Multiparametric Magnetic Resonance Imaging in the Assessment of Primary Brain Tumors Through Radiomic Features: A Metric for Guided Radiation Treatment Planning.

Purpose The definition of radiotherapy target volume is a critical step in treatment planning for all tumor sites. Conventional magnetic resonance imaging (MRI) pulse sequences are used for the definition of the gross target volume (GTV) and the contouring of glioblastoma multiforme (GBM) and meningioma. We propose the use of multiparametric MRI combined with radiomic features to improve the texture-based differentiation of tumor from edema for GTV definition and to differentiate vasogenic from tumor cell infiltration edema. Methods Twenty-five patients with brain tumor and peritumoral edema (PTE) were assessed. Of the enrolled patients, 17 (63 ± 10 years old, six female and 11 male patients) were diagnosed with GBM and eight (64 ± 14 years old, five female and three male patients) with meningioma. A 3 Tesla (3T) MRI scanner was used to scan patients using a 3D multi-echo Gradient Echo (GRE) sequence. After the acquisition process, two experienced neuroradiologists independently used an in-house semiautomatic algorithm to conduct a segmentation of two regions of interest (ROI; edema and tumor) in all patients using functional MRI sequences, apparent diffusion coefficient (ADC), and dynamic contrast-enhanced MRI (DCE-MRI), as well as anatomical MRI sequences-T1-weighted, T2-weighted and fluid-attenuated inversion recovery (FLAIR). Radiomic (computer-extracted texture) features were extracted from all ROIs through different approaches, including first-, second-, and higher-order statistics, both with and without normalization, leading to the calculation of around 300 different texture parameters for each ROI. Based on the extracted parameters, a least absolute shrinkage and selection operator (LASSO) analysis was used to isolate the parameters that best differentiated edema from tumors while irrelevant parameters were discarded. Results and conclusions The parameters chosen by LASSO were used to perform statistical analyses which allowed identification of the variables with the best discriminant ability in all scenarios. Receiver operating characteristic results showcase both the best single discriminator and the discriminant capacity of the model using all variables selected by LASSO. Excellent results were obtained for patients with GBM with all MRI sequences, with and without normalization; a T1-weighted sequence postcontrast (T1W+C) with normalization offered the best tumor classification (area under the curve, AUC > 0.97). For patients with meningioma, a good model of tumor classification was obtained through the T1-weighted sequence (T1W) without normalization (AUC > 0.71). However, there was no agreement between the results of both radiologists for some MRI sequences analyzed for patients with GBM and meningioma. In conclusion, a small subset of radiomic features showed an excellent ability to distinguish edema from tumor tissue through its most discriminating features.