NIMG-51. CONVENTIONAL MRI RADIOMIC FEATURES IMPROVE PROGNOSTICATION AND ARE PREDICTIVE OF H3 K27M STATUS IN DIPG

Abstract

Abstract BACKGROUND Genomic profiling of DIPG suggests distinct and clinically relevant molecular subgroups based on the presence and isoform of histone H3 K27M mutation. We evaluated the impact of radiomic features on the classification and prognostication of 81 histologically confirmed and centrally reviewed DIPG. METHODS We utilized a combination of manual and automatic segmentation (DeepMedic) to define tumor volume and Pyradiomics for computation of radiomic features. Imaging feature stability was assessed by calculating concordance correlation coefficient (CCC) for each radiomic parameter from two separate pretreatment MRIs. Bootstrapped least absolute shrinkage and selection operator (LASSO) was used for feature selection. Classification and prognostication models, incorporating H3 status and clinical variables, were developed using random forest, Cox proportional hazards, and deep learning algorithms and assessed for goodness of fit using the c-index. RESULTS Eighty of 386 imaging features demonstrated stability (CCC, p< 0.001) between pretreatment scans. LASSO identified 26 prognostic imaging features and 38 and 57 imaging features predictive of the presence of H3 K27M mutation and H3 K27M isoforms, respectively. Using five-fold cross validation, the accuracy of distinguishing H3 K27M mutant and WT tumors was 85% and 77% for H3.3 K27M, H3.1 K27M, and WT tumors. C-index for prognostication was 0.77 for Cox, 0.55 for random forest, and 0.72 for deep learning. All models were more predictive than the Jansen survival prediction model. CONCLUSIONS Stable, predictive radiomic features may be a surrogate for H3 status and enhance current prognostication of DIPG. Model validation in cohorts of prospectively treated patients with DIPG is ongoing.