Study of Shape Fusion Algorithms for 3D Time-Lapse Microscopy

In biomedical image processing, reference data used to benchmark segmentation algorithms are typically obtained by fusing expert annotations. We compare popular fusion algorithms (majority voting, STAPLE, and Robust Bayesian Fusion) and study (1) how their results depend on the variability of input annotations and (2) how well they preserve shapes w.r.t, biologically relevant measures (BRM) such as volume and average roundness. We also study (3) the correlation between BRM and measures common in image processing (namely, Jaccard index and Hausdorff distance). For our experiments, we created 3D time-lapse segmentation annotations of migrating cells with complex dynamic shapes (Fluo-C3DL-MDA231 dataset from the Cell Tracking Challenge). The full annotations are publicly available. The study found that none of the selected methods produced satisfactory results w.r.t. BRM, highlighting the need for new methods.