Nimg-04. pre-radiation progression prediction in glioblastoma patients

Abstract Pre-Radiation Progression (PRP) has been identified as an increase ( > 25%) in the size of the residual brain tumor or interval development of a new tumor after surgical resection but prior to radiation therapy. Glioblastoma (GB) patients with PRP have worse survival outcome compared to patients without PRP (non-PRP).The standardized approach to identify PRP in GB patients is to manually compare early post-operative MRI to radiation planning MRI. In this study, we investigate the efficacy of traditional radiomics and sophisticated multi-resolution fractal texture features to differentiate PRP from non-PRP. Radiation-planning T1C magnetic resonance imaging (MRI) of 13 PRP and 13 non-PRP patients is analyzed to develop the classification method. For each patient, around 600 conventional radiomics and fractal texture features are obtained from tumor tissue and entire brain areas. We propose a two-step feature selection process followed by ensemble classification. In the first step of feature selection, we apply a beta-mixture based graphical model to identify correlations between the extracted features. The features are the nodes of the graph and the edges of the graph denote the correlations among the features and adjacency to PRP node. After the first step feature selection, the 21 selected features are used in the second step feature selection based on accuracy, similarity, and stability of each feature to yield the final 12 features. Lastly, we utilize the final 12 selected features for ensemble classification of PRP from non-PRP. With leave-one-out cross validation, the predictive model has an area under curve (AUC) of 0.79 and accuracy of 80% which indicates the discriminatory function of imaging features to differentiate PRP patients from non-PRP. The experimental results show the diagnostic ability of radiation planning imaging features to predict PRP from non-PRP.