Background: Frequent somatic mutations of BRAF and CTNNB1 were identified in both histological subtypes of craniopharyngioma (adamantinomatous and papillary) which shed light on target therapy to cure this oncogenic disease. The aim of this study was to investigate the noninvasive MRI-based radiomics diagnosis to detect BRAF and CTNNB1 mutations in craniopharyngioma patients.
Methods: Forty-four patients pathologically diagnosed as adamantinomatous craniopharyngioma (ACP) or papillary craniopharyngioma (PCP) were retrospectively studied. High-throughput features were extracted from manually segmented tumors in MR images of each case. The modifications-robustness in region of interests and Random Forest-based feature selection methods were adopted to select the most significant features. Random forest classifier with 10-fold cross-validation was applied to build our radiomics model.
Results: Four features were selected to make pathological diagnosis between ACP and PCP with area under the receiver operating characteristic curve (AUC) of 0.89, accurancy (ACC) of 0.86, sensitivity (SENS) of 0.89 and specificity (SPEC) of 0.85. The other two features were applied to estimate BRAF V600E mutation with AUC of 0.91, ACC of 0.93, SENS of 0.83 and SPEC of 0.97. Accurate predication of CTNNB1 mutation by three selected features was realized with AUC of 0.93, ACC of 0.86, SENS of 0.86 and SPEC of 0.86.
Conclusions: We developed a reliable MRI-based radiomics approach to perform pathological and molecular diagnosis in craniopharyngioma patients with considerably accurate prediction, which could offer potential guidance for clinical decision-making.
Keywords: Craniopharyngioma; Machine learning; Molecular diagnosis; Non-invasiveness; Radiomics approach.