A reasonably recent version of my CV can be downloaded here [pdf].
A more up to date list of publications can be found on Google Scholar.
Research Overview
Our research focuses on advancing intelligent algorithms for multi-modal healthcare. We develop new generative and discriminative machine learning methods to enhance diagnostic decision-making, provide real-time guidance to human operators during diagnostic procedures, and explore emerging paradigms such as normative learning to ensure the safety of machine learning and bring its applications to the front lines of patient care.
Research Streams
๐ฅ Real-Time Clinical Guidance & Human-AI Collaboration
Can we democratize rare healthcare expertise through Machine Learning, providing guidance in real-time applications and second reader expertise in retrospective analysis? This stream focuses on developing AI systems that work alongside clinicians in real-time diagnostic procedures.
Key projects & spinouts:
- CHARMS (ERC Proof of Concept 101246053): Normative representation learning for ultrasound-based prenatal screening and earlier detection of congenital heart anomalies
- PRETUS Platform: Real-time ultrasound imaging and guidance systems for fetal and cardiac applications
- ThinkSono: Non-invasive deep vein thrombosis (DVT) diagnosis from ultrasound with machine learning guidance for novice users
- Fraiya: AI-powered prenatal ultrasound screening for fetal anomaly detection in real-time
- KidneyCaliper: Histopathology analysis platform for kidney biopsy assessment
- Fetal Echocardiography: AI-assisted real-time cardiac anomaly detection in fetal ultrasound
- Active Learning & Human-in-the-Loop: Optimizing human-machine collaboration for medical image analysis
๐ง Normative Learning & Population Health
Can we develop normative learning from large populations, integrating imaging, patient records and -omics, leading to data analysis that mimics human decision making? This research explores learning what is "normal" across large populations to better identify anomalies and diseases. Our ERC Consolidator Grant MIA-NORMAL focuses on anomaly-aware, open-world normative modeling.
Key projects:
- Unsupervised Anomaly Detection: Learning normal appearance for detecting rare conditions like Hypoplastic Left Heart Syndrome
- MOOD Benchmark: Public benchmark for out-of-distribution detection and localization on medical images
- Fetal & Neonatal Development: Normative models of cortical surface development and fetal organ growth
- RACOON Network: Collaboration with German university hospitals (UKER and 37 others) creating nationwide infrastructure for structured radiological data collection and AI-based medical assistance systems
- Federated Learning: Privacy-preserving multi-site disease detection across international centers
๐ Explainable AI & Causal Reasoning
Can we provide human interpretability and effective human-machine teamwork of machine decision making to support the 'right for explanation' in healthcare? This stream develops methods for understanding and explaining AI decisions through causal reasoning and multi-agent systems.
Key projects:
- ORACLE (ERC Proof of Concept 101326871): Out-of-distribution-aware reasoning agents for autonomous, explainable clinical learning and discovery
- Bayesian Decoding Games (NeurIPS 2025): Multi-agent debate systems using game theory to prevent collusion and enable smaller models to outperform larger ones through strategic consensus
- Hypothesis Hunting: Agentic ML systems that autonomously analyze massive multimodal datasets to derive and verify hypotheses about correlations in medical data
- Multi-Agent Cardiac Phenotyping (MICCAI 2025): Autonomous agents for cardiac data analysis combining imaging, omics, and clinical records
- Counterfactual Analysis: Deep twin networks and video generation for "what-if" scenarios in medical decision-making
- Attention Mechanisms: Learning to leverage salient regions in medical images with interpretable attention gates
- Causality in Medical AI: Review and methods for causal learning in medical imaging
- Uncertainty Quantification: Bayesian and stochastic methods for reliable predictions
๐ฏ Advanced Image Analysis & Reconstruction
Developing cutting-edge methods for medical image segmentation, reconstruction, and enhancement across multiple modalities and anatomies. This includes novel vision-language models for 3D medical imaging.
Key projects:
- BTB3D (NeurIPS 2025): Better Tokens for Better 3D - advancing vision-language modeling in 3D medical imaging with frequency-aware volumetric tokens for radiology report generation and text-to-3D synthesis
- Motion Compensation: Fetal MRI reconstruction with slice-to-volume registration and learning-based approaches
- Topology-Aware Segmentation: Cortical surface reconstruction with neural ODEs and topology-constrained learning
- Multi-Modal Synthesis: Generative models for echocardiogram synthesis and data augmentation
- Super-Resolution: Motion correction and super-resolution for abdominal MRI
๐ Data Augmentation & Synthetic Data
Creating realistic synthetic medical data to address data scarcity and improve model robustness.
Key projects:
- Diffusion Models: Feature-conditioned video diffusion for echocardiogram synthesis
- Lesion-Level Augmentation: LesionMix for pathology image segmentation
- Natural Synthetic Anomalies: Self-supervised methods for anomaly detection
- Foreign Patch Interpolation: Detecting outliers with synthetic perturbations
Publication Metrics
In Computer Science, peer-reviewed full papers at leading, top-ranked conferences are as important and sometimes more selective as journal publications. See e.g., http://bit.ly/2KzvvyZ or http://bit.ly/2ptl1tF for a discussion of this topic.
My research about human-centred AI in healthcare is at the interface of Computer Science, Medical Image Analysis, Machine Learning and Clinical Science. Thus, both journal and conference publications count equally much, and I have a good record in both categories.
The leading journals in my area are IEEE Transactions on Medical Imaging (IEEE Trans Med Imag), Elsevier Medical Image Analysis (Med Image Anal) and from a machine learning perspective the Journal of Machine Learning Research (JMLR). The leading conference is the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI) and its associated workshops.
Recent Highlights
NeurIPS 2026
GRASP has been accepted at NeurIPS 2026:
-
GRASP: Guided Residual Adapters with Sample-wise Partitioning
Cluster-specific residual adapters reduce gradient conflicts in long-tailed diffusion training, improving rare-class image quality, diversity, and downstream classification.
๐ NeurIPS | ๐ arXiv
MICCAI 2026
Our group has six papers at MICCAI 2026, spanning out-of-distribution localisation, volumetric generation, augmented-reality guidance, agentic phenotyping, rare-condition learning, and medical latent spaces:
-
WALDO: Wasserstein-Aligned Localisation for VLM-Based Distributional OOD Detection in Medical Imaging
Uses healthy reference distributions and vision-language models to localise rare abnormalities without labelled training data.
๐ Paper | ๐ arXiv | ๐ป Code
-
CTTok: Voxel-Abulary for Autoregressive 3D CT Volume Generation with Large Language Models
Extends a language model with anatomical tokens for fast, text-conditioned 3D CT generation without dedicated cross-attention.
๐ Paper | ๐ป Code
-
Multi-camera AR Guidance: Surgical Instrument Handling and Assembly
Combines marker-free multi-camera pose estimation with in-situ AR guidance, reducing task completion time and perceived workload for scrub nurses.
๐ Paper | ๐ arXiv
-
CPAgents: Agentic Composite Phenotype Generation for Cardiac Disease Association
Uses constrained multi-agent generation and verification to discover interpretable composite cardiac phenotypes and disease associations.
๐ Paper | ๐ arXiv | ๐ป Code
-
Fibers of Asymmetric Similarity: A Framework for Clinical and Imaging Data
Introduces asymmetric similarity fields and prototype learning for rare, severely imbalanced clinical and imaging tasks.
๐ Paper | ๐ป Code
-
The Learnability Gap in Medical Latent Diffusion
Shows that pretrained autoencoders can preserve diagnostic information while arranging latent representations in ways that remain difficult for classifiers to learn.
๐ Paper | ๐ arXiv | ๐ป Code
Acknowledgments
High-performance computing resources were provided by the Erlangen National High Performance Computing Center (NHR@FAU) at Friedrich-Alexander-Universitรคt Erlangen-Nรผrnberg (FAU), under the NHR projects b143dc and b180dc. NHR is funded by federal and Bavarian state authorities, and NHR@FAU hardware is partially funded by the German Research Foundation (DFG) โ 440719683.
We acknowledge the use of Isambard-AI National AI Research Resource (AIRR). Isambard-AI is operated by the University of Bristol and is funded by the UK Government's DSIT via UKRI; and the Science and Technology Facilities Council [ST/AIRR/I-A-I/1023].
Additional support was received from the ERC projects MIA-NORMAL 101083647, CHARMS 101246053, and ORACLE 101326871; DFG 512819079 and 513220538; and the state of Bavaria (HTA). Some researchers are supported by the JADS programme and the UKRI Centre for Doctoral Training in AI for Healthcare (EP/S023283/1).
Publications
This list is automatically generated from my BibTeX database based on Google Scholar.