Research

Research Themes

Three themes connect everything we do — publications, funded projects, datasets, code, courses and events are all tagged with them.

Background

Expertise should travel to patients, not the other way around. Our work began with that conviction and still follows it: we build medical imaging AI that brings specialist-level insight to the places, people and devices that have the least access to it.

Where it started: telemedicine in Gaza

Long before the lab existed, the question was concrete. Between 2007 and 2014, Shadi Albarqouni worked on telemedicine in Palestine. In a health system where checkpoints, permits and the separation of Gaza and the West Bank kept patients and physicians apart, he helped establish the first stable video link in 2008 between Caritas Baby Hospital in Bethlehem, Al-Shifa Hospital in Gaza and Al-Quds Open University. He then proposed a telemedicine centre for Gaza that would go beyond videoconferencing: it would connect pathology, endoscopy and EEG equipment to a secure network built on the patient’s health record.

Looking back in 2014 at a decade of Palestinian telemedicine pilots, he found that they failed for the same four reasons: infrastructure was missing, physicians were not engaged, administration and governance were not aligned, and the projects were not sustainable once the funding ended. Connecting hospitals was not enough. What was scarce was not bandwidth but expertise, and the trust, governance and local capacity needed to use it.

From connecting hospitals to carrying expertise

Medical image computing offered a new answer to the same problem. A model that has learned from expert annotations carries a piece of that expertise to wherever it runs: a district hospital, a smartphone, a portable scanner. Each of the four barriers of telemedicine returned in a new form, as a research question for AI:

  • Too few specialists → expert labels are scarce and expensive. We learn from crowds, weak labels, unlabeled data and, now, foundation models. (Computational imaging)
  • Heterogeneous, low-cost infrastructure → models break on new scanners, stains, sites and devices. We build robust, adaptable models that report their own uncertainty, and affordable imaging at the point of care. (Computational imaging, Responsible AI)
  • Physician engagement → clinicians cannot trust a black box that may be biased. We work on uncertainty, explainability and fairness, and helped write the FUTURE-AI guidelines. (Responsible AI)
  • Governance and privacy → patient data cannot leave the hospital. Federated learning moves the model instead of the data. (Federated learning)

Sustainability, the fourth barrier, is the reason for everything else we do: open benchmarks and code, and training the next generation of researchers where they are needed most.

A decade of answers

Each answer exposed the next question. Labels were scarce, so we learned from crowds and from what healthy anatomy looks like, without labels. Once data came from many scanners and stains, the distributions shifted, so we learned to normalise, translate and adapt. Clinicians needed to know when to trust a prediction, so we made models estimate their uncertainty and learn representations that are fair by design. The most valuable data could not leave hospitals, so we brought the model to the data and made federated learning work with non-identical sites and few labels. Large-scale self-supervision now gives us foundation models that need only a handful of labels, and we are learning how they fail across populations and institutions.

Where we are going: AI at the point of care

The three threads now converge on the problem telemedicine set out to solve. Foundation models reduce how much expert labelling a new clinic needs. Federated learning lets those models be adapted across hospitals, including within the German Medical Informatics Initiative (OMI), without moving patient data. Responsible and affordable AI makes sure they work for the patients who are usually left out. In practice this means:

  • portable ultra-low-field MRI for newborns, enhanced to diagnostic quality (MRIQT);
  • smartphone and fundus-camera screening for eyelid tumours and retinal disease (eyelid tumours, affordable retina);
  • auditing foundation models for bias and shortcuts before they reach screening programmes (fair mammography);
  • building research capacity across the world through SANAD, BILD and EEDA.

The question from Gaza remains the same: how do we bring the best possible care to people who cannot reach it? What has changed is that the expertise itself can now travel.

More on this journey: the DAAD’s Change by Exchange profile and the Tagesspiegel Background portrait.

Selected publications

  • Albarqouni et al., AggNet: deep learning from crowds for mitosis detection, IEEE TMI 2016 — paper
  • Rieke et al., The future of digital health with federated learning, npj Digital Medicine 2020 — paper
  • Sarhan et al., Fairness by learning orthogonal disentangled representations, ECCV 2020 — paper
  • Baur et al., Autoencoders for unsupervised anomaly segmentation in brain MR images, Medical Image Analysis 2021 — paper
  • Bercea et al., Federated disentangled representation learning for unsupervised brain anomaly detection, Nature Machine Intelligence 2022 — paper
  • Ogier du Terrail et al., FLamby: datasets and benchmarks for cross-silo federated learning, NeurIPS 2022 — paper
  • Nguyen et al., LVM-Med: learning large-scale self-supervised vision models for medical imaging, NeurIPS 2023 — paper
  • Lekadir et al., FUTURE-AI: international consensus guideline for trustworthy and deployable AI in healthcare, BMJ 2025 — paper
  • Germani et al., When multi-institutional dataset aggregation masks shortcut-like behavior in foundation models for medical imaging, Medical Image Analysis 2026 — paper

Research areas

The clinical and application areas of our computational medical imaging research — from pathology, haematology and radiology to ophthalmology, surgery and point-of-care imaging.