Purpose: Early identification of glioma molecular phenotypes can lead to understanding of patient prognosis and guide treatment planning. MR texture analysis has shown promising results in discriminating molecular subtypes (1–3). We aimed to develop a multiparametric MRI texture analysis model to predict several important prognostic biomarkers in patients with brain glioma including: isocitrate dehydrogenase-1 (IDH-1) mutation, O6-methylguanine–DNA methyltransferase (MGMT) methylation, epidermal growth factor receptor (EGFR) overexpression, alpha thalassemia/mental retardation syndrome X linked (ATRX) mutation, tumor protein p53 (TP53) mutation, and phosphatase and tensin homolog (PTEN) mutation statuses.
Methods: In this retrospective IRB approved study, patients were included if 1) had diagnosis of gliomas with known IDH-1, EGFR, MGMT, ATRX, TP53, and PTEN statuses from surgical pathology and 2) had preoperative MRI including FLAIR and T1c+. Tumor segmentation was performed by using volume-of-interest (VOI) analysis to include FLAIR hyperintense regions. VOIs were transferred for radiomic texture analysis obtained from FLAIR and T1c+ images using Olea Sphere software (Olea Medical). For each patient a total of 92 texture features including first-order, co-occurrence matrix, and run length matrix were calculated. Correlation between the biomarker statuses and texture features were assessed using Least Absolute Shrinkage and Selection Operator (LASSO) regression algorithm to prevent overfitting and increased interpretation. Receiver-operating characteristic (ROC) curve analysis was performed to determine the optimal parameters and threshold for predicting glioma prognostic biomarker statuses. Results: A total of 111 patients (age: 56 ± 15, M/F 66/45) were included. There were 21 patients with grade II-III gliomas and 90 patients with grade IV. The breakdown of genetic biomarkers is summarized in Figure 1. Figure 2 summarizes the AUC/sensitivity/specificity/accuracy measurements to predict each genetic marker from a combination of texture features that remained as significant contributors after logistic regression. The highest predictive performance was obtained for IDH-1 (AUC: 0.96, accuracy: 91.5%) from a combination of 6 features (informal correlation and maximum probability from FLAIR and cluster prominence, skewness, difference variance, and low gray level zone emphasis from T1c+). The second best performance was obtained for ATRX (AUC: 0.95, accuracy: 88.1%) from a combination of 11 features (gray level nonuniformity, cluster prominence, maximum probability, small area emphasis, dependence variance, and informal measure of correlation from FLAIR; and low gray level zone emphasis, matrix strength, high gray level emphasis, skewness, and difference variance from T1c+). The lowest predictive performance was obtained for MGMT methylation status (AUC: 0.65, accuracy: 62.4%) from a combination of 3 features (range and cluster prominence from FLAIR and interquartile range from T1c+).
Conclusion: Our multiparametric MR texture analysis can predict genetic biomarkers of brain glioma with good to excellent diagnostic accuracy nearing 90% for IDH-1 and ARTX, while only modest for MGMT.

Acknowledgements

No acknowledgement found.