of multi-parametric MRI images can be a promising noninvasive technique to predict H3 K27M mutation status in midline gliomas.