Lunit

The detection models are the engine of Lunit's core products, and the breast platform now sold around them carries value of its own. Lunit INSIGHT MMG (2D mammography), INSIGHT DBT (tomosynthesis) and INSIGHT CXR (chest X-ray) are deep learning products that do not exist without their models.

Since acquiring Volpara in 2024, though, Lunit sells a breast cancer platform: Scorecard for automated density assessment, Risk Pathways for risk assessment, Analytics and Live for mammography quality management, Patient Hub for risk based patient outreach and compliance, and mammography reporting. Several of those carry value without the detection models, and Lunit lists Analytics as a non-device data system. The company also sells SCOPE, a separate precision oncology line for pathology. Remove the INSIGHT models and a customer would still hold quality, density, risk and outreach software. The models drive the core modules but are not the whole of what is sold.

The US indication keeps the radiologist in charge; outside the US the company markets a larger role, and the flag threshold and the path after a miss are not published. INSIGHT MMG is cleared in the US (K211678) as detection and diagnosis support to be viewed after the interpreting physician completes an initial read, and not as a replacement for a full review.

It marks suspicious areas and gives an abnormality score per lesion and per breast; the company explains the score is a likelihood of malignancy, not severity, and is not linear. The product page says it supports single and double reading, and in Sweden's prospective ScreenTrustCAD study AI plus one radiologist was tested against two radiologists. That is the setting where AI stands in for a second reader, and in a double reading program it is the decision that matters.

SecondRead AI, a patient-paid add-on, states that all results are reviewed and communicated by the radiologist, and the models do not learn live. The company publishes no score at which a lesion is marked or a case flagged, and says nothing about whether a site can tune it or how a missed cancer is detected and fed back.

The models, their training and their versions are described specifically enough to check. INSIGHT MMG's 510(k) summary describes deep learning trained on large databases of biopsy proven cancers, benign lesions and normal tissue, and at launch the company gave more than 240,000 training cases, up to 50,000 of them cancers.

An evaluation of INSIGHT DBT led by Emory University's HITI Lab with Lunit staff among the coauthors (arXiv 2503.13581, a 2025 preprint absent from Lunit's own material) names the engine version (1.3.0) and product version (1.1.2.0), the architecture (a ResNet34 feature extractor with an aggregation module, 23.4 million trainable parameters) and the training set: about 160,000 exams from the US (87 percent) and South Korea (13 percent), 29 percent cancers, with the density and pathology mix given.

The product FAQ says the models do not learn in real time, that training data comes through formal collaboration agreements, and that each version is independently validated before release. So the models are identified, the training data characterized, and versioning and update practice stated. No change notice to customers is described for version updates, and the MMG cancer count is given only as an upper bound ('up to 50,000').

A structurally short chain for the models, with the delivery partners and hosts unnamed. Lunit builds its detection models in house on biopsy proven data obtained through formal collaboration agreements. The MMG FAQ says the models do not learn in real time, that customer images are retained only for processing and are not used for retraining, and that each version is validated before release, so no third party model sits between an image and a score and the cleared model is not swapped quietly.

What is not enumerated is the rest of the path: the cloud host and region used for cloud delivery, the equipment and viewer partners through which the software reaches many sites (a Fujifilm collaboration is promoted on the homepage), and any subprocessor list. An in house, cleared model with the data terms stated for the lead product.

Prospective, peer reviewed evidence exists for the headline screening claim. In ScreenTrustCAD (The Lancet Digital Health, 2023), the screening exams of 55,581 women at a Stockholm hospital were interpreted prospectively with INSIGHT MMG alongside radiologists in a paired reader, noninferiority design, and AI plus one radiologist was noninferior to two radiologists.

An independent external evaluation in JAMA Oncology (Salim and colleagues, 2020) compared three commercial algorithms on a retrospective Swedish screening cohort, and Lunit's had the highest accuracy. An Emory led evaluation of INSIGHT DBT with Lunit co-authors (arXiv 2503.13581, 2025 preprint) found AUC 0.91 on 163,449 screening exams with 1,368 screen detected cancers; at its 0.1 operating point it flagged 73 percent of those cancers and 31 percent (33 of 105) of interval cancers, with precision of 0.08.

The product page claims 18 percent higher accuracy in extremely dense breasts (Breast Cancer Research, 2024) and up to 40 percent of cancers detected earlier (interval cancers, Radiology, 2024), both retrospective. The 510(k) summary reports a standalone study on 2,412 mammograms (AUC 0.903; sensitivity 85.7, specificity 75.6 percent) and a reader study of 12 radiologists on 240 cases in which AUC rose from 0.754 to 0.805 with AI.

A European Radiology study found slightly lower specificity and higher recall than first readers at matched sensitivity. The prospective evidence is from one site and covers screening only; the chest X-ray line rests on retrospective work.

Training use and retention are stated for the lead product; de-identification is not. The INSIGHT MMG FAQ says images from commercial sites are retained only for the duration of processing and are not used for retraining, that the models do not learn in real time, and that training data is biopsy proven material sourced through formal collaboration agreements. Those are specific commitments a buyer can hold the company to, and they answer whether customer studies feed the models.

Not stated: whether studies are de-identified before processing, where cloud processing happens, and whether the same terms hold for DBT, CXR and the former Volpara products, which run on their own platforms. The corporate privacy notice covers websites, marketing, recruiting and support contacts, not patient images. Safety engineering, meaning how a missed or false finding is handled as a safety event, is not published. Categorical commitments with one of the three elements missing.

No business associate position is stated, and the only privacy document published covers the website. Lunit's privacy notice (effective October 2025) covers website visitors, marketing and events, job applicants and support contacts; it does not address patient images or protected health information in the products, beyond asking meeting participants not to share PHI in recorded calls.

No business associate agreement, role statement or tier is published, and the product FAQ's statement that images are retained only for processing is a handling commitment, not a status. A trust center hosted on TrustShare is linked from the footer and may hold a business associate statement or an agreement. Distribution through imaging equipment partners means the contracting party may not be Lunit at all.

ISO 27001 is claimed as the company's own, with a trust center and a bug bounty program alongside it. Lunit's Information Security and Medical Device Registration page states that Lunit.io is ISO/IEC 27001 certified for information security management, and that its manufacturing site holds ISO 13485:2016, MDSAP and Korean GMP certification.

Lunit International, the former Volpara business, states that its risk management conforms to ISO 14971 and ISO 27001:2022 and that BSI certifies its quality system. The footer links a trust center hosted on TrustShare and a bug bounty program. No SOC 2 report is claimed. A recognized certification named in the company's own material, with the certificate, its scope and its date not shown.

The US position is identifiable in the FDA database for the lead product and stated by regulation for the rest. Lunit INSIGHT MMG was cleared on 7 October 2021 (K211678; 21 CFR 892.2090, computer assisted detection and diagnosis, Class II, product code QDQ) to aid detection, localization and characterization of suspicious areas for breast cancer on screening mammograms from compatible full field digital systems, as an adjunct viewed after the physician's initial read and not a replacement for it.

Lunit INSIGHT CXR Triage was cleared in November 2021 (K211733, product code QFM) for triage, and INSIGHT DBT in November 2023. For the former Volpara products the company publishes its own regulatory table: Scorecard Class II (21 CFR 892.2050), Live Class I (892.1940), and Analytics as a non-device data system, with Canadian, EU, Australian and New Zealand listings.

Sold but not US cleared: the fuller INSIGHT CXR finding set sold under CE marking beyond the cleared triage findings, and the SCOPE pathology line. Double reading use, where AI stands in for a second reader, sits outside the US MMG indication.

Results compared across patient groups are published for INSIGHT DBT, one of the two screening reads on the breast platform. An evaluation led by Emory University's HITI Lab, with Lunit staff among the coauthors, tested the product as sold (engine 1.3.0, product 1.1.2.0) on 163,449 screening exams from the Emory Breast Imaging Dataset, 1,368 of them screen detected cancers (arXiv 2503.13581, a 2025 preprint not yet peer reviewed, which Lunit does not cite). Overall AUC was 0.91.

The authors report performance not statistically different across race (the cohort was 46 percent Black and 43 percent White patients), ethnicity or age, and significantly lower for non-invasive cancers (AUC 0.85), calcifications (0.80) and dense breasts (0.90). That is model behavior tested across the groups the product serves, with the weak spots named.

For INSIGHT MMG the product page claims 18 percent higher accuracy in extremely dense breasts and 8 percent across all density categories, citing a 2024 Breast Cancer Research comparison of radiologists and standalone AI; that is a claim about one group, and no results across groups are published for MMG or CXR. The FAQ's statements that each version is validated to prevent bias and that accuracy does not differ across equipment makers are claims, not results.

An error rate with its method and denominator is published for the lead product; nothing else stands behind the output. Lunit's own 510(k) summary for INSIGHT MMG (K211678, cleared 7 October 2021) reports a standalone study on 2,412 screening mammograms from Hologic, GE and Siemens systems, independent of the development and reader study data: ROC AUC 0.903, sensitivity 85.7 percent (95% CI 83.0 to 88.5) and specificity 75.6 percent (73.6 to 77.6).

The summary does not state how many of the 2,412 were cancers or define its reference standard in the text. It is a test set result, not performance in a buyer's screening program, and it covers MMG only: not INSIGHT DBT, INSIGHT CXR or the former Volpara products. The company publishes no warranty, indemnity or remediation commitment.

Distribution strategy substitutes for integration engineering here, and effectively. The company embeds its algorithms directly into imaging hardware through OEM relationships with major scanner manufacturers and distributes via major cloud infrastructure, meaning a customer can acquire the AI as part of equipment already being purchased rather than as a separate integration project. That is a materially different adoption path than PACS side deployment. No EHR integration applies or is claimed.

Multiple delivery paths are evidenced, spanning embedded OEM deployment inside imaging hardware and cloud distribution through a major provider, which gives buyers real choice about where inference runs. Specific tenancy, hosting, and data residency terms are not published, which matters given the global footprint and varying localization requirements.

No price is published, and the ways the product is bought are discoverable. Lunit sells INSIGHT and its breast platform direct and through imaging equipment and viewer partners, so the software can arrive inside or alongside a scanner or viewer purchase.

SecondRead AI is a patient-paid option: a provider offers an enhanced AI read at annual screening as an out-of-pocket service, with training and marketing materials supplied by Lunit, and no price for the patient or the provider is published. As a KOSDAQ listed company Lunit publishes financial results, which inform vendor stability but not product cost.

Deliberately concentrated on cancer detection rather than general radiology, spanning chest X-ray for lung abnormalities and mammography plus tomosynthesis for breast, with a separate oncology business line addressing therapy response. Settings center on organized screening programs and breast imaging centers, which is where the double reading economics the company targets actually exist. Emergency triage, neuro, musculoskeletal, and general radiology workflows are outside scope, distinguishing it from broad platform competitors.