Cross-disciplinary reuse of BioImage.IO AI models in Galaxy

This practice demonstrates transfer learning across scientific domains: a pretrained AI model from the BioImage.IO Model Zoo, trained on biological data, is reused to produce meaningful output on Earth-science data. AI enters the workflow through the Galaxy tool “Process image using a BioImage.IO model”, which performs inference on a given image using models from BioImage.IO. In the practice, the same model (NucleiSegmentationBoundaryModel) is applied both to its original task, cell nuclei segmentation in fluorescence microscopy, and to water body detection in Landsat imagery of the Alaska Arctic Coastal Plain. For the latter, a cloud-masked, peak-summer near-infrared band is inverted to match the model’s expected input morphology. Supported by upstream and downstream image analysis tools in Galaxy, the workflow standardizes the analysis and processes multiple images in parallel. The output of the AI inference tool is an enriched artefact: an image depicting bounded regions of cell nuclei or water bodies, dissecting them from the background.

The AI tool was created to enable inference from pretrained BioImage.IO models within Galaxy, an open, web-based execution platform. Combined with other image analysis tools in Galaxy, the resulting AI inference workflow standardizes the analysis, obviating the need for unreliable manual steps, and offers parallel processing of multiple images. The workflows were built for downstream use, so that any Galaxy user or researcher can run them. The practice emerged from collaboration between the Image Analysis, Earth Data Science and Machine Learning groups within the Galaxy community, all of which develop AI-enabled scientific workflows for biomedical images, climate and biological data. The team combines AI and domain expertise, with life and data scientists; skills were built through learning by doing and internal collaboration across groups. Smaller domain-specific groups, regular workshops and hackathons shaped the practice. It is disseminated through the Galaxy Training Network (GTN), which shares over 500 practices across more than 20 scientific domains, and a “Feedback” section invites users to report their experiences, which then become automatically available in the practice. Galaxy is open source and offers access to institutional HPC as well as national and European infrastructure. The AI models are pretrained and taken from the BioImage.IO Model Zoo, together with an example microscopy test image; the Landsat composite was purpose-built for the demonstration. Since the model is reused without retraining, no new training data were needed, and the only validation data come from an independent NDWI comparison across five existing satellite scenes. The practice follows FAIR principles and published community-based standards. The tool relies on the BioImage.IO RDF metadata schema, which specifies model inputs, outputs, axes and size requirements; this is the standard that makes cross-domain reuse possible. The tool does not itself assess or certify other AI systems. Domain knowledge is important for interpreting the outputs, so the practice provides an extensive explanation of the intermediate steps and the final output.

The practice enables reproducible science and lowers the barriers to running complex AI models for scientific analyses to only a few button clicks. It demonstrates transfer learning: a model trained on biological data can also produce meaningful output on datasets from a completely different domain. Benchmarked against the Normalized Difference Water Index (NDWI), an established remote-sensing method, the model’s estimates across five independent Landsat scenes ranged from 0.74 to 1.19 times the NDWI values, indicating that transfer quality is scene-dependent rather than uniform. Users can run the analysis entirely within a browser, without installing software packages or meeting hardware requirements; the associated community groups support users’ queries and organize online and in-person training events several times a year. The practice is updated periodically by the community, as indicated by its “Revision” counter. It follows FAIR standards and is licensed under Creative Commons Attribution 4.0 International, while the GTN framework is licensed under MIT; the tools in the AI workflow are open source on public GitHub repositories and can be adapted as needed. The tool’s help section discloses that its inputs are a pretrained AI/ML model and a test image, and that its output is an image. No formal ethical review is documented, as the data raise no privacy concerns. The main risk is epistemic: applying a model outside its original domain without retraining could lead to over-trusting results. The team mitigates this by justifying the transfer through a morphological rationale (bright objects on a dark background), validating against NDWI, and framing the case as an experimental demonstration rather than a validated production method. The raw model output is not directly usable: two output channels must be combined and the threshold tuned (0.6 instead of 0.5) to separate clustered objects. In terms of frugality, the practice uses inference only on shared infrastructure; as NDWI already exists for water detection, the AI model here serves a demonstrative purpose.

Tutorial (including workflow, dataset): https://gxy.io/GTN:T00500