T01: Kaapana Demo - HandsOn
DKFZ Heidelberg, Prof. Dr. Klaus Maier-Hein, Hanno Gao
Kaapana is an open-source medical image research platform for clinicians, data scientists and developers. It enables large scale machine learning research on real world data by bringing state of the art AI tools in clinics and research centers. Kaapana can support both central and decentralized use cases which enables interdisciplinary studies that can lead to significant advancements in medical research and patient care.
In this tutorial, we will first start with an overview of the Kaapana platform and its capabilities. We will explain and demonstrate the features and the various components of Kaapana. Then we will proceed with a hands on demo using Kaapana instances we provide to users. In the demo, a complete end-to-end analysis workflow of the platform is presented with a focus on creating and curating datasets, processing medical images, using machine learning models in the platform and exploring the ways for extending the platform in different use cases. Finally, the audience has the opportunity to explore the system freely and discuss their use cases.
T02: Foundation Models for Few‑Shot Medical Image Analysis
MEVIS Lübeck, Till Nicke, Raphael Schäfer, Johannes Lotz
Foundation models are rapidly reshaping medical imaging research by providing robust, transferable feature spaces that reduce annotation requirements while enabling rapid model development for clinically relevant tasks. This hands‑on tutorial uses the MedicalMultitaskModeling (MMM) foundation model[1][2][3] as an example and shows how to adapt pretrained backbones for (i) few-shot classification, (ii) few-shot segmentation, and (iii) multiple instance learning (MIL) on 2D, 3D, and gigapixel pathology data.
We will begin with a short conceptual primer: architecture recap, adaptation strategies, and considerations on supervised vs. self-supervised foundation model training. The core of the session is interactive and notebook-driven: preparing example datasets, constructing task heads, evaluating generalization, and making the pipeline work on real data.
Participants are encouraged to bring their own datasets for the second half of the session. Please make sure that the data can be loaded into a Jupyter notebook at the start of the session.
Prior deep learning experience is helpful; no prior exposure to foundation models is required.
[1]: Schäfer et al., "Overcoming data scarcity in biomedical imaging with a foundational multi-task model." Nat Comput Sci (2024). https://doi.org/10.1038/s43588-024-00662-z
[2]: Code: https://github.com/FraunhoferMEVIS/MedicalMultitaskModeling
[3]: Also available pre-trained for pathology, see Nicke et al. "Tissue concepts: Supervised foundation models in computational pathology." Comp. Biol. Med. (2025). [doi:10.1016/j.compbiomed.2024.109621](https://doi.org/10.1016/j.compbiomed.2024.109621)