Periodic Reporting for period 1 - AI4LIFE (Artificial Intelligence for Image Data Analysis in the Life Sciences)
Okres sprawozdawczy: 2022-09-01 do 2023-08-31
Machine learning (ML) has enabled and accelerated frontier research in the life sciences, but democratized access to such methods is, unfortunately, not a given. Access to necessary hardware and software, knowledge and training, is limited, while methods are typically insufficiently documented and hard to find. Furthermore, even though modern Artificial Intelligence (AI)-based methods typically generalize well to unseen data, no standard exists to enable sharing and fine-tuning of pretrained models between different analysis tools. Existing user-facing platforms operate entirely independently from each other, often failing to comply with FAIR (Findable, Accessible, Interoperable, Reusable) data and OpenScience standards. The field of AI and ML is developing at a staggering pace, making it impossible for the non-specialist to stay up to date. To enable the life science communities to benefit from AI/ML-powered image analysis methods, AI4Life builds bridges, providing urgently needed services on the common European research infrastructures. Our goal is to build an open, accessible, community-driven repository of FAIR pre-trained AI models (the BioImage Model Zoo) and develop services to deliver these models to life scientists, including those without substantial computational expertise. Our direct support and ample training activities will prepare life scientists for responsible use of AI methods, while contributor services and open standards drive community contributions of new models and interoperability between analysis tools. Open calls and public challenges will provide state-of-the-art solutions to yet unsolved image analysis problems in the life sciences. Our consortium brings together AI/ML researchers, developers of popular open source image analysis tools, providers of European-scale storage and compute services and European life sciences Research Infrastructures -- all united behind the common goal to enable life scientists to fully benefit from the untapped but potentially tremendous power of AI-based analysis methods.
The first year of AI4Life has been dedicated to the establishment of the services and the development of the missing community standards. Our central service - the BioImage Model Zoo - has gone out of the prototype stage, serving pre-trained AI-based models to thousands of users in Europe, USA, China and Japan. All the model bundles follow the same open metadata standard that we have developed in consultation with the community, allowing method developers and advanced users to access and execute all the models through the same Application Programming Interface (API), in Python or in Java. For users that prefer point-and-click Graphical User Interface (GUI)-based tools - the majority of the bioimage analysis community - we have expanded the partnerships of the Model Zoo to so far nine end-user facing software tools. The Zoo models can therefore now be executed without code in deepImageJ/Fiji, ilastik, ImJoy, Icy, QPath and StarDist.
The Model Zoo and the AI4Life project in general aims to serve both method developers and end users. To increase the engagement of method developers, we are working on contributor services that will streamline, simplify and robustify the deployment of new AI-based methods through the Model Zoo. Here, the main accomplishments of the first year include the development of Python and Java libraries to operate with models and validate them locally before submission to the Zoo, as well as the development of the first version of continuous integration tools for the Zoo, automatically testing all submitted models. The ongoing integration of JupyterLite and exploration of container technologies within the Model Zoo front-end and back-end has laid the foundation for a new deployment service of Jupyter notebooks as user-facing web-apps, further lowering the barrier of entry for model developers wishing to address end users directly.
For end users, we have greatly increased the accessibility of the website, created multiple tutorials and documentation pages and performed overall improvements to the User Interface of the interaction with models. Furthermore, the user community was heavily involved in the first Open Call, submitting more than 70 projects which highlighted outstanding problems across a broad range of applications and modalities. We developed the necessary procedures for project selection and openly published them on Zenodo to seed future initiatives. Finally, we developed a standard for annotated data to enable submission of such data - usually as ground-truth for training of AI-based methods - into the BioImage Archive. As one of the first tests, the standard will be used for data annotated through Open Calls.
The Model Zoo and the AI4Life project in general aims to serve both method developers and end users. To increase the engagement of method developers, we are working on contributor services that will streamline, simplify and robustify the deployment of new AI-based methods through the Model Zoo. Here, the main accomplishments of the first year include the development of Python and Java libraries to operate with models and validate them locally before submission to the Zoo, as well as the development of the first version of continuous integration tools for the Zoo, automatically testing all submitted models. The ongoing integration of JupyterLite and exploration of container technologies within the Model Zoo front-end and back-end has laid the foundation for a new deployment service of Jupyter notebooks as user-facing web-apps, further lowering the barrier of entry for model developers wishing to address end users directly.
For end users, we have greatly increased the accessibility of the website, created multiple tutorials and documentation pages and performed overall improvements to the User Interface of the interaction with models. Furthermore, the user community was heavily involved in the first Open Call, submitting more than 70 projects which highlighted outstanding problems across a broad range of applications and modalities. We developed the necessary procedures for project selection and openly published them on Zenodo to seed future initiatives. Finally, we developed a standard for annotated data to enable submission of such data - usually as ground-truth for training of AI-based methods - into the BioImage Archive. As one of the first tests, the standard will be used for data annotated through Open Calls.
AI4Life develops the only model zoo that targets users in the life sciences, who are not necessarily deep learning experts. The key difference lies in the direct integration of our models with end-user facing software tools that can be used without coding. Many of the key tools in the bioimage analysis community can run the Zoo models, including Fiji/deepImageJ, ilastik, Icy, ImJoy, QuPath, StarDist and ZeroCostDL4Mic. The Zoo is already serving a wide variety of models that have in total been downloaded more than 45.000 times. While a lot still has to be accomplished to bring the Zoo to production technology readiness level, thousands of users all over the world are already actively downloading models.