Helmholtz Imaging at HAICON26: Connecting Imaging, AI and Scientific Discovery

Helmholtz Imaging was delighted to be part of HAICON26 in Munich, the conference of our partner platform Helmholtz AI. The event brought together researchers, AI experts, and research support communities from across Helmholtz and beyond to exchange ideas, discuss current developments in AI for science, and explore new opportunities for collaboration.

For Helmholtz Imaging, HAICON26 was a valuable opportunity to connect with peers working at the interface of imaging, AI, and scientific discovery.

Hands-on workshops for the information & data science community

The conference started with a busy Workshop & Tutorial Day, where Helmholtz Imaging hosted and contributed to three sessions covering key topics in AI-based image processing, foundation model adaptation, and benchmarking.

In the workshop “Novel Helmholtz Imaging Tools for AI image processing along the pipeline”, participants explored two tools developed within Helmholtz Imaging that address opposite ends of the AI pipeline in scientific imaging: PixelPatrol and the Helmholtz Model Zoo (HMZ).

PixelPatrol supports researchers in understanding and curating scientific image datasets before computationally intensive analysis or AI model training. The open-source tool systematically profiles imaging data and generates interactive dashboard reports with statistics and visualizations, helping users identify outliers, artifacts, metadata inconsistencies, and acquisition issues at an early stage.

The Helmholtz Model Zoo, in turn, supports a later stage in the AI pipeline: sharing, deploying, and running inference with trained deep learning models. The platform enables researchers across the Helmholtz Association to make models accessible through a web interface and REST API, while integrating with Helmholtz infrastructure such as Helmholtz ID, dCache, and DESY’s HPC resources. During the workshop, participants learned how to prepare and deploy models and gained insight into how the HMZ can support AI-based research workflows.

Helmholtz Imaging also hosted “A Practical Tour of PEFT & Co.”, a hands-on tutorial on adapting foundation models. Starting from a fixed pretrained checkpoint, participants explored methods such as linear probing, LoRA, prompt-based approaches, and related techniques. The session provided practical intuition for different adaptation strategies and showed how they affect model parameters, representations, and performance.

A third contribution focused on the broader question of benchmarking in AI. In the session “Current status of the benchmarking field: lessons learned from the first half of the UNLOCK initiative”, participants discussed best practices for establishing benchmarks across emerging research fields. The session brought together perspectives on benchmark design, AI safety, energy consumption, scientific impact, and community building, highlighting the growing importance of meaningful benchmarks for evaluating AI models in science.

Sharing tools, support and scientific images

During the main conference days, the Helmholtz Imaging booth served as a meeting point for researchers engaging in discussions on image data analysis, AI-based imaging workflows, and the support and resources Helmholtz Imaging offers to the community. The booth facilitated numerous conversations and exchanges with researchers across disciplines. Visitors could learn more about our consulting and support opportunities, training formats, funding options, and solutions for imaging-related research.

A particular highlight were the highly popular posters featuring award-winning images from the Helmholtz Imaging Best Scientific Image Contest 2025, showcasing the diversity and beauty of scientific imaging across research fields.

Another highlight was Hans Werners’ poster spotlight presentation on the Helmholtz Model Zoo. It was great to see his poster selected for a spotlight talk, giving the project additional visibility among the HAICON26 community and opening up further conversations about model sharing, deployment, and inference for AI-based research.

Supporting imaging research across Helmholtz

HAICON26 once again showed how strongly imaging, AI, data science, and domain research are connected. For Helmholtz Imaging, the conference served not only as a platform to present our work, but also to listen to researchers’ needs, discuss concrete challenges, and connect with colleagues from our partner platforms across the Helmholtz Information & Data Science Framework.

We thank Helmholtz AI and everyone involved in organizing HAICON26 for creating such an inspiring environment for scientific exchange, collaboration, and new ideas.

Thanks to Ella Bahry, Philipp Heuser, Annika Reinke, Thomas Roßberg, Deborah Schmidt, Jeremias Traub, Dasha Trofimova Hans Werners from Helmholtz Imaging, and Florian Grötsch from the Helmholtz Head Office for hosting and contributing to the Workshop & Tutorial Day!

Are you a Helmholtz researcher? Contact us to get professional imaging support at no cost.

Interested in PixelPatrol? Starting August 13, you can register for the two-part online workshop on September 10 & 17!

Explore the Helmholtz Model Zoo: hmz-hub.desy.de