Purpose: PTEN mutation status is a pivotal biomarker for glioblastoma. This study aimed to establish a radiomic signature to predict PTEN mutation status in patients with glioblastoma, and to investigate the genetic background behind this radiomic signature.
Methods: In this study, a total of 862 radiomic features were extracted from each patient. The training (n = 69) and validation (n = 40) sets were retrospectively collected from the Cancer Genome Atlas and the Chinese Glioma Genome Atlas, respectively. The minimum redundancy maximum relevance (mRMR) algorithm was used to select the best predictive features of PTEN status. A machine learning model was then built with the selected features using a support vector machine classifier. The predictive performance of each selected feature and the complete model were evaluated via the area under the curve from receiver operating characteristic analysis in both the training and validation sets. The genetic background underlying the radiomic signature was determined using radiogenomic analysis.
Results: Six features were selected using the mRMR algorithm, including two features derived from contrast-enhanced images and four features derived from T2-weighted images. The predictive performance of the machine learning model for the training and validation sets were 0.925 and 0.787, respectively, which were better than the individual features. Radiogenomics analysis revealed that the PTEN-associated biological processes could be described using the radiomic signature.
Conclusion: These results show that radiomic features derived from preoperative MRI can predict PTEN mutation status in glioblastoma patients, thus providing a novel noninvasive imaging biomarker.
Keywords: Glioblastoma; Machine learning; Phosphatase and tensin homolog (PTEN); Radiogenomics.