Trustworthy and Robust Anomaly-Detection in Clinical Environments(TRACE)

• Funding: MERCUR / University Alliance Ruhr
• Duration: 12 months
• Start: January 2026
• Partners: TU Dortmund University / Lamarr Institute, Ruhr University Bochum, University Medicine Essen / Institute for AI in Medicine (IKIM), and associated research partners.

TRACE aims to develop robust and interpretable methods for detecting anomalies in 3D medical imaging data. The project investigates how unsupervised anomaly detection can be combined with causal concept-based models to make predictions more understandable and clinically meaningful. A further objective is to improve robustness and data provenance through security-related methods such as digital watermarking. The developed approaches will be evaluated using real-world clinical imaging data in collaboration with medical experts.

• WP1 – Robust 3D Anomaly Detection

Development and evaluation of unsupervised methods for detecting unusual or pathological patterns in 3D medical images, with a focus on brain MRI.

• WP2 – Causal Concept Bottlenecks

Development of interpretable machine-learning models that connect model predictions with clinically meaningful concepts and investigate causal relationships between them.

• WP3 – Clinical Validation

Evaluation of the developed methods on real-world clinical imaging data in collaboration with medical experts, with particular attention to reliability, interpretability, and clinical relevance.

• WP4 – Coordination and Trustworthy Clinical AI Collaboration

Coordination of the interdisciplinary collaboration between the participating institutions and integration of expertise in machine learning, clinical AI, and data security to support trustworthy AI research in healthcare.