Radiology · Mammography screening reading support
Lunit INSIGHT MMG
Lunit Inc.
The mammography AI that first took a human reader's chair: in the prospective ScreenTrustCAD trial of 55,581 Swedish women, one radiologist plus Lunit's algorithm detected 261 cancers to standard double reading's 250 — non-inferior on the primary endpoint — and the host hospital went on to make AI-as-second-reader its routine practice. What the record still lacks, its strongest rival can show: a randomised trial.
Performance Metrics
Clinical Evidence
The evidence base is prospective, interventional and largely independent — one grade short of randomised, and the distinction matters when comparing this device with its closest rival. The anchor is ScreenTrustCAD (Lancet Digital Health 2023): a prospective, population-based, paired-reader non-inferiority study at Capio Sankt Göran Hospital in Stockholm, in which every one of 55,581 consecutive screening exams was assessed by two radiologists, by one radiologist plus AI, and by AI alone. One radiologist plus INSIGHT MMG (version 1.1.6) detected 261 cancers against double reading's 250 — relative proportion 1.04 (95% CI 1.00–1.09), non-inferior and nominally superior — while AI alone, at 246 versus 250, was itself non-inferior to two human readers. Because the paired design gives every exam to every reading arm, it answers the accuracy question directly; what it cannot do, without randomisation and interval-cancer follow-up, is deliver the endpoint the MASAI randomised trial delivered for the competing Transpara system. That gap is the honest reason this entry carries two AurisMarks to Transpara's three. The second prospective pillar is AI-STREAM (Nature Communications 2025), a multicentre cohort study of 24,543 women inside Korea's national screening programme, where mammography is read by a single radiologist. With AI assistance, cancer detection rose 13.8% overall with no increase in recall, and the gain concentrated where support is most needed: general (non-breast -specialised) radiologists improved from 3.87 to 4.89 cancers per 1,000 screens, a 26.4% increase. The retrospective foundations are consistent. The manufacturer-authored development study (Kim et al., Lancet Digital Health 2020; 170,230 exams from Korea, the US and the UK) reported a standalone AUROC of 0.959, and in its 320-case multi-reader study 14 radiologists improved from 0.810 to 0.881 with AI assistance. The independent Swedish external validation that preceded ScreenTrustCAD (Salim et al., JAMA Oncology 2020; 8,805 women, 739 cancers) compared three anonymised commercial algorithms and found the best-performing — AUC 0.956 against 0.922 and 0.920 — significantly ahead of the other two; the ScreenTrustCAD investigators selected Lunit's system on the strength of that comparison.
| Study | Design | n | Sensitivity | Specificity | AUC | Published |
|---|---|---|---|---|---|---|
| Dembrower K, Crippa A, Colón E, Eklund M, Strand F (ScreenTrustCAD Trial Consortium; investigator-led, Karolinska Institutet) | ProspectiveProspective | 55,581 | — | — | — | Lancet Digital Health, Oct 2023 (PMID 37690911); paired-reader non-inferiority — AI + 1 radiologist 261 vs double reading 250 cancers (relative proportion 1.04, 95% CI 1.00–1.09); AI alone 246 vs 250, also non-inferior (INSIGHT MMG 1.1.6) |
| AI-STREAM investigators (Chang YW et al.; prospective multicentre cohort, Korean national screening programme) | ProspectiveProspective | 24,543 | — | — | — | Nature Communications, Mar 2025 (PMID 40050619); cancer detection +13.8% with AI assistance at no increase in recall; general radiologists improved 3.87 → 4.89 cancers per 1,000 screens (+26.4%) |
| Salim M, Wåhlin E, Dembrower K, et al. (Karolinska Institutet; independent external evaluation of 3 anonymised commercial algorithms) | RetrospectiveRetrospective | 8,805 | — | — | 0.956 (best of three; vs 0.922 and 0.920) | JAMA Oncology, Oct 2020 (PMID 32852536); 739 cancers among 8,805 women; the top performer was subsequently selected for the ScreenTrustCAD trial (Lunit INSIGHT MMG) |
| Kim HE, Kim HH, Han BK, et al. (Lunit-affiliated — manufacturer-authored development and reader study) | RetrospectiveRetrospective | 320 | — | — | 0.959 standalone (170,230-exam validation); readers 0.810 → 0.881 with AI | Lancet Digital Health, Mar 2020 (PMID 33334578); algorithm developed on 170,230 exams from Korea, the US and the UK; 320-case multi-reader study with 14 radiologists |
Clinical Pulse
Verified clinician sentiment
Verified clinician reviews launching soon · Apply to contribute →
Reviews are signed by clinicians verified against their regulator's public register.
Inside the algorithm
Editorial featureHow Lunit INSIGHT MMG reaches a verdict.
Five stages — from raw input to verdict — drawn from manufacturer documentation and the public regulatory record.
- INGEST
- NORMALISE
- DETECT
- LOCALISE
- VERDICT
Stage 01 · INGEST
Screening mammograms arrive from the workflow.
INSIGHT MMG receives full-field digital mammography studies as objects routed from the acquisition workstation or . Analysis runs in the background as exams arrive; source images are not modified and no additional imaging or contrast is required. Digital breast tomosynthesis studies are out of scope — they belong to the separate INSIGHT DBT product.
Input
Mammography DICOM
Inference
Background, per-exam
Inside the Auris+ Listing
Five more sections complete this device’s Auris+ Listing.
Decision Ledger
ProPro unlocks a private, cross-vendor log of every case you read on this device — what the AI called, what you concluded, and a one-line reason if you overrode it. Ready for the EU AI Act's deployer logging obligations when they land in 2028.
Clinical Evidence Deep Dive
ProPro unlocks the structured clinical-evidence summary — study count, target patient population, and a tabular accuracy-metrics view drawn from peer-reviewed sources.
Peer-Reviewed Publications
ProPro unlocks the curated peer-reviewed publication list with PubMed cross-links — the citation backbone of every editorial verdict.
Post-Market & Regulatory Conditions
ProPro unlocks the post-market surveillance summary, recall record, and the conditions of approval that bound real-world use.
AI Algorithm Version History
ProPro unlocks the chronological record of algorithm version changes — what changed when, drawn from manufacturer changelogs and regulatory filings.
Regulatory Approvals
K211678
Class II
Safety Record
No recalls, field safety notices or FDA safety communications attributable to Lunit INSIGHT MMG were found in indexed public sources as of July 2026 — read as "none found", not an exhaustive audit, since the MAUDE and FDA recall databases could not be queried directly from this environment. The structural risks are those of the reader-replacement configuration rather than the algorithm itself. In the European deployment one of the two human reads is gone: a cancer the AI does not flag has one fewer chance of being caught, and ScreenTrustCAD's non-inferiority bound — not superiority on every subgroup — is the evidential floor under that trade. The US configuration carries the opposite risk profile: as an adjunct to a single reader's interpretation, automation bias (over-reliance on unflagged exams) is the failure mode to monitor, and the AI-STREAM data showing the largest gains among general radiologists cuts both ways — the readers helped most are also those least equipped to overrule a miss.
Intended Use & Indications
Lunit INSIGHT MMG is a software-only device that analyses full-field digital mammography studies and presents region marks with abnormality scores to the interpreting physician. What the device is permitted to be differs sharply by jurisdiction, and the difference is the single most important thing a procurement reader should hold onto. The US 510(k) clearance (K211678, product code QDQ — computer-assisted detection and diagnosis software) covers an adjunct reading aid: the radiologist reads the mammogram, and the AI output supports that interpretation; the device is not authorised in the US as an independent reader. In Europe, under the CE mark, the same system runs in a configuration the US clearance does not cover: at Capio Sankt Göran Hospital in Stockholm, following the ScreenTrustCAD trial, INSIGHT MMG operates as one of the two readers in the double-reading workflow of a population screening programme — the first live deployment of its kind anywhere. In every configuration the final recall decision rests with the humans in the loop; in the European deployment, consensus discussion of flagged cases remains a radiologist task. Digital breast tomosynthesis is handled by a separate Lunit product (Lunit INSIGHT DBT) with its own clearances — this entry covers the 2D mammography device only.