Fig. 2: Benchmarking of CELLECT with state-of-the-art algorithms in the Cell Tracking Challenge.
From: CELLECT: contrastive embedding learning for large-scale efficient cell tracking
a–d, Performance evaluation on the mskcc-confocal dataset. e–h, Performance evaluation on the nih-light sheet dataset. a,e, The average number of errors per 1,000 ground truth edges for each error type. Errors were categorized into five conditions: false positive edges, false negative edges, identity switches, false positive divisions and false negative divisions. b,f, The proportion of error-free tracks (marked as tracking accuracy) versus tracking duration in terms of frames. c,g, Comparisons of average computational time cost for each frame among different algorithms. d,h, Hierarchical circular diagram of lineage tracing, with gray lines indicating cell division. Bold black lines and circles indicate false negative identity switch and division, respectively. The scores for the three models other than CELLECT and the accuracy curves of linajea are derived from a previous report35. ‘linajea + csc + sSVM’ refers to an enhanced version of linajea that incorporates a csc to model biological states (for example, division, continuation, polar body) and an sSVM to learn tracking cost functions from data. All evaluations were conducted using the first 270 frames of each of the three subvolumes for both datasets (mskcc-confocal and nih-light sheet), following a cross-validation protocol. The CELLECT models used in each comparison were trained on one of the two other subvolumes. Further details on the training and evaluation splits are provided in Supplementary Tables 1 and 2. Scale bars, 10 μm.
