Pushing the Limits of Cell Segmentation Models for Imaging Mass Cytometry

In recent years, learning-based segmentation methods have enabled precise quantification of cell type and morphology, but typically rely on large datasets with fully annotated ground truth (GT) labels. This paper explores the effects of imperfect labels on learning-based segmentation of imaging mass cytometry (IMC) data, and evaluates the generalisability to different tissue types. We show that removing 50% of cell annotations from GT masks only reduced the dice similarity coefficient (DSC) score by 0.015, when compared to a model trained on uncorrupted data. Furthermore, training a single-tissue model on imperfect labels only decreased DSC by 0.031 on an unseen tissue type, as compared to its multi-tissue counterpart. Additionally, bootstrapping the worst-performing model improved the DSC score from 0.720 to 0.829. Our findings show that considerable time savings may be afforded by using very weak labels, with minimal impact on overall performance. Our source code is available on GitHub: https://github.com/kimberley/ISBI2024.