Shariq Mohammed, Ph.D.
Assistant Professor
Department of Biostatistics
Boston University School of Public Health
Quantification of heterogeneity in the tumor microenvironment is extremely important as differentiating disease types visually is challenging. In this talk, we will present a statistical framework that quantifies spatial interactions in biomedical images to build prediction models for clinical phenotypes. We first build a spatial regression model to assess spatial interactions in regions of interest in the image. The heterogeneity in the spatial interactions in each image is then represented as a probability density function (serving as a signature quantifying spatial interactions). These density functions are analyzed using a Riemannian-geometric framework to include them as covariates in models that predict clinical outcomes of interest. We present our methodology with applications to radiology imaging in brain cancer to predict isocitrate dehydrogenase mutation, and to pathology imaging in pancreatic cancer to distinguish between different pancreatic disease subtypes.