Individual Prediction of Brain Tumor Histological Grading Using Radiomics on Structural MRI

The accurate diagnosis of brain tumors is of primary importance for optimal therapy planning. In clinical practice, this is determined on a biopsy, exposing the patient to the risk of complications. Moreover, sampling bias and performer variability may influence the result. Several studies have investigated the histological grading of brain tumors in a non-invasive way by extracting features from medical images. A multicenter study where both tumor grade and cell type are simultaneously predicted is however lacking. In this study we collected structural MRI-scans from 294 patients with glioma acquired in different centers (the local hospital and two online databases). The goal was to predict tumor grade and cell type of individual patients using a radiomics study with Random Forests. In a multiclass design, we obtain a global accuracy of 59.9% to predict tumor grade and 53.4% to predict cell type. Converting the problem to binary classification, we obtain an accuracy of 98.3% to distinguish between meningioma and glioma, and 84.5% to distinguish between low-grade glioma and glioblastoma. A high degree of diagnosis variability and overlap between different low-grade classes might cause the reduced prediction accuracy. Our results however show that radiomics on structural MRI is a suitable approach for non-invasively assessing brain tumor diagnosis and might be used for individual treatment planning.