Tumor volume features predict survival outcomes for patients diagnosed with diffuse intrinsic pontine glioma

BACKGROUND Diffuse intrinsic pontine glioma (DIPG) is a fatal childhood central nervous system tumor. Diagnosis and monitoring of tumor response to therapy is based on magnetic resonance imaging (MRI). MRI-based analyses of tumor volume and appearance may aid in the prediction of patient overall survival (OS). METHODS Contrast-enhanced T1- and FLAIR/T2-weighted MR images were retrospectively collected from children with classical DIPG diagnosed by imaging (n = 43 patients). MRI features were evaluated at diagnosis (n = 43 patients) and post-radiation (n = 40 patients) to determine OS outcome predictors. Features included 3D tumor volume (T$_{wv}$), contrast-enhancing tumor core volume (T$_{c}$), T$_{c}$ relative to T$_{wv}$ (T$_{C}$/T$_{wv}$), and T$_{wv}$ relative to whole brain volume. Support vector machine (SVM) learning was used to identify feature combinations that predicted OS outcome (defined as OS shorter or longer than 12 months from diagnosis). RESULTS Features associated with poor OS outcome included the presence of contrast-enhancing tumor at diagnosis, >15% T$_{c}$/T$_{wv}$ post-radiation therapy (RT), and >20% ∆Tc/T$_{wv}$ post-RT. Consistently, SVM learning identified T$_{c}$/T$_{wv}$ at diagnosis (prediction accuracy of 74%) and ∆T$_{c}$/T$_{wv}$ at <2 months post-RT (accuracy = 75%) as primary features of poor survival. CONCLUSIONS This study demonstrates that tumor imaging features at diagnosis and within 4 months of RT can predict differential OS outcomes in DIPG. These findings provide a framework for incorporating tumor volume-based predictive analyses into the clinical setting, with the potential for treatment customization based on tumor risk characteristics and future applications of machine-learning-based analysis.