Angela Yuan
Pronouns: She/her
Research Mentor(s): Arvind Rao
Research Mentor School/College/Department: Computational Medicine and Bioniformatics / Medicine
Program: UROPF
Session: Session 7 (4:40pm – 5:30pm)
Authors: Angela Yuan, Yiou Wang, Avery Maddox, Arvind Rao
Abstract
Standard hematoxylin and eosin (H&E) stained tissue images are a key data point used in the clinical management of many diseases, especially cancer. As the clinically relevant features within cancerous tissue become increasingly complex, new methods are needed to efficiently detect these features and leverage them towards improved patient outcomes. One such feature is the spatially patterned interaction of different cell-types in and around cancerous tissue. Our research aims to use computational techniques to directly detect such patterns from H&E images and derive indicators of key clinical outcomes. Here we develop and apply our method on a novel glioblastoma (GBM) dataset. Our method first requires tumor/non-tumor segmentation of GBM H&E images. This segmentation is essential for comparing how cell-types are patterned in relation to the tumor region. This segmentation will also serve as an initial key test of our method to differentiate between cell-types in a relatively simple 2-class scenario. The ground-truth label is derived from spatial RNA profiles obtained through Visium 10x. We seek to later train a model to predict the abundance of different cell-types across the H&E image. Then, we will apply spatial analysis to see how different cell-types associate with one another and derive clinically meaningful insights. Our methods will be implemented in Python and its relevant libraries. We will present figures showing model performance, cell-type abundance over the original H&E tissue image, and the relation between spatial cell-type association measurements and clinical outcomes. We are still finalizing our results since we spent much time becoming familiar with deep learning. We hope to find a correlation between our spatial cell profiles and clinical outcomes, such as tumor grade, and to make comparisons between different cell-type plots for deeper insights.
