Drop an image anywhere, or choose an example.
Drop an image anywhere, or choose an example.
StarDist finds cell nuclei in 2D fluorescence microscopy by predicting, for every pixel, the probability that it lies inside a nucleus and the distances to the nucleus boundary along 32 rays. Each candidate is a star-convex polygon; non-maximum suppression keeps the most probable polygon where candidates overlap.
This app runs StarDist's pretrained 2D_versatile_fluo model, trained on a subset of the 2018 Data Science Bowl, with the model's tuned thresholds. The image is normalized to its 1st and 99.8th percentiles as in the StarDist examples. The model was converted to ONNX and runs with ONNX Runtime on the CPU. Non-maximum suppression and label drawing are ports of StarDist's code. On the example, the labels equal Python StarDist 0.9.1 to the pixel.
The model expects one fluorescence channel with bright nuclei on a dark background, at roughly the pixel size of the training images (nuclei about 10 to 40 pixels across). It is not for brightfield or H&E images, whole cells or 3D stacks. The table gives each nucleus's centroid in pixels (x to the right, y down), its area in pixels and its mean intensity in the segmented channel.
Images are processed locally in this browser and are never uploaded. The model is downloaded once from the Neurodesk dataset on Hugging Face and checked against its checksum.