A Detector-Independent Quality Score for Cell Segmentation Without Ground Truth in 3D Live Fluorescence Microscopy

Deep-learning techniques have enabled a breakthrough in the robustness and execution time of cell segmentation algorithms for fluorescence microscopy datasets. However, the heterogeneity, dimensionality and ever-growing size of 3D+time datasets challenge the evaluation of measurements. Here we propose an estimator of cell segmentation accuracy that is detector-independent and does not need any ground-truth nor priors on object appearance. To assign a segmentation quality score, our method learns the dynamic parameters of each cell to detect inconsistencies in local displacements induced by segmentation errors. Using simulations that approximate the dynamics of cellular aggregates, we demonstrate the score ability to rank the performance of detectors up to 40% of false positives. We evaluated our method on two experimental datasets presenting contrasting scenarios in density and dynamics (stem cells nuclei in organoids and carcinoma cells in a collagen matrix) using two state-of-the-art deep-learning-based segmentation tools (Stardist3D and Cellpose). Our score is able to appropriately rank their performances as reflected by accuracy (centroid matching) and precision (segmentation overlap).