Robustness of Spatial Cell Point-Cloud Features for Cancer
Subtype Prediction in Multiplex Immunofluorescent Lung Tumor Images – Love Nordling

The spatial organization of cells within cancer tissue carries rich information about tumor biology and immune interactions. Using multiplex immunofluorescence, individual cells can be segmented and assigned phenotypes based on marker expression, after which each cell is reduced to a coordinate and a label; together forming a spatial point cloud of the tissue. These abstractions open the door for quantitative spatial statistics, such as clustering measures and neighborhood analysis, as well as geometric deep learning approaches that treat tissue as graphs.

In this project, we ask how much biologically meaningful information is retained in cell point clouds, and whether such abstractions can be used to distinguish cancer subtypes. We compare interpretable spatial statistics with geometric learning methods, and relate their performance to that of pathologists tasked with recognizing lung adenocarcinoma from point cloud representations. A further focus is on robustness: we analyze how different segmentation and classification pipelines (InForm, CellProfiler, and Cellpose) affect the resulting point clouds and the downstream spatial features. By framing tissue as
a set of spatially organized points, the project aims to clarify both the potential and the limitations of point-cloud-based analysis for computational pathology.

About Love Nordling