Computer-aided pediatric brain tumor detection, prediction and statistical validation using structural MRI and gene expression data

Brain tumor is one of the leading causes of solid tumor cancer-related deaths in children under the age of 20. Automatic detection and prediction of pediatric brain tumors is a non-trivial task because of the heterogeneity in tumors' structural phenotype and genotype information. The urgency of early detection of such life-threatening disease among children can not be overemphasized. However, the current state of specific detection, prediction and classification of most prevalent types of pediatric brain tumors such as posterior fossa (PF) and glioblastoma multiform (GBM) is rather inadequate. This inadequacy and the resulting opportunity to improving pediatric patient management provide the motivation to pursue this dissertation research in investigating and improving computer-aided pediatric brain tumor diagnosis and prediction for children. In this dissertation work, computer-aided pediatric brain tumor detection is investigated based on: (a) structural image feature values from brain magnetic resonance images (MRI), and (b) molecular information from microarray gene expression measures.;One of the major novel contributions of this dissertation is the modeling of tumor texture in brain MRIs as multi-fractional Brownian motion (mBm). This multi-fractal feature is exploited to detect PF tumor in brain MRI. A novel modification of AdaBoost classifier is further proposed for effective non-patient-specific tumor classification. A novel statistical modeling of gene expression based tumor prototype is proposed for automatic tumor prediction. The novelty of the proposed modeling stems from the foot that the tumor prediction scheme explores correlations among genes and probable new subgroupings within known tumor types. A novel gene visualization plot is also proposed for qualitative analyses of marker gene selection process. A novel gene selection method based on samples from the posterior distributions of class-specific gene expression measures is also proposed. In that work, a hierarchical Bayesian framework for a random effect ANOVA model is constructed that allows us to obtain the posterior distributions of the class-specific gene expressions. Further, a novel class prediction scheme is formalized based on the samples from new posterior distributions of group specific gene expressions. Finally, a novel tumor prediction model is proposed wherein the information from structural brain MRI features and Microarray gene expression measures are combined. In this work, desirable features from both genotype and phenotype modalities are combined into a single seamless framework for effective GBM tumor prediction.;The primary contribution of this dissertation is that a few computer-aided statistical and computational models are proposed for improved specificity and sensitivity of pediatric brain tumor detection, prediction and classification using both MRI structural and microarray gene expression data. Extensive comparisons of the proposed techniques are shown with existing methods in literature and demonstrate the efficacy of the techniques.