CITrackastra
Public microscopy · actual predictions

Follow objects
through time.

Inspect predicted centroid paths on calibrated public data. Short and full-length examples verify the workflow; they do not establish biological accuracy.

ACTUAL MODEL PREDICTION

Fluo-N2DH-GOWT1 (92 frames)

GFP-GOWT1 mouse stem-cell nuclei · 0.24 µm/pixel · 300.0 s/frame.

Source microscopy frame with predicted centroid trajectories

Display: shared 1–99.8 percentile stretch. Paths: actual general_2d / greedy predictions. Input labels are CTC TRA reference markers, which may be smaller than full cells or nuclei.

2058 observations · 29 tracklets · 2031 links. Dataset & attribution ↗ · CC BY 4.0.

Longer division example: 92 source frames; 2 documented binary reference divisions and predicted division events: 1. Counts alone are not an accuracy score.

Conventional 2D tracking movie with IDs and predicted division events.

Divisualisation spacetime movie: the vertical display axis is time.

ACTUAL MODEL PREDICTION

Fluo-N2DH-SIM+ (65 frames)

Simulated fluorescent HL60 nuclei · 0.125 µm/pixel · 1740.0 s/frame.

Source microscopy frame with predicted centroid trajectories

Display: shared 1–99.8 percentile stretch. Paths: actual general_2d / greedy predictions. Input labels are CTC TRA reference markers, which may be smaller than full cells or nuclei.

2607 observations · 84 tracklets · 2565 links. Dataset & attribution ↗ · CC BY 4.0.

Longer division example: 65 source frames; 28 documented binary reference divisions and predicted division events: 21. Counts alone are not an accuracy score.

Conventional 2D tracking movie with IDs and predicted division events.

Divisualisation spacetime movie: the vertical display axis is time.

ACTUAL MODEL PREDICTION

Fluo-N2DH-GOWT1 (6 frames)

GFP-GOWT1 mouse stem-cell nuclei · 0.24 µm/pixel · 300.0 s/frame.

Source microscopy frame with predicted centroid trajectories

Display: shared 1–99.8 percentile stretch. Paths: actual general_2d / greedy predictions. Input labels are CTC TRA reference markers, which may be smaller than full cells or nuclei.

138 observations · 23 tracklets · 115 links. Dataset & attribution ↗ · CC BY 4.0.

Divisualisation spacetime movie: the vertical display axis is time.

ACTUAL MODEL PREDICTION

Fluo-N2DH-SIM+ (6 frames)

Simulated fluorescent HL60 nuclei · 0.125 µm/pixel · 1740.0 s/frame.

Source microscopy frame with predicted centroid trajectories

Display: shared 1–99.8 percentile stretch. Paths: actual general_2d / greedy predictions. Input labels are CTC TRA reference markers, which may be smaller than full cells or nuclei.

180 observations · 30 tracklets · 150 links. Dataset & attribution ↗ · CC BY 4.0.

Divisualisation spacetime movie: the vertical display axis is time.

ACTUAL MODEL PREDICTION

PhC-C2DH-U373 (6 frames)

Phase-contrast glioblastoma-astrocytoma cells · 0.65 µm/pixel · 900.0 s/frame.

Source microscopy frame with predicted centroid trajectories

Display: shared 1–99.8 percentile stretch. Paths: actual general_2d / greedy predictions. Input labels are CTC TRA reference markers, which may be smaller than full cells or nuclei.

36 observations · 6 tracklets · 30 links. Dataset & attribution ↗ · CC BY 4.0.

Divisualisation spacetime movie: the vertical display axis is time.

Read the examples correctly

Source frames and reference markers are preserved. These original time series are packaged as single-well, single-field NGFF test plates. Their physical acquisition calibration is retained. The model uses intensities and object geometry; reference IDs are not association scores. CTC training data may overlap pretrained-model training. No independent generalization or segmentation accuracy is claimed. Synthetic integration tests separately exercise HCS isolation, divisions, gaps and rejection of unsupported inputs.

Calibration comes from CTC acquisition descriptions. Original TIFF archives, checksums and converted OME-Zarr stores remain local, outside Git. This site distributes derived display frames, predicted centroid paths and review movies, attributed to the CTC datasets under CC BY 4.0.