Description
Multiplexed tissue imaging provides an unprecedented understanding of tissue biology by capturing the spatial relationships between cells and their protein expression profiles. Common approaches use highly multiplexed fluorescence microscopy to measure tissue slides as images where each channel represents the intensity of a particular protein marker. Single-cell features are generated by segmenting the cells and aggregating all pixel intensities per cell for all measured channels. The resulting feature table is then used for further analysis, including quality control, preprocessing, dimensionality reduction, and clustering. Cell segmentation masks play a critical role in the generation of feature tables. Thus, their accuracy is crucial for downstream analysis. However, segmentation remains challenging, with existing solutions being prone to errors. The performance of these segmentation methods is evaluated using metrics such as mean average precision, intersection-over-union, and F1 scores. In this work, we empirically examine how segmentation errors influence the outcomes of downstream analysis, providing insights into the epistemic uncertainty induced by segmentation errors. Specifically, we use a perturbation procedure to generate augmented segmentation masks, allowing for systematic investigations. Using this approach, we can quantify the propagated error in the subsequent pipeline, which includes neighborhood preservation, unsupervised clustering and phenotyping. We demonstrate that even small perturbations significantly impact cell type inference and that the commonly used IoU-F1 metric hides segmentation inaccuracies.