501 Radiomic Phenotypes Distinguish ATRT from Medulloblastoma

Abstract

INTRODUCTION: Presurgical distinction of posterior fossa atypical teratoid/rhabdoid tumo (ATRTs) from medulloblastoma (MB) is challenging because of shared, low T1- and T2-weighted intensities, variable enhancement, and reduced diffusion. METHODS: The T2- and gadolinium-enhanced T1-weighted images of 48 posterior fossa ATRT and 96 match-paired MB were assembled from seven institutions. 1800 features were extracted from segmented tumor volumes. Feature reduction was performed by LASSO regression and cross-validation. Retained features were inputted to six training models, including support vector machine, logistic regression, k-nearest neighbor, random forest, eXtreme Gradient Boosting, and neural net. The optimal classifier maximized the area under the curve (AUC) on a holdout test set. Finally, two human experts reviewed the MRIs for categorical measurement of enhancement (0%; <50%, and =50% volume). RESULTS: The final feature set included one T1-MRI and five T2-MRI variables. Among classifier models, logistic regression produced the highest AUC of 0.858, with sensitivity, specificity, and F1 score of 0.800, 0.818, and 0.851, respectively. The top three predictive features were 90th Percentile Voxel Intensity, Inverse Difference Moment Normalized, and Kurtosis—all from T2-MRI. The last three features included T1-MRI and T2-MRI measurements for Elongation and Flatness. Separate qualitative human assessment identified greater volume enhancement for ATRT versus MB (p = 0.0135). CONCLUSION: Overall, T2-MRI-based features were more relevant for distinguishing ATRT from MB, with ATRT identified by brighter and more heterogeneous voxel intensities, consistent with ATRT’s myxoid and broad histologic composition. Additionally, ATRT were with more linear and planar morphology, consistent with ATRT flattening within the cerebellopontine angle and MB expansion within the fourth ventricle. Earlier distinction could aid family counseling and surgical planning.