Peer-reviewed article

DIPG-48. MRI volumetric and machine learning based analyses predict survival outcome in pediatric diffuse midline glioma

· 1 Jun 2022

Abstract

Abstract INTRODUCTION: Diffuse midline glioma (DMG) is a fatal childhood CNS tumor. Magnetic resonance imaging (MRI) is the gold standard for DMG diagnosis and monitoring of response to therapy. Leveraging novel MRI analytical approaches, including volumetric and machine learning based analyses, may aid in the prediction of patient overall survival (OS) and help to identify high-risk cases. METHODS: T1- and T2-weighted MR images were retrospectively collected from children and young adults diagnosed with DMG (n=43). MRI features, including manually determined 3D tumor volume (T2), T1 contrast-enhancing tumor volume, T1 relative to T2 volume (T1/T2), tumor relative to whole brain volume, tumor average intensity, and tumor heterogeneity (i.e., intensity skewness and kurtosis), were evaluated at upfront diagnosis. MRI features were analyzed to identify significant predictors of OS outcome, which was defined as OS shorter, or longer, than one year from diagnosis. A support vector machine was used to predict OS outcomes using combinations of these features. RESULTS: The presence of T1 contrast-enhancing tumor at diagnosis (p=0.01), and a high T1/T2 ratio (>25%, p=0.009), predicted significantly shorter median OS. Moreover, feature selection identified T2 mean intensity (p<0.001), T2 image intensity skew (p=0.006), T1/T2 ratio (p=0.02), and T1 volume relative to whole brain (p=0.03) as significant predictors of OS outcome (short versus long). Combining T2 mean intensity, T2 image skew, T1 segment kurtosis and patient gender resulted in OS outcome prediction accuracy of 83.3% (sensitivity=85%, specificity=81.8%, n=42 cases).CONCLUSION: We have identified MRI volume and imaging features that significantly predict OS outcome in children diagnosed with DMG. Our findings provide a framework for incorporating MRI volumetric and machine learning analyses into the clinical setting, allowing for the customization of treatment based on tumor risk characteristics.