Evaluation of cell segmentation methods without reference segmentations

Cell segmentation is a cornerstone of many bioimage informatics studies, and inaccurate segmentation introduces error in downstream analysis. Evaluating segmentation results is thus a necessary step for developing segmentation methods as well as for choosing the most appropriate method for a particular type of sample. The evaluation process has typically involved comparison of segmentations with those generated by humans, which can be expensive and subject to unknown bias. We present here an approach to evaluating cell segmentation methods without relying upon comparison to results from humans. For this, we defined a number of segmentation quality metrics that can be applied to multichannel fluorescence images. We calculated these metrics for 14 previously described segmentation methods applied to datasets from four multiplexed microscope modalities covering five tissues. Using principal component analysis to combine the metrics, we defined an overall cell segmentation quality score and ranked the segmentation methods. We found that two deep learning–based methods performed the best overall, but that results for all methods could be significantly improved by postprocessing to ensure proper matching of cell and nuclear masks. Our evaluation tool is available as open source and all code and data are available in a Reproducible Research Archive.