Pushing the Limits of Cell Segmentation Models for Imaging Mass Cytometry

Imaging mass cytometry (IMC) is a relatively new technique for imagingbiological tissue at subcellular resolution. In recent years, learning-basedsegmentation methods have enabled precise quantification of cell type andmorphology, but typically rely on large datasets with fully annotated groundtruth (GT) labels. This paper explores the effects of imperfect labels onlearning-based segmentation models and evaluates the generalisability of thesemodels to different tissue types. Our results show that removing 50annotations from GT masks only reduces the dice similarity coefficient (DSC)score to 0.874 (from 0.889 achieved by a model trained on fully annotated GTmasks). This implies that annotation time can in fact be reduced by at leasthalf without detrimentally affecting performance. Furthermore, training oursingle-tissue model on imperfect labels only decreases DSC by 0.031 on anunseen tissue type compared to its multi-tissue counterpart, with negligiblequalitative differences in segmentation. Additionally, bootstrapping theworst-performing model (with 5improves its original DSC score of 0.720 to 0.829. These findings imply thatless time and work can be put into the process of producing comparablesegmentation models; this includes eliminating the need for multiple IMC tissuetypes during training, whilst also providing the potential for models with veryfew labels to improve on themselves. Source code is available on GitHub:https://github.com/kimberley/ISBI2024.