A menagerie of deep-learning models

EMBL scientists collaborate to build easy-to-use ‘Zoo’ of pre-trained AI models to help biologists and microscopists better analyse their biological images

Imagine a scientist studying how plants respond to environmental stress. She wants to hone in on cell membrane structures in numerous samples to get a better sense of how they are affected by long-term drought conditions. She can do this manually, but it will require months and months of work, and she lacks the skills to build a computer algorithm that could do it for her.

Another scientist might have a hypothesis that star-shaped cells are a biomarker for a specific disease, but he needs more examples to build a deep-learning algorithm that can help him probe this idea further. 

These are just a few kinds of problems that the BioImage Model Zoo, one of the core facets of the EU-funded AI4Life project, helps to solve. A crowd-sourced repository of pre-trained deep-learning models, the Zoo provides ready-to-use algorithms that can, for example, identify particular cellular or sub-cellular features – a process known as segmentation. 

The open-access tool is designed to be easy for scientists to use, making it an important time-saver. A crowd-sourced tool, the Zoo will grow as other scientists contribute algorithms tied to new model organisms, imaged with different microscopes, and featuring more diverse segmentation techniques.

BioImage Model Zoo began as a collaboration between EMBL Group Leader Anna Kreshuk,  Florian Jug, a computational biologist and Senior Group Leader at the Human Technopole Foundation in Milan, Italy, and Wei Ouyang, an assistant professor at KTH Royal Institute of Technology in Sweden and a SciLifeLab fellow. As scientific coordinators of AI4Life, Kreshuk and Jug work together with other partners across Europe, including the project’s main coordinator Euro-BioImaging, to bridge the gap between the computational and life sciences communities. 

The first phase of AI4Life, which began in 2022, ended in August 2025, so we caught up with Kreshuk and Jug to learn more about the project, its use, and what they hope its impact will be. 

Why was the BioImage Model Zoo created?

The point of BioImage Model Zoo is to enable the sharing of pre-trained deep-learning models for microscopy image analysis. The goal was also to share them in a standardised way, so that multiple user-facing tools could use the same models. This makes the models interoperable and also makes it easier for contributors to define metadata standards and assemble the models. Similarly, it makes it easy for end users to find the right model to further their research. 

Who can use the Zoo?

The Zoo can be accessed and used by everyone. We have, at the moment, a very strong focus on the bioimaging community. In the future, we might incorporate more medical imaging and even non-imaging technologies. 

How can AI help with segmentation questions?

We are inclined to think that computers, artificial intelligence, and deep learning tools have the potential to outperform humans in many tasks, but in segmentation, humans remain better. Humans can differentiate different cell types in photos, particularly those from different kinds of organisms, better than computers, and it has been challenging to build algorithms to automate the process. However, just as the adage goes, ‘practice makes perfect.’ Sharing repeated but variable microscopic imagery can help overcome this weakness for AI algorithms. That’s why human scientists have been ‘training’ the technology, so this work can become increasingly automated, saving time and expediting scientific findings. The BioImage Model Zoo is a way to share the fruits of those labours.

How is the BioImage Model Zoo situated within the wider AI4Life project? Were there any challenges the project faced?

The BioImage Model Zoo is basically a website that lets you browse and test run the deep learning models. AI4Life encompasses the Model Zoo, as well as all the technologies and the tools, the APIs, the model runner – everything that runs the Model Zoo, but can also be used and reused in many different contexts.

Every project of this size faces some challenges. It’s a consortium that is spread all over Europe, so there was the challenge of synchronising with each other and just working together efficiently. 

In the end, however, we have created something we actually like, and we also still like working with each other. I think having arrived at that point in such a big collaboration is not trivial.

Why is this such an important resource to have, especially at this time?

The state-of-the-art in computer vision is made possible because of the ongoing progress in AI methods. While new tools keep getting published, many advances remain inaccessible to life scientists who need them most. The Zoo provides pre-trained models in an easy-to-use form and thus reduces the gap between the amount of data that scientists can generate and the amount they can realistically process. We also place a lot of value on reproducibility. Together with our model metadata specification, we provide links to training data, code, and a full list of validation steps. More broadly, the Zoo also reflects EMBL’s commitment to democratising access to AI, ensuring that powerful methods are available to a wide community of researchers and not limited to AI specialists.

How does the BioImage Model Zoo tie into other AI work happening at EMBL?

We see the Zoo as a way for us to share the neural networks we train in the collaborative projects we run together with biologists. The Zoo is also strongly tied to the EMBL-EBI BioImage Archive, where Matthew Hartley’s BioImage Archive team is working on defining standards for training data annotations. 

What impact do you hope this project will have in the future?

It’s hard to predict whether your work will be impactful –  you may have some ideas, but other people might pick up on completely different things. Still, I hope all the effort we have put into making this model FAIR (findable, accessible, interoperable, reusable) will actually make an impact on the community. 

AI4Life has received funding from the European Union’s Horizon Europe research and innovation programme under grant agreement number 101057970.

Tags:
artificial intelligence, computational biology, deep learning, kreshuk, microscopy