Vara AI
Vara AI is CE-marked Class IIb breast imaging AI for radiologists and organised screening programmes. It triages normal mammograms, drafts their reports automatically, raises a Safety Net on suspicious cases, and rests on peer-reviewed clinical evidence.
What is Vara AI?
Vara AI is breast imaging AI published by MX Healthcare GmbH, a Berlin company. It reads screening and diagnostic mammograms, and the homepage positions it as Independent AI for European Breast Screening. It is CE marked under Regulation (EU) 2017/745 (MDR) — the footer names the devices covered as Vara MG and Vara AI — and the site describes it as Class IIb, among the few breast imaging AI certified as an independent reader for screening and diagnostic use.
It comes in two shapes. Vara Workflow is a web-based platform for screening teams, with a worklist, viewer-side triage, automated reporting and departmental dashboards. Vara Plugin is a lightweight deployment connecting to an existing PACS via API, adding AI without changing the workflow.
The clinical concept behind both is decision referral: the algorithm makes a statement only where it is confidently predicting well, leaving the rest to human expertise, in three pillars. Normal triaging pre-labels high-confidence normal exams, and the site claims 80% of normal cases are pre-reported. The safety net fires when a radiologist marks negative a case the AI finds highly suspicious, and points to the specific region. Unclassified cases go entirely to the radiologist.
The outputs are an exam-level score from 0 to 100, the scoring categories Unremarkable, Neutral and Suspicious, CAD-SR lesion markings with per-lesion scores, and a structured Secondary Capture report. DICOMs go in and results come back in under 30 seconds. Tomosynthesis is supported with the same report across synthetic 2D, slabs and the full 3D stack, and is CE marked under EU MDR for DBT. Deployment is cloud by default, with on-premise available for partnerships where data cannot leave the site; it works with any PACS and any mammography device and needs no new on-site infrastructure.
The evidence base is unusually visible: six peer-reviewed publications, headed by PRAIM in Nature Medicine, which the site calls the world's largest prospective AI study in healthcare, with 460,000 participants across 12 centres. After deployment, predictions are monitored continuously, quarterly per-radiologist reviews check overrides and safety-net catches against biopsy outcomes, and cancer-registry linkage supplies final diagnoses. A company news post of 02/09/2026 announces a further CE certification called Autonomous Triage — reporting a screening mammogram as normal with no radiologist reading it — backed by ATMON, the company's safety-monitoring system, which reverts a site to full radiologist reading when signals drift outside defined limits.
What it does
- Send mammography exams straight from the PACS: DICOM images go out, an AI report comes back.
- Receive an exam-level classification — unremarkable or suspicious — with an overall AI score from 0 to 100.
- Let high-confidence normal cases be triaged and pre-reported automatically, then review, sign and send.
- Get a Safety Net alert, with the region marked, when the radiologist calls a case negative but the AI is confident it is suspicious.
- Read lesion-level CAD-SR markings with per-lesion scores instead of binary CAD marks.
- Tune the operating point per site, shifting thresholds toward sensitivity or toward specificity.
- Track the deployment through dashboards, continuous prediction monitoring and quarterly review sessions.
When to use Vara AI / When not to
A quick filter to help you decide if Vara AI is the right fit.
When to use Vara AI
- Organised population breast-screening programmes reading at national scale — the site names Germany and describes expansion across Europe.
- Radiologists reading high mammography volumes who want normal-case triage plus a miss-catching Safety Net inside their existing viewer.
- Screening centres looking to raise throughput without hiring, in the site's own framing: screen more women at higher quality without adding staff.
- Sites adopting tomosynthesis that cannot absorb the extra reading time a 3D workflow creates.
- Health authorities, quality and regulatory functions, and hospital IT teams that need configurable thresholds, documented post-market surveillance, quarterly outcome reviews, and AI delivered as a PACS plugin rather than as a new system.
When not to use Vara AI
- Consumers and patients: this is not a patient-facing product, and there is no signup, no app store listing and no self-serve entry point at all.
- Anyone looking for general-purpose radiology AI — the scope is breast imaging only, covering mammography and tomosynthesis.
- Buyers who need public pricing or a trial before speaking to sales: no price is published anywhere and every call to action is a demo request.
- Organisations outside the CE regulatory framework: the footer states that CE marking applies only to the devices and configurations whose conformity assessment is complete, and that availability may vary by country.
- Sites that cannot integrate at the PACS/DICOM level, or that expect the AI to replace the radiologist in the general case — outside the specific certified configurations the site describes, it works alongside the reader.
How to use Vara AI
A typical end-to-end flow, from setup to results.
- Start from the site: there is no self-serve path, and every Book a demo button leads to the contact form, which asks for name, e-mail, company and message.
- Expect an organisational procurement rather than a signup — the buyers the site names are screening programmes, hospitals, health authorities and distribution partners.
- Scope the project with the vendor, since no public documentation, API reference, help centre or status page exists to consult beforehand.
- Choose the integration route: Vara Plugin, which connects to the existing PACS via API and returns the AI report into the same workflow, or Vara Workflow, the web-based platform with worklist, triage, reporting and dashboards.
- Decide on hosting at deployment time — cloud by default, or fully on-premise for partnerships where data cannot leave the premises.
- Send DICOMs from the PACS and read the AI report back in the usual workflow; the site stresses that nothing changes in the workflow and that no new on-site infrastructure is required.
- Configure the operating point for your site, tuning thresholds toward sensitivity or toward specificity.
- Revisit performance in quarterly review sessions, where thresholds and outcomes are examined again.
Pros & Cons
Pros
- An unusually strong published evidence base for this market: six peer-reviewed publications, including PRAIM in Nature Medicine, described by the site as the largest prospective AI study in healthcare, with 460,000 participants across 12 centres.
- CE marked under EU MDR 2017/745 at Class IIb, with the independent-reader certification the site says few competitors hold.
- Deep real-world deployment: the site states that more than 50% of Germany's national screening programme runs on Vara.
- Post-market accountability is treated as a product feature rather than an afterthought — continuous prediction monitoring, quarterly per-radiologist reviews, and cancer-registry linkage for biopsy outcomes.
- Adjustable thresholds instead of a fixed black-box output, and lesion-level context instead of binary CAD marks.
- Low integration friction: PACS- and device-agnostic, plugin deployment, no new on-site infrastructure, results in under 30 seconds, and an on-premise option where imaging data cannot leave the site.
- Tomosynthesis reuses the same report and the same integration as 2D, so adopting DBT does not mean rebuilding the workflow.
Cons
- No pricing whatsoever is published — no page, no range, no model; every path leads to a demo request form.
- No terms of service, no EULA and no data processing agreement are published anywhere on the site.
- The only privacy policy published governs website visitors; nothing is published about how patient imaging data is handled, retained, hosted or used by the clinical product.
- No security or trust page, no ISO 27001 or SOC 2 claim, no published sub-processor list for the clinical platform, and no status page.
- The site contradicts itself on its own numbers: stat widgets show 100,000 PRAIM participants against 460,000 in the prose of the same pages, and 5+ team members against the ten people named on that same page.
- No public API or integration documentation exists although the site claims a PACS API, so technical evaluation requires contacting sales.
- The offer is narrow by design — breast imaging only — availability is explicitly said to vary by country, and no support e-mail is published, support running instead through a generic contact form.
Pricing & Plans
No pricing is published by this vendor. A sweep for currency symbols and codes across every archived page of the site returned zero hits, and /pricing returns a genuine HTTP 404, as do /terms, /faq, /demo, /products and /integrations. That emptiness is a verified fact about the site, not a gap in research. No tier, no list price and no billing model is stated, and no free plan or free trial is mentioned anywhere — neither offered nor ruled out. The commercial model is contact-sales: every call to action reads Book a demo and points at the contact form, and the buyers named by the site are screening programmes, hospitals, health authorities and distribution partners. Access therefore runs through an enterprise medical-device sales conversation rather than through a subscription signup.
Data, GDPR & hosting
A consolidated view of how Vara AI handles your data.
GDPR overview
The footer lists GDPR Compliant under a Certifications heading — a self-declaration, not an audit; no certification body or certificate number is named. The privacy policy is framed as an Art. 13 and Art. 14 notice, covers website visitors only, and carries no effective date or version. Data subject rights are enumerated with their articles: withdrawal 7(3), objection 21, access 15, rectification 16, erasure 17, restriction 18, portability 20, Art. 22, plus the Art. 77 complaint right. Third-country transfers rely on adequacy decisions, otherwise on Art. 46 clauses or binding corporate rules, with Art. 49 as a last resort; the policy warns that foreign authorities may gain access. US recipients are covered only under the EU-US Data Privacy Framework (Google, LinkedIn, Meta). Section 25 TDDDG is cited for device storage. No Art. 27 representative is named or needed, and no ISO 27001, SOC 2 or security page exists.
Who owns the data?
The only data document this vendor publishes is a website privacy policy, and it governs visitors to the marketing site, not patient imaging. Within that scope the named controller is MX Healthcare GmbH, legal representative Jonas Muff, Max-Urich-Str. 3, 13355 Berlin, Germany, with a designated data protection officer reachable at dataprotection@vara.ai or at the same postal address. No terms of service, no EULA and no data processing agreement is published anywhere on the site, and nothing at all is stated about who owns or controls the mammograms the clinical product handles. That question has to be settled contractually with the vendor.
Reuse rights
On the website, the privacy policy enumerates each purpose with its Art. 6 GDPR legal basis: service provision, contact handling, newsletter, job applications, necessary and functional tools, analytical tools, marketing tools, social-network business pages and data-subject requests. The data categories listed include connection data (HTTP header, user agent, IP address), device data, usage data, identifiers, profile data, location data, communication data and applicant data. Marketing tools are described as building interest profiles, serving personalised advertising, retargeting, tracking conversions across sites and devices, and matching hashed e-mail addresses with social media partners — all on consent. Joint controllership is declared with LinkedIn Ireland for business page insights and with Meta Platforms Ireland for Meta Pixel event data. The policy states that automated decision-making producing legal or similarly significant effects does not take place, and that automated decisions are limited to personalised content and advertising. None of this describes the clinical product: nothing is published about how patient mammograms are used or reused. The site does say that AI performance is monitored continuously, that the product is continuously improved through real-world feedback, and that quarterly review findings are used to learn and improve — but it never states whether customer or patient images train the models, and that question remains open.
Data retention & training
Hosting summary
What this vendor publishes about hosting is a deployment model, not a location. The site states that the cloud deployment is the default, with monitoring and analytics included, and that a fully on-premise option is offered for partnerships where data cannot leave the premises, the full 3D solution then running on the customer's own infrastructure. No hosting country and no hosting region is named anywhere on the site, which is why both fields are left empty here: the geography is simply unstated. The privacy policy does discuss third-country transfers, but only for website tooling such as Google, Hotjar, LinkedIn, Meta and Wix; that is marketing-site analytics and must not be read as a statement about where mammograms are stored or processed. Nor does the DNS resolution of the marketing site tell you anything: it points at a content delivery edge serving the website, not at clinical infrastructure. A buyer who needs a jurisdiction guarantee for patient imaging will have to obtain it contractually from the vendor, since no such commitment appears anywhere the site publishes.
Things to keep in mind
Risks and trade-offs to weigh before adopting Vara AI.
- The site contradicts its own numbers. Stat widgets display 100,000 PRAIM participants against the 460,000 stated in the prose of the same pages, and 5+ team members while ten people are named on that same page; other counters read 1 M Reads and 1 years of continuous deployment, which cannot be literal.
- The practical consequence: treat every figure displayed on the site as needing vendor confirmation before you rely on it in a business case or a tender.
- The deployment share differs by source — more than 50% of Germany's screening programme on the main site, against more than 60% and 250,000+ screenings per month in the 02/09/2026 post on the company's own news site. Both are first-party.
- CE marking is explicitly conditional: it applies only to the devices and configurations whose conformity assessment is complete, and availability of devices and features may vary by country.
- The only privacy policy published covers website visitors. Do not read it as a statement about patient imaging data: its analytics and advertising sections describe the marketing site, not the clinical platform.
- No terms of service, no data processing agreement, no sub-processor list for the clinical platform, no security certification and no public API documentation are published, even though the site claims a PACS API. All contractual and technical due diligence happens off-site.
- The company's founding date could not be established: the About page displays the year 2018, while the first web archive capture of the domain is 26/09/2020, so this record leaves the field unfilled rather than pick one.
Setup & Integrations
Technical difficulty
Moderate and entirely vendor-mediated. There is no self-serve setup: onboarding starts with a demo request and runs as a vendor-led integration project. The technical work is a PACS/DICOM integration: the Plugin connects via API and returns a structured report into the same workflow. The site claims any PACS, any mammography device, no new infrastructure, no new report to learn and no added steps for the radiologist, so the reader's burden is light. It falls on hospital IT, heavier for on-premise, and on clinical governance: a Class IIb device entering a regulated screening pathway. No public documentation exists to scope it.
Deployment
Behind Vara AI
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Alternatives
Tools that compete with or complement Vara AI.
Frequently asked questions
What does Vara AI do?
Is it a certified medical device?
Does it work with tomosynthesis?
How does it integrate with existing systems?
Where is the data hosted?
How much does it cost?
What clinical evidence supports it?
How is performance monitored after deployment?
Is there an API or developer documentation?
Who publishes Vara AI and how can they be contacted?
Should you pick Vara AI?
Vara AI arrives with the strongest published evidence base in its category: six peer-reviewed publications, headed by PRAIM in Nature Medicine, which the site describes as the largest prospective AI study in healthcare, at 460,000 participants across 12 centres. It is CE marked under Regulation (EU) 2017/745 at Class IIb, and the site describes it as certified to operate as an independent reader for screening and diagnostic use, a status it says few competitors hold. That marking is explicitly conditional: it applies only to the devices and configurations whose conformity assessment is complete, and availability of devices and features may vary by country. A post of 02/09/2026 on the company's own news site announces a further CE certification called Autonomous Triage, for reporting a screening mammogram as normal with no radiologist reading it, backed by the ATMON monitoring system; that is the company's own announcement, and it is not repeated on the marketing site. Deployment depth is real — the main site states that more than 50% of Germany's national screening programme runs on Vara, while the news site puts it above 60% at 250,000+ screenings per month, a later and slightly different claim. Post-market monitoring is treated as part of the product rather than a compliance chore.
Against that, the commercial and contractual surface is opaque. No pricing, no terms of service, no data processing agreement, no security page and no policy on clinical data are published, and the only privacy policy covers website visitors. The site also contradicts its own figures in its stat widgets, which is reason enough to confirm any number with the vendor. Evaluation therefore cannot be completed from the website alone: it requires a sales conversation.
- Choosing a selection results in a full page refresh.
- Opens in a new window.