AI analysis for bioimages—what’s missing?

Lack of incentives and low adoption of metadata standards limit AI’s potential for bioimage analysis: community initiative offers solutions.

AI can detect subtle patterns in millions of microscopic images or compare one patient’s scan to thousands of others in seconds. However, several technical and cultural barriers, affecting metadata, incentives, formats and accessibility, stand in the way.

Matthew Hartley (MH), BioImage Archive team leader, and Teresa Zulueta Coarasa (TZC), bioinformatician, explain how community recommendations could help.

What is metadata and why is it important in bioimaging today?

MH: Metadata is essentially the context around images and annotations. It explains what we see and how the image was captioned: when, where and under what conditions.

For AI training, metadata is what makes a data set interpretable, reusable, and valuable beyond the lab that generated it. The challenge is that different labs record metadata in different ways, making it difficult for others to reuse their data. Agreeing on standards helps everyone speak the same language.

Your article offers guidelines for improving the reuse of bioimages for AI applications. How did you develop the guidelines?

MH: The idea came from a workshop we organized in 2023 as part of the AI4Life project, which brought together 45 participants from across the community, including data producers, AI scientists and bioimage analysts. We identified four groups of recommendations, summarized as MIFA, which stands for Metadata, Incentives, Formats and Accessibility. The document describing our recommendations was published in Natural methods.

How can we improve the reuse of bioimages for AI applications?

TZC: For metadata, we are proposing a new standard focused on image annotations, building on metadata standards such as REMBI, which we developed in 2021. This is important because annotations, such as segmentation masks, are an essential part of these datasets, and scientists need to understand what they are and how they were generated.

In terms of incentives, we recognize that it takes significant time and effort to produce annotated datasets. Currently, there are few incentives to encourage laboratories to produce metadata or share their images in open repositories such as BioImage Archive. This needs to change, and it will require funders, journals, data archives, and the bioimaging community working together on strong incentives.

Microscopy equipment uses a range of formats, depending on the manufacturer. We need to ensure common and interoperable data formats so that laboratories can easily share and reuse images.

These are not just abstract ideals; these are practical recommendations developed with input from the people producing the data and those who need it for AI training.

What impact could widespread adoption of MIFA have?

TZC: Life scientists spend months generating stunning, painstakingly annotated datasets, which AI developers often struggle to interpret. Bringing the two sides together could help bridge this divide. With standardized metadata, AI models trained on one dataset could be validated on others, improving reproducibility. This would unlock the ability to compare models, reproduce results, and accelerate discovery. In short, this makes bioimaging AI scalable.

MH: There is real momentum in the community for this to happen. For example, journals recommend that datasets be deposited in public archives. This in itself inspires researchers to structure and share their data more thoughtfully. We believe that if data producers adopt these guidelines, we will have a virtuous cycle: better datasets, better AI, and better science. A good first step is to read our recommendations and see if you can incorporate them into your workflows.

Community Voices

Below are some comments from colleagues who have successfully used the MIFA guidelines in their work.

“Using the MIFA Guidelines and BioImage Archive, I was able to easily locate new datasets suitable for a project that studied the transferability and appropriate selection of pre-trained image segmentation models. Access to well-structured metadata made working with multiple datasets for training and evaluating neural networks much simpler and faster. Our results are available here,” says Joshua Talks, Ph.D. EMBL. student.

“We hope that by sharing our images and annotations in accordance with MIFA guidelines, we will maximize the reusability of our datasets to train new AI tools and increase the visibility of the AI ​​tool we trained using these datasets. Our ReSCU-Nets are recurrent neural networks that integrate segmentation and tracking of cellular, subcellular and supracellular results from multidimensional confocal microscopy. sequences,” explains Rodrigo Fernandez-Gonzalez, professor at the University of Toronto.