T01: Validating AI for Medical Image Analysis
Dr. Annika Reinke, Dr. Evangelia Christodoulou (Div. Intelligent Medical Systems, German Cancer Research Center (DKFZ))
This tutorial provides an introduction to the validation of AI methods in medical image analysis. Reliable validation is essential to decide whether algorithms are working as intended beyond standard leaderboards, and even whether they have a potential for clinical translation. It ensures that model performance estimates are meaningful, reproducible, and reflecting the underlying research questions. The course introduces key principles and practical techniques for validating AI models. Participants will learn how to select and interpret appropriate performance metrics, quantify model performance uncertainty, and assess the robustness of rankings in algorithm benchmarking. The tutorial combines theoretical lectures with guided hands-on exercises.
Preliminary schedule:
09:00 - 10:30 | Performance Metrics
- Introduction to performance metrics for classification and segmentation tasks
- Common pitfalls and limitations of performance metrics
- Selecting appropriate metrics based on the research question
- Hands-on exercises with practical use cases
10:30-11:00 | Coffee break
11:00 - 12:30 | Uncertainty and Model Performance Reporting
- Reporting variability and uncertainty in model performance estimates
- Confidence interval calculation and model comparisons
- Assessing the validity of model outperformance claims in scientific studies
- Good practices for reporting AI model performance
- Hands-on exercises with practical use cases
T02: Bias and Fairness in Medical Image Analysis
Dr. Eike Petersen (Medizinische Hochschule Hannover, Fraunhofer-Institut für Digitale Medizin MEVIS), M.Sc. Chiara Tappermann (Fraunhofer-Institut für Digitale Medizin MEVIS)
As AI transforms healthcare, ensuring fairness in medical imaging is crucial to prevent health disparities. This tutorial provides a comprehensive introduction to algorithmic fairness, covering key concepts, sources of bias, and practical mitigation strategies. Through expert-led lectures and hands-on demonstrations, participants will explore real-world case studies, fairness evaluation techniques, and tools for bias assessment and mitigation. By the end, attendees will be equipped to integrate fairness principles into their machine learning workflows, fostering medical AI systems that work for every patient.
Preliminary schedule:
14:00 - 15:30 | Causes of Bias
- Under-representation, differences in task difficulty, shortcut learning, label bias
- Disentangling different sources of bias
- Quantifying subgroup differences in model performance
- Hands-on exercises with practical use cases
15:30 - 16:00 | Coffee break
16:00 - 17:30 | Bias mitigation
- General robustness improvements, synthetic data augmentation, model selection, loss-based methods, calibration
- Trade-offs (there are few that matter)
- Conclusions & recommendations
- Hands-on exercises with practical use cases
T03: Vibe Coding and Software Engineering in the Age of AI
Prof. Dr. Andreas Maier (Universität Erlangen), Prof. Dr. Mattias Heinrich (Universität zu Lübeck)
Coding agents now write a large share of research software, including in medical image computing. Vibe coding, building software by describing intent to an AI agent instead of typing code line by line, makes prototypes fast, but it does not make them correct, maintainable or safe. This tutorial shows that the classical methods of software engineering are what make agent-assisted development reliable: requirements and specifications that agents can execute, architectural design, test-driven development and quality gates, version control, and explicit hand-over points between humans and agents. The first part introduces these methods following the textbook "Vibe Coding: Software Engineering in the Age of AI" (Springer LNCS, 2027). The second part is hands-on: participants work at several round tables on typical medical-imaging tasks.
Preliminary schedule:
14:00 - 15:30 | From prompt to process
- Vibe coding: origins, current practice and limits
- Requirements and specifications that agents can execute
- Architecture, agentic design patterns and tool interfaces
- Testing, quality gates and human oversight
15:30 - 16:00 | Coffee break
16:00 - 17:30 | Hands-on round tables
- Building a tool from a single prompt, and what breaks
- Porting legacy MATLAB code to Python with tests
- Reliable deployment and testing of a segmentation tool
- Project hand-over via README and specification
- Joint discussion and wrap-up