Stardist Nuclei Segmentation

This repository packages the StarDist algorithm into a Cytomine Task compatible with the Cytomine App Engine. It detects and segments cell nuclei in H&E stained histology images using the pre-trained 2D_versatile_HE model, producing nucleus outlines as GeoJSON polygons along with a per-nucleus detection probability score for each input image.

GPU Support

GPU acceleration is supported but limited to NVIDIA GPUs only (via CUDA). AMD and other GPU vendors are not supported.

ROI Support

Each input image may optionally be paired with a GeoJSON file defining a region of interest (ROI). When provided, only nuclei whose centroid falls inside that region are included in the output. To use it, place a file named <index>.geojson alongside the image file in the images/ input array directory (e.g. images/0.geojson for the first image).

How to run it manually

Build the docker image yourself

docker build -t com/cytomine/nuclei-segmentation/stardist:1.1.0 .

Run on data

From this repository, run:

docker run --gpus all -v ./examples/inputs:/inputs -v ./local-outputs:/outputs --rm -it com/cytomine/nuclei-segmentation/stardist:1.1.0

You can then explore the results in the ./local-outputs directory.

Build bundle to upload on Cytomine

1. Build the docker image as described above
2. Save it as a tar archive:

docker save -o com.cytomine.nuclei.segmentation.stardist-1.1.0.tar com/cytomine/nuclei-segmentation/stardist:1.1.0

3. Build the bundle:

zip com.cytomine.nuclei.segmentation.stardist-1.1.0.zip descriptor.yml com.cytomine.nuclei.segmentation.stardist-1.1.0.tar

4. Upload the bundle on Cytomine