MaMMa Klinika added Kheiron's Mia AI as a third reader to Hungary's national breast-screening programme. Across 25,065 women (2021-23), Mia found 24 more cancers (+7%, 83% invasive) versus double human reading, with only 70 extra recalls (+0.28%). Published in Nature Medicine.
+7 %
Additional cancers detected vs double human reading (Apr 2021-Jan 2023)
MaMMa Klinika (MaMMa Egészségügyi Zrt.), with Kheiron Medical Technologies and Imperial College London, within Hungary's national breast-screening programme
Period
April 2021 – January 2023 (three-phase prospective study); AI reader in live clinical use since the final phase
Region (NUTS)
HU110
Keywords
healthcare, medical imaging, AI diagnostics, public health screening
Context
MaMMa Klinika, a private-law provider operating within Hungary's state-organised, population-based breast-cancer screening programme for women aged 45-65, added Kheiron Medical Technologies' Mia AI system as an additional reader alongside its existing double-human-reading mammography protocol, developed in partnership with Imperial College London.
Activities
Between April 2021 and January 2023, Mia was rolled out in three phases across four screening sites — a single-centre pilot, a wider multicentre pilot, and a full live clinical rollout — covering 25,065 women. When Mia's assessment disagreed with the two human readers, a senior radiologist arbitrated whether to recall the patient, without materially increasing radiologists' workload.
Results
Across the study, Mia's involvement led to 24 additional cancers being detected compared with double human reading alone, a 7% relative increase, 83% of which were invasive. This came at the cost of only 70 additional patient recalls, a 0.28% relative increase in the recall rate. Detection gains attenuated across phases: 13% in the single-centre pilot, 10% in the wider pilot, and 5% in the full live rollout.
Conclusions
Published as a peer-reviewed prospective implementation study in Nature Medicine, this is one of the first papers to report AI mammography performance from routine clinical use rather than a retrospective or simulated dataset. It lacks a randomised control arm, so it describes a real-world service evaluation rather than definitive causal proof, and governance sits with a private clinical provider delivering a publicly mandated screening service rather than directly with a government agency.
Implementation
Indicative cost
Medium (€50k–€500k) — Costs are not itemised in the source; running an additional AI reader plus senior-radiologist arbitration implies incremental cost per screening round beyond standard double reading, but no per-screen or programme-wide figure is published.
Time to results
Medium (1–3 years) — Phased rollout over roughly 21 months (April 2021 to January 2023), moving from single-centre pilot to full live clinical use; the AI reader has been in live use since the final phase.
Staffing & skills
MaMMa Klinika radiologists (double human readers plus senior-radiologist arbitrator), Kheiron Medical Technologies (AI system developer), Imperial College London (development/research partner)
Conditions for success
AI positioned as a third reader that only triggers senior-radiologist arbitration on disagreement, avoiding a large increase in radiologist workload
Phased rollout (single-centre pilot, multicentre pilot, full live rollout) allowing performance to be checked before wider deployment
Integration within an existing double-human-reading protocol rather than replacing human readers
Common failure modes
Detection gains attenuated moving from pilot to routine practice (13% pilot vs 5% full rollout), the expected pattern as a system moves from controlled conditions to real-world use
No randomised control arm, so causal impact is not definitively established
Where it fits
Governance type
private clinical provider within a state-mandated public screening programme
Scale
four screening sites
Income level
high-income
Commonly funded by
National / regional programmes
Indicative funding routes for practices of this type — always check each programme's current calls and eligibility rules.
Do you run this practice?
Claim it —
verified implementers get a public contact pathway and can propose corrections.
Data sources
Where this practice's information was retrieved from, and when.