Bio

Shahira Abousamra is a Postdoctoral scholar in the department of Biomedical Data Science at Stanford University. She earned her PhD in Computer Science from Stony Brook University in 2024.

In her research, she integrates mathematical modeling with computer vision to create more robust solutions, particularly in the context of advancing cancer research and enhancing our understanding of the tumor microenvironment. She leverages topological data analysis (TDA) and spatial statistics to provide spatial semantic grounding to complement machine learning models.

She publishes in top computer vision, artificial intelligence, and medical image analysis conferences including CVPR, ECCV, ICCV, AAAI, and MICCAI.

Topological Data Analysis to Enhance Deep Learning Models of the Tumor Microenvironment

My research is at the intersection of computer vision, topological data analysis (TDA), and biomedical data analysis. One key aspect of biomedical data is their inherent anatomical and cellular structures. For example, within the tumor microenvironment, the arrangement and spatial co-localization of different cell types are recognized as critical factors influencing cancer progression and drug response. By inspecting pathology images, humans easily identify topological structures formed by cells in the form of loops, tunnels, and clusters. However, modern machine learning frameworks for biomedical image analysis are primarily adapted from methods designed for natural images. They do not explicitly integrate these visually apparent spatial context features in their learning, rendering them lacking essential spatial semantic grounding.

In my research I developed methods that integrate mathematical modeling with computer vision to create more robust solutions, particularly in the context of advancing cancer research. I quantify the clusters and loops formed by different cell types using TDA. By leveraging TDA and spatial statistics, I describe structural patterns formed by the cells. Through innovative integration of these spatial descriptors, I enhance deep learning models for computational pathology in the areas of: data generation, representation learning, and model optimization, with applications to spatially conditioned image generation, context-aware cell classification, and cell localization and instance segmentation. These works primarily focused on patch-level analysis. In my current research, I target mathematical modeling at the whole slide image (WSI) level, aiming to uncover novel biomarkers for survival and treatment response while guiding deep learning models for WSI-level tasks. With the innovative and interdisciplinary nature of research, there are so many opportunities and applications to explore. I look forward to establishing collaborations to explore these problems and address their challenges.