AI for biology and medicine.

Six directions, one question: how do we turn biomedical measurements into something that can be trusted, quantified and deployed?

01

Representation learning & neural representations

What does a network actually encode? We study the internal representations of neural networks and use continuous, implicit representations to describe biological and physical data.

We could show that a single latent channel is sufficient for a segmentation task by using implicit neural representations of microscopy volumes and laryngeal aerodynamics, and by developing methods to compare what two networks have learned.

All 7 publications on this topic

02

Multimodal learning & computational oncology

Clinical questions rarely come with a single modality. We combine histopathology, imaging, blood-based immunophenotyping and structured clinical records into models that can be interrogated rather than merely scored.

Our work in head & neck oncology is a joint effort with national and international clinical partners.

All 5 publications on this topic

03

AutoML, efficient & embedded AI

A model that cannot run where the data is measured is of limited use. We develop evolutionary neural architecture search with hardware in the loop: every candidate is deployed to the target microcontroller and measured for accuracy, memory, latency and energy.

We also study the statistics of NAS benchmarks themselves, and how mixed precision changes the cost of training and inference.

All 17 publications on this topic

04

Generative AI & foundation models

Large language models, text-to-speech systems and segmentation foundation models are useful — but how far do they actually generalise?

We test them against clinical reality: unstructured electronic health records of dysphagic patients, LLM-generated segmentation baselines, prompt generation for segmentation foundation models, and synthetic sustained phonation for speech therapy.

All 8 publications on this topic

05

Biomedical imaging & quantitative analysis

Turning images and videos into numbers a clinician or biologist can act on: glottis segmentation and glottal midline detection in high-speed videoendoscopy, dendritic spine quantification in light microscopy, swallowing studies, coronary segmentation, and 3D reconstruction from monoscopic endoscopy.

A recurring theme is quality: not only how good a segmentation is on average, but whether we can predict when it fails.

All 27 publications on this topic

06

Biomedical signals & wearable health

One-dimensional biosignals — audio, accelerometry, aerodynamics — carry a great deal of clinical information if acquired and analyzed well.

With collaborators at McGill University, we developed a wearable airway symptom detection; we build the annotation tools that make such longitudinal data usable, and open platforms for acquiring laryngeal high-speed video and audio at the same time.

All 12 publications on this topic

An open, ambitious and internationally oriented group of researchers and students.

A few pieces of work that show what we do. The complete list is on the publications page.

• HANCOCK

  A multimodal dataset for precision oncology in head and neck cancer

  A multimodal dataset for precision oncology in head and neck cancer, built with clinical partners at Universitätsklinikum Erlangen.

  • Dörrich et al., Nat. Commun. 2025 ·DOI: 10.1038/s41467-025-62386-6

  • DOI: 10.1038/s41467-025-62386-6
  • Project

• EdgeVolution

  EdgeVolution: democratizing multi-objective neural architecture search and end-to-end deployment on microcontrollers

  Democratizing multi-objective neural architecture search and end-to-end deployment on microcontrollers.

  • Groh et al., Commun. Eng. 2026 ·DOI: 10.1038/s44172-026-00708-2

  • DOI: 10.1038/s44172-026-00708-2
  • GitHub
  • Documentation

• DeepD3

  DeepD3, an open framework for automated quantification of dendritic spines

  An open, deep-learning framework for automated quantification of dendritic spines, trained on data from multiple experts and experimental conditions.

  • Fernholz et al., PLoS Comput. Biol. 2024 ·DOI: 10.1371/journal.pcbi.1011774

  • DOI: 10.1371/journal.pcbi.1011774
  • GitHub
  • Data
  • Project

• Impact of Mixed Precision

  Impact of Mixed Precision Techniques on Training and Inference Efficiency of Deep Neural Networks

  How mixed-precision techniques affect training and inference efficiency of encoder–decoder deep neural networks in biomedical image segmentation.

  • Dörrich et al., IEEE Access 2023 ·DOI: 10.1109/access.2023.3284388

  • DOI: 10.1109/access.2023.3284388

• OpenHSV

  OpenHSV: An open platform for laryngeal high-speed videoendoscopy

  An award-winning, open hard- and software platform for laryngeal high-speed videoendoscopy with real-time, deep-learning-powered analysis.

  • Kist et al., Sci. Rep. 2021 ·DOI: 10.1038/s41598-021-93149-0

  • DOI: 10.1038/s41598-021-93149-0
  • GitHub
  • Docs
  • Award

Interested in collaborating, or in a thesis project? Get in touch.