41 Topology-Guided Multi- Class … Computer Vision News and arrangement of cells are crucial in pathologists’ decision-making process. This aspect has not been incorporated into deep learning models before and is usually saved for analysis after data generation. Shahira introduces a novel approach by integrating these patterns into the deep learning process, developing a model that could effectively capture the complex structure of the tumor microenvironment and satisfy the desired conditions. “ First, we look at the cell layout as a point map of different classes , ” she explains. “ We see these different distributions that we want to capture. We also see that they form some clusters, holes, and gaps, and want to capture them all. We can capture the spatial colocalization using spatial statistics like cross K-functions, which we’ve used successfully in a previous cell classification paper. We talked to pathologists, who told us thatwhen they classify a cell, they don’t look at just one but at the whole region . We didn’twant the model to be fixed on the morphology and texture of just one cell but to have a larger context and used the cross K-function to model this spatial context in the image. ” Realizing that clusters, holes, and patterns are like topological features, Shahira used the persistent homology algorithm , a topology data analysis algorithm, to capture these characteristics and model the cell layout. Another challenge was how to train the model, given these features, to generate data that satisfies these spatial and topological characteristics. She attempted to use adversarial learning and different formulations with little success. Recognizing the need for an alternative approach, she discovered
RkJQdWJsaXNoZXIy NTc3NzU=