HOCT

HOCT reconstructs cell trajectories and divisions from time-lapse microscopy images and instance segmentation masks. It scores candidate links with a Higher-Order Cell Tracking Transformer and selects a consistent lineage with an integer linear programming (ILP) solver.

This package provides pretrained-model inference, graph construction, tracking, and incremental correction. It supports 2D and 3D time series and exports GEFF graphs or Cell Tracking Challenge (CTC) results.

Start here

• Install HOCT and choose the dependencies for your workflow.
• Track cells from images and masks using the CLI or Python.
• Correct links and fine-tune in the Napari demo, or connect your own UI.
• Look up the Python API for graph, prediction, and correction parameters.
• Develop and build documentation from a checkout.

What you need

Supply an instance mask for every image frame: each cell has a distinct positive integer label within a frame, and background is zero. Labels do not need to be consistent across time; discovering those links is the tracking task. HOCT expects spatially aligned images and masks with matching shapes.

Pretrained weights download on first use. CUDA speeds up inference, while CPU also works. Tracking requires an available ILP backend; see installation.

For the model and correction method, see the HOCT paper. The source code and examples are on GitHub.