Developing a Predictive Grading Model for Children with Gliomas Based on Diffusion Kurtosis Imaging Metrics: Accuracy and Clinical Correlations with Patient Survival

SIMPLE SUMMARY: The most frequent brain tumors in children are solid tumors. A significant fraction of pediatric brain tumors is represented by gliomas, which are heterogeneous. Diffusion kurtosis imaging metrics (MK, AK, RK, FA, and ADC) have shown promising results for glioma grading in adult patients; however, it is unclear whether this technique is accurate for diagnosing high grade pediatric gliomas and if it is correlated with patient survival. In our study, we performed a retrospective whole-tumor analysis on 59 children affected by gliomas and tested (1) if DKI metrics are accurate for grading pediatric gliomas and (2) if DKI metrics are correlated with patient overall survival and progression-free survival. ABSTRACT: Purpose: To develop a predictive grading model based on diffusion kurtosis imaging (DKI) metrics in children affected by gliomas, and to investigate the clinical impact of the predictive model by correlating with overall survival and progression-free survival. Materials and methods: 59 patients with a histological diagnosis of glioma were retrospectively studied (33 M, 26 F, median age 7.2 years). Patients were studied on a 3T scanner with a standardized MR protocol, including conventional and DKI sequences. Mean kurtosis (MK), axial kurtosis (AK), radial kurtosis (RK), fractional anisotropy (FA), and apparent diffusion coefficient (ADC) maps were obtained. Whole tumour volumes (VOIs) were segmented semi-automatically. Mean DKI values were calculated for each metric. The quantitative values from DKI-derived metrics were used to develop a predictive grading model to develop a probability prediction of a high-grade glioma (pHGG). Three models were tested: DTI-based, DKI-based, and combined (DTI and DKI). The grading accuracy of the resulting probabilities was tested with a receiver operating characteristics (ROC) analysis for each model. In order to account for dataset imbalances between pLGG and pHGG, we applied a random synthetic minority oversampling technique (SMOTE) analysis. Lastly, the most accurate model predictions were correlated with progression-free survival (PFS) and overall survival (OS) using the Kaplan–Meier method. Results: The cohort included 46 patients with pLGG and 13 patients with pHGG. The developed model predictions yielded an AUC of 0.859 (95%CI: 0.752–0.966) for the DTI model, of 0.939 (95%CI: 0.879–1) for the DKI model, and of 0.946 (95%CI: 0.890–1) for the combined model, including input from both DTI and DKI metrics, which resulted in the most accurate model. Sample estimation with the random SMOTE analysis yielded an AUC of 0.98 on the testing set. Model predictions from the combined model were significantly correlated with PFS (25.2 months for pHGG vs. 40.0 months for pLGG, p < 0.001) and OS (28.9 months for pHGG vs. 44.9 months for pLGG, p < 0.001). Conclusions: a DKI-based predictive model was highly accurate for pediatric glioma grading. The combined model, derived from both DTI and DKI metrics, proved that DKI-based model predictions of tumour grade were significantly correlated with progression-free survival and overall survival.