itasc-cellpose

Standalone Cellpose segment + track tool. It works on one or two channels:

  • One channel: run a local Cellpose-SAM model for native masks, then link them across time with laptrack into tracked labels, a self-contained “segment then track” product.

  • Two channels: joint mode (and only joint mode). The first channel is the anchor (segmented + tracked); the second channel’s foreground is flowed onto it, giving one second-channel object per anchor object, sharing its track id.

It does not need the rest of the ITASC pipeline. Every result is added to the napari viewer as a layer, and there is no output directory: you save whichever layers you want via napari’s own Save Selected Layers. There is no nucleus/cell vocabulary; conventionally Channel 1 is the nucleus and Channel 2 the cell, but nothing assumes it.

The integrated ITASC app uses this distribution’s Cellpose runner and divergence maps for its in-app stage; those modules still ship here and are imported directly by the orchestrator. This README covers the standalone napari surface, which is the segment + track tool.

Install

pip install itasc-cellpose                       # package + Qt-free helpers
pip install "itasc-cellpose[cellpose]"           # + the Cellpose-SAM model
pip install "itasc-cellpose[cellpose,laptrack]"  # + the laptrack tracker

This pulls in itasc-core. Everything installs into the shared itasc.* namespace (PEP 420), so import itasc.cellpose works with or without the full orchestrator. The model (cellpose, torch, torchvision) is the [cellpose] extra and laptrack/pandas are the [laptrack] extra, both imported lazily, so importing the package does not require them.

Use

  • napari plugin: add the Cellpose Segment + Track widget. Each channel takes its input from the active image layer: select an image in the viewer and click the channel’s ⧉ pill to bind it. There is no file loading and no layout to declare. That pill then doubles as a status light: it stays lit while its bound layer is in the viewer and goes dark (releasing the channel) when the layer is removed.

    • Channel 1: the anchor stack (typically the nucleus).

    • Channel 2: optional second stack (typically the cell).

    Every plane is segmented individually, so the input shape needs no 2D/2D+t/3D/ 3D+t selection. Axis identity only matters for tracking, and is inferred: the last two axes are Y, X; of the remaining leading axes the shorter is Z and the longer is time (the preview status shows the inferred T/Z).

    Channel 1 carries the single-channel pipeline and surfaces three maps from a single Cellpose pass: the native masks, the sigmoid probability map, and the flow (HSV-coloured by direction). Preview (▷) runs them on the current frame only so you can tune diameter / min-size / gamma and the Prob threshold (a [0, 1] cutoff read in the prob-map image’s own space, 0.5 = Cellpose’s default; reversed through the inverse sigmoid to the raw cellprob Cellpose thresholds, where lower finds more and larger masks and higher is stricter), the Flow error tolerance (Cellpose’s flow_threshold, relabelled so its direction reads right: it is the per-mask flow-error budget, so higher is more permissive; 0.4 default, 0 disables the QC and keeps every mask; this, not Prob threshold, is usually the knob that gates how many masks survive) and Iterations (niter flow steps, 0 = auto) against what you see; Segment (▶) runs the whole stack and streams each frame into the viewer as it is computed, so the masks/prob/flow layers fill in live instead of appearing only at the end. Track (⊳) links the masks axis-by-axis: it stitches the z axis by overlap (so an object spanning planes becomes one), then tracks time by motion with laptrack (max-distance / frame-gap, in the Channel 2 & tracking parameters section). Results land as layers tagged [Channel 1] (… masks, … prob, … flow, … tracked, and … preview / … prob preview / … flow preview); save whichever you want via napari.

    Channel 2 is never segmented on its own: it is always run jointly. Once a second channel is set its two actions mirror Channel 1’s: Preview (▷) runs the joint assignment on the current frame so you can tune the Channel-2 params first, and Run (▶) commits it over the whole stack. Either way Channel 1 is segmented + tracked, then each Channel-2 foreground pixel is flowed along Cellpose’s flow field (blended with a pull toward the nearest Channel-1 object) and assigned to one. You get one Channel-2 object per Channel-1 object, sharing its track id: [Channel 1] tracked and [Channel 2] tracked are paired by construction (Channel 2 is tracked by inheriting Channel 1’s tracks, not a separate tracker). Channel 2 & tracking parameters: Diameter/Min size/Gamma shape its Cellpose flow field; FG threshold (foreground cutoff on the sigmoid), Flow weight (Cellpose flow vs. pull-to-anchor) and Max assign radius (foreground farther than this from any anchor is left unassigned) drive the assignment; Max distance / Max frame gap tune the Channel-1 tracker that both the joint anchor and Channel 1’s own Track action run.

    The embedded Correction panel (the ultrack/OverlapDB-free cell corrector) edits whatever Labels layer is currently active (typically [Channel 2] tracked) in place, with the full DB-free toolkit: select (left-click), spawn (middle-click empty space), erase (middle-click a cell or Delete), merge (Ctrl+left), swap / attach to track (Ctrl+right), grow / link the selected track (Ctrl+middle), draw / split (Shift+ left / right-drag), plus fill-holes and stranded-fragment cleanup and Ctrl+Z undo. A built-in retracker re-links the tracks from the current frame outward by geometric similarity: E forward, Q backward (the Retrack max dist parameter gates a match). It targets 2D+t (single-Z) labels; save the corrected layer via napari.

  • Headless / scripting:

    import tifffile
from itasc.cellpose import cellpose_runner, native_masks, track_laptrack

stack = cellpose_runner.to_tzyx(tifffile.imread("cell.tif"), "2D+t")
params = cellpose_runner.CellParams(diameter=0.0, min_size=0, gamma=1.0)
masks = native_masks.run_cell_masks_stack(stack, params)       # (T, Z, Y, X)
tracked = track_laptrack.track_masks(masks, max_distance=15.0)  # tracked labels

I/O contract

  • Input: any 2-D..4-D image layer per channel, bound from the active layer via the channel’s ⧉ pill (there is no file loading). There is no required 0_input/ layout and no layout to declare. Every plane is segmented individually; for tracking the shorter leading axis is read as Z, the longer as time. Channel 1 is required; Channel 2 is optional (and turns the run into joint mode). The headless API below still reads .tif files directly.

  • Output: napari layers, not files. Masks/tracked/preview are added as int32 Labels layers tagged [Channel 1] / [Channel 2] (singleton-Z squeezed to (T, Y, X) for 2D+t data); the user saves them with napari’s Save Selected Layers. The headless API above still returns plain (T, Z, Y, X) arrays.