Quantification of the detection, segmentation and tracking performance.

Measurements were performed on two manually annotated time series of the Fluo-N2DL-HeLa data set that is part of the Cell Tracking Challenge [76] and was most similar to our application scenario (acquisition was performed with an Olympus IX81 microscope using a 10×/0.4 objective lens, a physical spacing of 0.645×0.645μm and 30 minute time intervals). The used metrics are DET (detection score), SEG (segmentation score), TRA (tracking score), OPCTB (average of SEG and TRA) and OPCSB (average of DET and SEG). All metrics have a value range between 0 and 1 with 1 being the optimum (see [76] for details on the measures). While Cellpose [16] yields better segmentation scores, it misses cells occasionally. On the other hand, our classical approach based on multi-scale Laplacian-of-Gaussian blob detection is able to find most cells with a slightly worse segmentation accuracy. Combining both approaches, i.e., complementing the Cellpose segmentation with additional cells that were only found by the classical method, yielded consistently the best results (highlighted in bold face letters). The experiments described in the main paper were consistently acquired with 3 minute time intervals, which further improves the tracking performance. Please note that neither parameter tuning of the classical method nor any retraining of Cellpose was performed. Thus, the top-scoring methods listed on the Cell Tracking Challenge leader board are still slightly higher but we expect our method to generalize better to unseen data sets.
(XLSX)