EXA MODE · Project

AI That Analyzes Massive Medical Image Archives Without Manual Labeling

healthTestedTRL 5

Hospitals produce mountains of medical images and reports every day — far more than any human team could ever sort through. Imagine if a smart assistant could read both the images and the written reports together, learn what matters without someone hand-labeling every single slide, and then help doctors spot patterns faster. That's what EXA MODE built: AI tools that learn from weakly labeled medical data — especially digital pathology slides — so hospitals and companies can unlock the knowledge buried in their archives without needing armies of annotators.

By the numbers

>2000 exabytes

Healthcare data production projected by 2020

EUR 4,333,281

EU funding for development

Consortium partners

Countries represented

Total project deliverables

Industry partners in consortium

The business problem

What needed solving

Hospitals and medical companies sit on vast archives of images and reports — over 2000 exabytes of healthcare data produced annually — but extracting useful knowledge from this data traditionally requires expensive, slow manual labeling by medical experts. Most AI tools demand thousands of hand-annotated examples before they can learn anything, making large-scale medical image analysis economically impractical for many organizations.

The solution

What was built

The project delivered 23 deliverables including a dynamic visual analytics prototype that lets experts interact with AI learning processes and understand intermediate results — field-tested and refined with real users. Core outputs include weakly supervised deep learning methods, a semantic middleware for image compression, segmentation, and classification, and decision support prototypes aligned to medical ontologies.

Audience

Who needs this

Digital pathology platform companies needing to train AI without massive manual annotationHospital groups wanting to make their medical image archives searchable and actionableHealth IT vendors building clinical decision support systemsPharma companies analyzing heterogeneous image data across multi-site clinical trialsMedical AI startups looking for validated weakly supervised learning methods

Business applications

Who can put this to work

Digital Pathology

mid-size

Target: Companies developing or operating digital pathology platforms

If you are a digital pathology company struggling with the cost and speed of annotating thousands of tissue slides for AI training — this project developed weakly supervised deep learning methods that extract diagnostic knowledge from existing reports and images together, cutting the need for manual expert labeling. The tools were tested with hospital partners across a 10-partner consortium spanning 5 countries.

Hospital IT & Health Informatics

enterprise

Target: Hospital groups or health IT vendors managing large medical image archives

If you are a hospital IT department drowning in over 2000 exabytes of healthcare data produced annually and struggling to make that data searchable and useful — this project built a semantic middleware that compresses, segments, and classifies medical images automatically. The visual analytics prototype was field-tested with expert users and refined based on their feedback.

Pharmaceutical & Life Sciences

enterprise

Target: Pharma companies running large-scale image-based drug discovery or clinical trials

If you are a pharma R&D team that needs to analyze heterogeneous image datasets from multiple clinical sites without standardized labels — this project created multimodal knowledge discovery tools that fuse text reports with image features and align them to medical ontologies. The consortium included 4 industry partners ensuring real-world applicability.

Frequently asked

Quick answers

What would it cost to license or adopt this technology?

The project was funded as a Research and Innovation Action (RIA) with EUR 4,333,281 in EU funding. Licensing terms would need to be negotiated directly with the consortium partners. Based on available project data, no commercial pricing model has been publicly disclosed.

Can this work at industrial scale with real hospital data volumes?

The project was specifically designed for extreme-scale analytics — their objective explicitly targets exascale data volumes. Healthcare data production was cited at over 2000 exabytes by 2020. The tools include data compression methods to accelerate processing of massive distributed datasets.

Who owns the IP and how can I access it?

As an EU-funded RIA project, IP is typically owned by the consortium partners who generated it. The coordinator is Haute Ecole Spécialisée de Suisse Occidentale in Switzerland, with 10 partners across 5 countries. Licensing arrangements would need to be discussed with the relevant partner holding specific IP rights.

Has this been tested with real users in real clinical settings?

Yes. The demo deliverable — a dynamic visual analytics prototype — was explicitly tested with expert users in the field. The prototype was refined based on their feedback, confirming usability in actual clinical workflows.

What exactly was delivered — software, models, or methodology?

The project produced 23 deliverables in total, including a dynamic visual analytics prototype for interacting with learning methods and understanding intermediate learning states. The objectives also included decision support prototypes and a semantic middleware for image compression, segmentation, and classification.

Does this comply with medical data regulations?

Based on available project data, specific regulatory certifications (CE marking, FDA clearance) are not mentioned. As an EU-funded research project with hospital partners, it operated under EU data protection rules. Any commercial deployment would require separate regulatory assessment.

How does the weakly supervised approach compare to fully supervised AI?

The core advantage is reducing dependence on expensive manual annotation. Instead of requiring large labeled datasets, EXA MODE extracts knowledge from existing clinical reports and image features using document-level semantic networks. This makes it feasible to train AI on data volumes where full manual labeling would be prohibitively expensive.

Consortium

Who built it

The EXA MODE consortium brings together 10 partners from 5 countries (Bulgaria, Switzerland, Italy, Netherlands, Poland), with a healthy 40% industry ratio — 4 industry partners including 2 SMEs alongside 4 universities and 2 other organizations. This mix signals that the research was designed with commercial applicability in mind, not purely academic. The Swiss coordinator (Haute Ecole Spécialisée de Suisse Occidentale) is an applied sciences university, which typically bridges research and industry more effectively than traditional universities. With EUR 4,333,281 in EU funding and partners spanning Western and Eastern Europe, the consortium had both the resources and the geographic diversity to test solutions across different healthcare systems.

HAUTE ECOLE SPECIALISEE DE SUISSE OCCIDENTALECoordinator · CH
ONTOTEXT ADparticipant · BG
STICHTING RADBOUD UNIVERSITEITparticipant · NL
STICHTING RADBOUD UNIVERSITAIR MEDISCH CENTRUMparticipant · NL
UNIVERSITA DEGLI STUDI DI PADOVAparticipant · IT
SURF BVparticipant · NL
ONTOTEXT ADparticipant · BG

How to reach the team

Coordinator is Haute Ecole Spécialisée de Suisse Occidentale (Switzerland). SciTransfer can facilitate a direct introduction to the project team.

Order a report on this consortium See HAUTE ECOLE SPECIALISEE DE SUISSE OCCIDENTALE profile →

Next steps

Talk to the team behind this work.

Want to explore how EXA MODE's medical image AI can work for your organization? Contact SciTransfer for a tailored briefing and introduction to the research team.

Order a report on this topic More in Health & Biomedical