Method / tool

ilastik

Interactive machine learning toolkit for bioimage analysis — users annotate a handful of pixels or objects with brush strokes and a Random Forest classifier generalises to the full image, enabling no-code pixel classification, object detection, cell tracking, and 3D carving.

The problem — Many segmentation tasks are too sample-specific for a generalist deep-learning model and too labour-intensive for hand-drawn masks. The gap is the case where the biologist knows what to label but lacks the training data or the engineering time to train a supervised model from scratch.

What it is / how it works — ilastik is an interactive machine learning toolkit built around a Random Forest classifier. The user paints brush-stroke annotations on a handful of images; ilastik extracts a rich set of hand-crafted texture, edge, and Gaussian-derivative features, trains a Random Forest on those labels, and propagates predictions to the rest of the dataset in real time. The core workflow is Pixel Classification — every pixel receives a class probability — but the same interaction pattern drives Object Classification, automatic and manual cell tracking, and semi-automatic 3D carving of volumetric structures. A napari plugin (ilastik-napari) brings the same pixel classification loop into the scientific-Python viewer ecosystem. Berg et al. (2019) describe the full feature set and benchmark its reach across modalities from widefield fluorescence to electron microscopy.

Where it fits in the pipeline — ilastik is most useful at the segment stage as a rapid prototype segmenter: when a generalist model like Cellpose misfires on a novel morphology, ilastik's interactive loop lets you build a reasonable classifier in an hour, assess whether the problem is tractable, and export probability maps that a downstream tool can threshold or feed into a structured pipeline.

Where it breaks — ilastik's reproducibility story is the main production concern. Classifier state lives in an .ilp project file that captures labels and hyperparameters but not the full software environment, and small changes to feature selection or training pixels produce silently different outputs. Batch headless mode exists but requires the full ilastik environment; it does not slot into a conda-based pipeline with the same ease as a Python library. Scalability is a second ceiling: Random Forests hit memory limits on large 3D volumes before deep-learning methods. For a Screen-scale Cell Painting run or a terabyte light-sheet volume, ilastik is a prototyping entry point, not the production segmenter — the custom-vs-off-the-shelf decision still applies once the prototype confirms the task is viable.

Treat an ilastik classifier as a proof of concept, not a deployable component. Lock the `.ilp` file into version control, record the ilastik version, and plan a re-validation step before any result built on it is reported — the classifier is only as stable as the labelling session that produced it.

References

Appears in these notes

  • Choosing a Bioimage Analysis Pipeline — Off-the-Shelf vs CustomMost microscopy questions are answered by composing existing, validated tools — and a custom pipeline is justified only when off-the-shelf options fail on a specific, identifiable axis. This note gives the decision criteria and the signals that you have genuinely crossed into custom territory.
  • The Bioimage Analysis Ecosystem — Where Each Tool FitsBioimage analysis is not one tool but a landscape organized along two axes — interactive vs scriptable, and generalist vs specialist. Fiji, napari, CellProfiler, QuPath, ilastik, and the deep-learning segmenters each occupy a different cell of that grid, and choosing well means knowing which cell your problem lives in.

← Back to the constellation