Objectives: To predict ATRX mutation status in patients with lower-grade gliomas using radiomic analysis.
Methods: Cancer Genome Atlas (TCGA) patients with lower-grade gliomas were randomly allocated into training (n = 63) and validation (n = 32) sets. An independent external-validation set (n = 91) was built based on the Chinese Genome Atlas (CGGA) database. After feature extraction, an ATRX-related signature was constructed. Subsequently, the radiomic signature was combined with a support vector machine to predict ATRX mutation status in training, validation and external-validation sets. Predictive performance was assessed by receiver operating characteristic curve analysis. Correlations between the selected features were also evaluated.
Results: Nine radiomic features were screened as an ATRX-associated radiomic signature of lower-grade gliomas based on the LASSO regression model. All nine radiomic features were texture-associated (e.g. sum average and variance). The predictive efficiencies measured by the area under the curve were 94.0 %, 92.5 % and 72.5 % in the training, validation and external-validation sets, respectively. The overall correlations between the nine radiomic features were low in both TCGA and CGGA databases.
Conclusions: Using radiomic analysis, we achieved efficient prediction of ATRX genotype in lower-grade gliomas, and our model was effective in two independent databases.
Key points: • ATRX in lower-grade gliomas could be predicted using radiomic analysis. • The LASSO regression algorithm and SVM performed well in radiomic analysis. • Nine radiomic features were screened as an ATRX-predictive radiomic signature. • The machine-learning model for ATRX-prediction was validated by an independent database.
Keywords: Biomarkers; Genetics; Glioma; Machine learning; Magnetic resonance imaging.