CVPR Daily - Tuesday

Pathologists analyze cell samples to determine the presence and type of cancer, assess its stage and aggressiveness, and evaluate patient responsiveness. However, automating these tasks is challenging due to the limited data available on the tumor microenvironment. Unlike natural images, which are abundant and anyone can annotate, pathology data, like all biomedical data, requires patient information and expert annotation . In overcoming this limitation, this work aims to create a generative model capable of producing synthetic labeled data to augment training and facilitate downstream tasks. “ We realized that when pathologists look at an image or data from pathology, they look at the distribution of different types of cells, how they colocalize, and the patterns they make, which has not been considered in a generative model before, ” Shahira tells us. “ We’re trying to address this by generating realistic cell layouts that capture this biology . With a reference sample, you can generate new samples that exhibit the same spatial and structural patterns or characteristics and use them efficiently in training for downstream tasks. This paper is about how to model this kind of tumor microenvironment, the cell layout, and these patterns. Given some specific characteristics, how do we train the model to generate data that satisfies these characteristics? ” Shahira’s approach differs from existing work on generating synthetic data in several ways. Previous methods focused on creating visually appealing images, whereas this paper takes it back to the beginning, focusing on the location and distribution of different cell types . The spatial organization and Shahira Abousamra is a senior PhD student at Stony Brook University, advised by Chao Chen and Dimitris Samaras, and working closely with Joel Saltz’s group. She speaks to us about her work on pathology data ahead of its poster this morning Topology-GuidedMulti-Class Cell Context Generation for Digital Pathology 8 DAILY CVPR Tuesday Poster Presentation