Periodic Reporting for period 2 - AI4LIFE (Artificial Intelligence for Image Data Analysis in the Life Sciences)
Periodo di rendicontazione: 2023-09-01 al 2025-08-31
Artificial intelligence (AI) and machine learning (ML) have enabled and accelerated frontier research in the life sciences; however, access to such methods remains limited. Access to hardware, software, knowledge, and training is scarce, while methods are often insufficiently documented and hard to find. Furthermore, even though modern AI-based methods can generalise to unseen data, no widely adopted standard exists to enable the sharing and fine-tuning of pretrained models between different analysis tools. Existing user-facing platforms operate independently, often failing to comply with FAIR (Findable, Accessible, Interoperable, Reusable) and Open Science standards.
The fast pace of AI/ML makes it hard for non-specialists to stay current, and many published methods lack standardised metadata, limiting reproducibility. To enable life science communities to benefit from AI-powered image analysis, AI4Life builds bridges, providing urgently needed services on European research infrastructures. We are building an open, accessible, community-driven repository of FAIR pretrained models (the BioImage Model Zoo) and delivering them to life scientists, including those without computational expertise. Our training and support empower responsible AI use, while contributor services and open standards foster community participation and tool interoperability. Through Open Calls and Public Challenges, we help tackle unsolved image-analysis problems. Bringing together AI/ML experts, open-source tool developers, compute providers, and research infrastructures, our consortium works toward enabling life scientists to fully harness AI. We expect thousands of European researchers to benefit directly from these resources, amplifying AI’s impact across the life sciences.
The fast pace of AI/ML makes it hard for non-specialists to stay current, and many published methods lack standardised metadata, limiting reproducibility. To enable life science communities to benefit from AI-powered image analysis, AI4Life builds bridges, providing urgently needed services on European research infrastructures. We are building an open, accessible, community-driven repository of FAIR pretrained models (the BioImage Model Zoo) and delivering them to life scientists, including those without computational expertise. Our training and support empower responsible AI use, while contributor services and open standards foster community participation and tool interoperability. Through Open Calls and Public Challenges, we help tackle unsolved image-analysis problems. Bringing together AI/ML experts, open-source tool developers, compute providers, and research infrastructures, our consortium works toward enabling life scientists to fully harness AI. We expect thousands of European researchers to benefit directly from these resources, amplifying AI’s impact across the life sciences.
A major part of AI4Life’s work has been dedicated to establishing services and community standards. Our central service — the BioImage Model Zoo (BMZ) — has transitioned from prototype to production-ready, serving pre-trained models to users worldwide, with >29k visits and >38k page views (Aug 2024–Aug 2025) from 117 countries.
All model bundles use a community-developed metadata standard, enabling developers and advanced users to run models through a single API in Python or Java. For GUI-based users — the majority — we expanded BioImage Model Zoo integration to nine end-user tools: deepImageJ/Fiji, ilastik, ImJoy, Icy, QuPath, StarDist, CAREamics, SpotMax, and BiaPy.
To support both developers and end users, we created contributor services that streamline deployment of new AI methods. Key achievements include Python and Java libraries for local validation, continuous-integration testing that automatically verifies all submitted models, and a GUI tool for easier metadata entry. Through DL4MicEverywhere, container integration preserves and runs notebooks as exact snapshots of developer environments. So far, 13 community partners have contributed 122 models, 46 datasets, and 65 applications.
We improved website accessibility, tutorials, documentation, and redeveloped the online model-testing platform for better scalability. To engage the community, we organised 10 hackathons, 5 conference workshops, and multiple training courses, totalling over 180 dissemination activities that reached thousands of life scientists. The community submitted 151 eligible Open Call applications across 3 rounds, from which we selected 34 projects for consultations and 22 for full support. We established, refined, and published our selection procedures to guide future initiatives towards most impactful applications. All solutions have been made available on github. For method developers, AI4Life hosted 3 Public Challenges, receiving 242 submissions from 137 registered participants.
Finally, we developed a community standard for annotated/ground truth data, enabling submission of such data into the BioImage Archive (BIA). The corresponding BIA Reference Data Service has been accessed by 2634 unique visitors per month (average over the last 2 years), generating ~23,000 web page accesses per month.
All model bundles use a community-developed metadata standard, enabling developers and advanced users to run models through a single API in Python or Java. For GUI-based users — the majority — we expanded BioImage Model Zoo integration to nine end-user tools: deepImageJ/Fiji, ilastik, ImJoy, Icy, QuPath, StarDist, CAREamics, SpotMax, and BiaPy.
To support both developers and end users, we created contributor services that streamline deployment of new AI methods. Key achievements include Python and Java libraries for local validation, continuous-integration testing that automatically verifies all submitted models, and a GUI tool for easier metadata entry. Through DL4MicEverywhere, container integration preserves and runs notebooks as exact snapshots of developer environments. So far, 13 community partners have contributed 122 models, 46 datasets, and 65 applications.
We improved website accessibility, tutorials, documentation, and redeveloped the online model-testing platform for better scalability. To engage the community, we organised 10 hackathons, 5 conference workshops, and multiple training courses, totalling over 180 dissemination activities that reached thousands of life scientists. The community submitted 151 eligible Open Call applications across 3 rounds, from which we selected 34 projects for consultations and 22 for full support. We established, refined, and published our selection procedures to guide future initiatives towards most impactful applications. All solutions have been made available on github. For method developers, AI4Life hosted 3 Public Challenges, receiving 242 submissions from 137 registered participants.
Finally, we developed a community standard for annotated/ground truth data, enabling submission of such data into the BioImage Archive (BIA). The corresponding BIA Reference Data Service has been accessed by 2634 unique visitors per month (average over the last 2 years), generating ~23,000 web page accesses per month.
AI4Life develops the only image analysis model zoo specifically for life sciences, targeting users who are not deep learning experts. The added value lies in direct integration with GUI-based software tools usable without coding. Many key tools in bioimage analysis can run BMZ models, including Fiji/deepImageJ, ilastik, Icy, ImJoy, QuPath, CAREamics, StarDist, SpotMax, and BiaPy.
The BMZ already serves models covering multiple tasks and modalities. By Reporting Period 2 of the AI4Life project, the BMZ recorded >29k visits and >38k page views from 117 countries, with >45k model downloads via Zenodo and S3.
Beyond the state of the art, AI4Life combines:
- Global reach: active use in Europe, North America, and Asia.
- Seamless integration: GUI access in nine open-source tools lowers barriers for non-specialists.
- Standards: all models packaged with open metadata, ensuring reproducibility.
- Community contributions: 122 models, 46 datasets, 65 applications openly shared, supported by validation pipelines.
- Benchmarks: three public challenges delivered evaluation datasets and leaderboards for transparent progress.
These advances establish AI4Life as the reference infrastructure for AI in bioimage analysis. To ensure uptake and long-term success, further efforts are needed in (i) sustainability of hosting and support, (ii) standardisation of metadata and annotated datasets, and (iii) integration with wider European infrastructures such as EOSC and the Health Data Space. In the future, the approach could extend beyond imaging, supporting other biomedical data types and workflows.
The BMZ already serves models covering multiple tasks and modalities. By Reporting Period 2 of the AI4Life project, the BMZ recorded >29k visits and >38k page views from 117 countries, with >45k model downloads via Zenodo and S3.
Beyond the state of the art, AI4Life combines:
- Global reach: active use in Europe, North America, and Asia.
- Seamless integration: GUI access in nine open-source tools lowers barriers for non-specialists.
- Standards: all models packaged with open metadata, ensuring reproducibility.
- Community contributions: 122 models, 46 datasets, 65 applications openly shared, supported by validation pipelines.
- Benchmarks: three public challenges delivered evaluation datasets and leaderboards for transparent progress.
These advances establish AI4Life as the reference infrastructure for AI in bioimage analysis. To ensure uptake and long-term success, further efforts are needed in (i) sustainability of hosting and support, (ii) standardisation of metadata and annotated datasets, and (iii) integration with wider European infrastructures such as EOSC and the Health Data Space. In the future, the approach could extend beyond imaging, supporting other biomedical data types and workflows.