P13.14.A PREDICTION OF OVERALL SURVIVAL IN HIGH-GRADE GLIOMA USING MULTIPARAMETRIC MRI-BASED RADIOMICS. APPLICATION TO A MULTICENTER COHORT

Abstract

Abstract BACKGROUND High-grade gliomas are the most common type of primary brain tumor in adult patients. Despite a multimodal treatment, the prognosis remains poor. Accurate stratification of survival is crucial for choosing the best course of therapy. The objective of this study was to use radiomics analysis based on multiparametric MRI combined with clinical data of high-grade gliomas patients with incomplete resection to identify a predictive signature for the stratification of patients according to their overall survival. MATERIAL AND METHODS In this multicentric retrospective study, we included a sub-cohort of 237 patients treated for high-grade gliomas, from 2009 to 2021, at three French medical centers. This data collection was carried out as part of a national funding to promote the integration of AI tools into clinical routine in oncology (AI.DReAM project). Clinical and pathological data, together with the MR images before treatment and at the time of relapse were collected.Patients with gross tumor resection were excluded to homogenize the automatic contouring. Three tumor subregions were generated by a deep learning framework based in the BraTS challenge on the MR images at the time of radiotherapy: label 1, corresponding to the necrosis part of the tumor; label 2, the hypersignal in the T2/FLAIR sequence; and label 3, corresponding to the enhancing part after the injection of gadolinium. The cohort was split into a training data set of 189 patients with cross-validation to ensure robustness and evaluated using a test set of 48 patients. The survival stratification of the two classes was based on the median in month of overall survival (OS) in our cohort with short survival defined as OS inferior to 17 months. The open-source software Pyradiomics and the ensemble classifiers XGBoost, in conjunction with grid search were used for calculating radiomic parameters. RESULTS In total, 976 patients were collected. For this analysis, 237 patients with high-grade gliomas were included, 82 % being IDH-wildtype glioblastoma. One-hundred and nineteen patients underwent subtotal tumor resection. A total of 936 radiomics were extracted from the three tumor subregions and 14 clinical data features were added to the algorithm. The areas under the ROC curve (ROC-AUC) for survival prediction based on radiomic and clinical parameters alone were 0.61 and 0.60 respectively on the test set. The combination of radiomic features with clinical features showed superior performance with ROC-AUC of 0.67 in the 5-fold cross-validation and 0.64 in the test set. CONCLUSION In this preliminary study, we achieved a reproducible estimation of patient survival using radiological and clinical features. In future analyses, we will seek to better account for the biological heterogeneity of high-grade glioma by increasing the number of labels and having them corrected by experts in the field.