Brain cancer has complex underlying mechanisms that promote tumors through multiple signaling networks hosted in a number of molecularly discrete diseases. A quantitative understanding of the relationships and interactions among histopathology, genetics and genomic expression can lead to a better interpretation of the effects of these signaling networks. The morphology of tumor nuclei, for instance, plays a pivotal role in the classification of diffuse gliomas. Necrosis is linked with tumor progression through well-characterized molecular mechanisms. Investigation of morphological properties at the tissue, cellular and sub-cellular scales and correlation of these properties with genomic profiles and clinical data can significantly improve prognostic accuracy and characterization of treatment response.
Morphology based integrative analyses pose significant challenges on computational accommodation, data curation and algorithmic development. Moderate numbers of digitized microscopy specimens quickly lead to formidable information synthesis and management problems. Our group carries out research and development in three areas: optimized analysis algorithms and analytic pipelines, high performance computation middleware and information models and data management systems, in order to enable investigators to carry out large-scale comparative analyses of brain tumor histological features from whole slide images synergized with patterns of protein and gene expression data.
Click on each header below to see our publications in the area.
Larger scale histological structures, such as necrosis and hypoxia-driven blood vessel growth (angiogenesis or microvascular hyperplasia), are linked with tumor development through increasingly well-characterized molecular mechanisms. Investigations of necrosis in TCGA frozen sections have revealed links between necrosis and recognized transcriptional classes of glioblastoma. Investigations of the correlations between transcription and necrosis in TCGA have illustrated that hypoxia induces key transcriptional factors that regulate these transcriptional classes, suggesting that the observable microenvironment is a critical factor in establishing tumor subclass. (Read more... )
Morphologic variations of disease observed at microscopic resolution are often linked to underlying molecular events and patient outcome, suggesting that quantitative morphometric analysis may provide further insight into disease mechanisms. There is currently intense interest in developing subclassifications of disease, with the hope that finer distinctions will lead to improved personalized therapies. However most, if not all, of these efforts have focused on genomics. Using image analysis techniques, we develop signatures to represent patient-specific tumor morphology through the analysis of hundreds of millions of cells in digitized whole slide images. Clustering of these signatures aggregates tumors into groups with cohesive morphological characteristics. We demonstrate this methodology with an analysis of glioblastoma, using data from The Cancer Genome Atlas to identify a prognostically-significant morphology-driven subclassification, in which clusters are correlated with transcriptional, genetic and epigenetic events. (Read more... )
We have carried out studies investigating the utility of imaging-based biomarkers in predicting overall survival, pathology features and molecular signatures. In collaboration with an NCI led consortium, we have analyzed pre-surgical MRI images of GBM patients whose tissues had been molecularly analyzed in the Cancer Genome Atlas study. We jointly carried out a qualitative assessment of imaging based tumor characteristics such as necrosis and contrast enhancement and found that these characteristics significantly predict overall survival. We have also found that molecularly defined tumor subtypes show observable differences in macroscopic appearance. (Read More... )
Software to support the extraction and interpretation of information from large imaging datasets has to deal with expensive data processing requirements, thousands of multi-gigabyte images and trillions of microscopic objects and their features. Detailed characterization of morphology in a large image dataset requires coordinated use of many interrelated analysis pipelines and comparison of analysis results from multiple analysis pipelines and analysis runs. A single analysis run involves pipelines of cascaded methods including: 1) data transformation tasks such as thresholding, tessellation, color and illumination normalization, 2) segmentation of structures such as cells and nuclei, 3) characterization of shape and texture features of segmented structures, and 4) machine-learning methods that integrate information from features to accomplish classification tasks. (Read More... )