The rapid development of artificial intelligence (AI) has revolutionized traditional microscopy imaging in many ways. Examples include ubiquitous tasks such as image denoising and restoration, image segmentation, object detection and classification, virtual staining, and more. While advances in all of these areas are being published in high-profile journals, this development has not yet been matched with standards and/or a community consensus on how to properly report results, conduct evaluations, and deal with ethical issues that come with technical AI advances. Hence, in this in-depth Dagstuhl Seminar, our discussions will be centered around two central topics, (i) the ethical use of AI in bioimaging and (ii) performance evaluation and reporting when using bioimage AI methods.

Who is invited to the seminar? We aim at bringing together people representing three types of expertise that all have an interest in finding solutions to the topics introduced above, namely, (i) AI researchers and bioimage AI method developers, (ii) users of AI facilitated methods and tools (life scientists, biologists, microscopists), and (iii) editors of journals handling bioimaging AI manuscripts.

What is the format of the seminar? Both main discussion topics will follow an “introduction and related works → brainstorming → outlook summary” format. We hope this seminar will serve as a kick-off meeting to initiate further community efforts on these and related topics.

For the topic of “ethical use of AI in bioimaging”, we first would like to raise the awareness that generative AI models have become so sophisticated that AI-generated bioimages can be indistinguishable from experimentally acquired images. Technically, we will discuss the state-of-the-art of image generation methods and how related AI fields deal with generated fake images (e.g., deepfake detection or image forgery detection). Then, the brainstorming session will aim to identify potential challenges in the bioimaging context and prospective solutions, followed, in the outlook session, by a discussion of potential research synergies and how community-driven guidelines on the ethical use of AI can be established.

For the topic of “AI performance evaluation and reporting”, we will first exchange insights about the status quo in bioimaging (e.g., common factors affecting microscopy image qualities, related biological applications, etc.) and then introduce the status quo on relevant topics in AI (e.g., domain shift, out-of-distribution detection, generalization gap, etc.). This will be followed up by introducing “application-appropriate validation” in bioimaging and possible evaluation strategies in related AI fields, such as medical imaging. The subsequent brainstorming session will begin with “case studies” from participants in order to understand and dissect the complexity and diversity of bioimaging evaluation problems, and then discuss potential solutions. Finally, the outlook session will focus on identifying future research synergies to address bottlenecks in bioimaging AI performance and the potential of establishing community standards for bioimaging AI performance evaluation and reporting.

What to expect at the end of the seminar? We expect all participants will gain a comprehensive overview of the demands, issues, potential solutions, and missing links to enable ethical, trustworthy, and publishable bioimaging AI methods, tools, and applications. We also aim to have this seminar kick off further meetings on these topics and create collaborations in the bioimaging AI community to follow up on ideas discussed during this event. Our hope is to follow up with participants online to transform the discussions in this seminar to an eventual joint community-driven perspective paper on these topics.

Creative Commons BY 4.0

Jianxu Chen, Florian Jug, and Susanne Rafelski