Automahout hands cell-tracking judgement to an agent: no human input on three public datasets, TRA 0.95–0.96

The judgement steps of cell tracking can now be carried out by an agent on the Mastodon/ELEPHANT platform: reading the data, choosing or training a detector, reviewing its output, correcting the lineage and deciding when to stop. ELEPHANT had already put sparse-annotation training and human correction into one loop, but what to annotate, when to train, which predictions were wrong, how to fix them and when to stop were still judgements a person made from the images for every new dataset. Run with no human input on training sequence 01 of three Cell Tracking Challenge datasets, TRA was 0.954 on the 2D fluorescence HeLa culture, 0.954 on the 2D phase-contrast PSC culture and 0.965 on the 3D C. elegans embryo; a baseline applying one pretrained detector as it is and linking with Trackastra scored TRA 0.929 and 0.900 on the two cultures and 0.202 on the embryo.

It reads the data first and then picks a route: a pretrained detector that passes its check is used as it is, and only when none passes does it train its own from random weights, advancing in time windows of 10 frames. Each window re-reads the cell radius from the image by scale selection, taking the median radius at which scale-normalised Laplacian-of-Gaussian blobs peak, and re-centres ELEPHANT's detection radius range and linking distance on it. The embryo grows from 4 to 362 cells while the median nuclear radius shrinks from 4.3 to 1.9 µm; on Fluo-N3DH-CE 01, every 30th frame gave estimated radii of 4.58, 3.66, 2.34, 1.88, 1.88, 1.88 and 1.50 µm against ground-truth median radii of 4.32, 3.55, 2.54, 2.57, 2.38, 2.28 and 1.92 µm.

Review and stopping rules let the agent judge its own work without ground truth: missed and false detections go back as ELEPHANT's own correction labels, and it stops only after two rounds running meet the criteria. Stopping requires, in two rounds running, at most 3% of references missed, false hits at most 2%, predictions at most 1.5 times the references and a size ratio of 0.67–1.5; after 20 rounds without Go it keeps the round that found the most references rather than the last. Every window on the embryo reached Go, most within a few rounds; the vision-language model was consulted once, at t = 5, and judged one doubtful reference not to be a nucleus although it was a real cell, so the reference was dropped.

The demonstrated scope is training sequence 01 of three Cell Tracking Challenge datasets, one run each, covering a 2D fluorescence culture, a 2D phase-contrast culture and a 3D C. elegans embryo, with no human annotation, region of interest, parameter or answer supplied. The authors propose a broader evaluation next: further challenge datasets, their test sequences through a regular submission, repeated runs, which operations matter, and a comparison with an expert using ELEPHANT.

The authors state the scores come from training sequence 01, which was also used during development, so they are not comparable with the challenge's published test-set results. Reference cells are bright blobs on a dark background, so the review applies to fluorescence; on phase contrast it currently depends on a pretrained detector passing its check. The stopping rule tolerates up to 3% missed references and can therefore leave a small number of real cells unlearned; none of its criteria is ground truth, and a result meeting all of them can still be wrong. The vision-language model judges large, well-separated cells well and small, crowded ones poorly, and can be wrong even after its check; on the embryo it rejected a real cell. A rejected reference is only dropped, never taught as background.