Validated prototype
Core components and AI-assisted workflows evaluated in representative settings.
Startup Voucher 2025
Operation, activities and results
A public reporting view of BreastScreening-AI's human-centred clinical AI development, validation activities and supporting research.
Operation
Clinical decision support designed around radiologists
BreastScreening-AI develops and validates artificial intelligence tools that support the interpretation of breast imaging and clinical decision-making without replacing professional judgement. The operation combines multimodal imaging, human-computer interaction, clinical workflow integration and responsible AI evaluation.
Objectives
- Improve the consistency and efficiency of breast imaging assessment.
- Evaluate AI assistance in realistic clinical workflows with healthcare professionals.
- Reduce avoidable diagnostic errors while preserving clinician control and accountability.
- Prepare the technology, evidence and governance required for further clinical validation.
Technology readiness
Progression from TRL 5 toward TRL 6
The Hospital da Luz work provided evidence of integration and usability testing in a relevant clinical environment. This supports a reported evolution from TRL 5 to TRL 6, while not constituting formal TRL certification or clinical deployment approval.
Relevant-environment demonstration
Integration through Siemens syngo.via Frontier/OpenApps and usability activities with physicians.
Broader validation
Larger prospective studies, consolidated CHTMAD outcomes and formal regulatory evidence.
TRL 6 is an evidence-based project interpretation, not an external certification.
Clinical evidence
Hospital da Luz exploratory pilot
The pilot involved seven physicians and assessed workflow integration, usability and decision support. Results are preliminary, use a small sample and should not be interpreted as clinical efficacy claims.
Exploratory improvement in triage-level decision accuracy.
Decisions that remained unchanged with AI assistance.
Statistical result for the complementary triage analysis.
Participants in integration and usability activities.
Where the pilot showed measurable progress
The clearest observed effect was a +11.82 percentage-point improvement in the complementary triage-level analysis, supported by the exploratory statistical result of p = 0.0036.
BI-RADS remained the primary clinical reference throughout the pilot. The triage analysis provides an additional view of decision support and is not intended to replace BI-RADS or clinical judgement.
- Approximate subset
- 11 patients
- Image volume
- 23 images
- Paired observations
- 110
- Environment
- Siemens syngo.via Frontier/OpenApps
| Evidence period | Indicator | Reported value | Status |
|---|---|---|---|
| Previous, 2022 | False positives | 27% decrease | Peer-reviewed research |
| Previous, 2022 | False negatives | 4% decrease | Peer-reviewed research |
| Previous, 2022 | Positive expectations and satisfaction | 91% of clinicians | Peer-reviewed research |
| Current, 2026 | Triage-level decision accuracy change | +11.82 percentage points | Exploratory pilot |
| Current, 2026 | Decision stability | 81.82% | Descriptive pilot result |
| Current, 2026 | Clinical classification framework | BI-RADS retained | Primary clinical reference |
Field validation
CHTMAD / ULSTMAD research activities
CHTMAD activities investigated clinical reporting, documentation and human-machine-readable report structures using anonymised cases under an ethics-approved research protocol.
Current reporting status
Three fieldwork periods are documented: 17-21 November 2025, 5-9 January 2026 and 26-30 January 2026.
The available evidence supports reporting the activities and methodology. No consolidated quantitative outcome data was reported yet, so no CHTMAD timeline is published here.
Scientific comparison
Results reported in peer-reviewed research
These figures provide research context only. They come from a different study design and sample and must not be combined statistically with the Hospital da Luz pilot.
Artificial Intelligence in Medicine, 2022
BreastScreening-AI: Evaluating Medical Intelligent Agents for Human-AI Interactions
A real-world case study compared clinician-only and clinician-AI scenarios with 45 clinicians from nine institutions.
- 27% decrease in false positives.
- 4% decrease in false negatives.
- 91% of clinicians reported positive expectations and perceptive satisfaction.
- Time-to-diagnose decreased by three minutes per patient.
| Published indicator | Reported result |
|---|---|
| False positives | 27% decrease |
| False negatives | 4% decrease |
| Positive expectations and satisfaction | 91% of clinicians |
| Time-to-diagnose | 3 minutes less per patient |
Specialist support
Contributors to operation readiness
External organisations supported distinct technical, funding, legal, regulatory and intellectual-property workstreams.
SNAP
Proposal development, EIC project management and Grant Agreement Preparation support.
Leyton
Startup Voucher and PRR technical reporting, cost-eligibility and communication compliance support.
SAVEAS
Intellectual-property consultancy, including Freedom-to-Operate and term-sheet work.
KGSA
Our legal provider, supporting the company with corporate, contractual and legal matters.
Complear
Regulatory and independent-validation discussions for medical AI readiness.
AAVANZ
Supported the preparation of our EIC Pathfinder and Horizon Europe proposals.
Reporting note
Transparent by design
This page distinguishes confirmed pilot results, research activities and peer-reviewed evidence. Values are not pooled across studies. Clinical results remain preliminary until larger prospective validation and formal institutional review are completed.