Mindpeak

Deep learning models are the whole company. The portfolio is a set of biomarker quantification algorithms operating on H and E, immunohistochemistry, and multiplex immunofluorescence images, with no platform, scanner, or assay of its own. That purity is also a commercial dependency: the algorithms need someone else's workspace to reach a pathologist, which is why they are distributed through PathAI's platform as well as directly.

Assistive quantification with the pathologist in control of the result, and the failure side unpublished. Mindpeak ships both region of interest products, where the pathologist selects the area analyzed, and whole slide analysis. Its General Terms and Conditions (5 July 2025) describe the quantitative analysis as only a suggestion that the user can influence and change by changing the settings, and require that users only analyze images for which the service was developed, meeting stated requirements for staining method, sharpness, resolution and section thickness. The independent studies the company cites frame the tools as improving agreement between pathologists rather than replacing the reader.

The part missing is what happens when a result is wrong or an image falls outside the requirements: no thresholds, rejection behavior for out of specification slides, or route for reporting a wrong result are published.

The oversight structure is described with the failure handling missing.

The technical claim is specific and checkable: BreastIHC was described at launch as the first deep learning solution able to distinguish tumorous from non tumorous structures at the cellular level in single stained slides, classifying cells into positively stained tumor and unstained tumor populations. The company markets plug and play deployment without per site calibration, which is a meaningful robustness claim.

Training data scale and architecture are not published, and the robustness claim is asserted rather than independently benchmarked, but independent studies have tested the tools across image variance.

Mindpeak runs its own models on its own hosted service, and its terms leave the rest of the chain open. The General Terms and Conditions (5 July 2025) say Mindpeak makes its computational pathology service available on a central server or several servers and hosts it on redundant server systems, and the company builds its own models, including what it calls its patented Pathology Frontier Model, so no outside model provider sits behind a result. The algorithms are also offered inside PathAI's AISight platform, where images pass through PathAI's environment. The same terms, and the separate image upload policy, grant Mindpeak and its affiliates a perpetual right to use uploaded images and to sublicense that right, for purposes including advertisement, product improvement and the retraining of AI. No hosting provider, affiliate, sublicensee or subprocessor is named, and no retention or deletion term governs images once they arrive.

The architecture is described and no other party is named, with an open right to pass images on.

Independent evidence exists. A study by Abele and colleagues assessed the Breast Ki-67 and ER/PR products for quantifying Ki-67, ER, and PR, finding reliability across a wide range of image variance and pathologist confirmation, and concluding the tools can improve interobserver agreement in real world clinical settings.

A separate PubMed indexed concordance study at Aga Khan University Hospital compared automated Ki67 scoring against expert pathologist eyeball assessment across 60 invasive ductal carcinoma cases and reported strong positive concordance. Both are third party rather than vendor run, though sample sizes are modest and endpoints are concordance rather than patient outcome.

What happens to uploaded images is stated, and it runs against the uploading laboratory. The General Terms and Conditions (5 July 2025, section 7) and the separate image upload policy grant Mindpeak and its affiliates a non-exclusive, royalty free, perpetual right, unrestricted as to territory, time and content, to copy, transmit, reproduce, publicly display, publicly perform, edit and reformat images analyzed by the service, with the right to sublicense, for documentation duties, service provision, advertisement and product improvement, in particular the retraining of artificial intelligence. Because the clause sits in the general terms, it applies to clinical use, not only to trial uploads. The customer must warrant that it owns the necessary rights and has obtained all applicable consents. The upload policy page is marked noindex, so search engines do not surface it.

Nothing addresses de-identification of images, a retention period or deletion, and no safety engineering is described for handling a wrong output as a safety event.

Training use and indefinite rights are stated, with de identification and retention unaddressed and the terms operating against the laboratory.

No HIPAA position is stated, and a document reaching the service is published, one that runs against HIPAA's usual limits. The footer carries a privacy policy, the image upload policy and the General Terms and Conditions, and none states business associate status or mentions HIPAA. The general terms and the upload policy do reach the product: they grant Mindpeak a perpetual, sublicensable right to use analyzed images for advertisement and the retraining of AI, and require the customer to warrant it holds all consents. Mindpeak has a US entity in Cambridge, Massachusetts and names US laboratory customers, so a US laboratory needs to know whether uploaded images are de-identified and outside HIPAA, or are protected health information, in which case those rights and a business associate's obligations conflict. Nothing published says which.

A document reaching the service with no business associate position stated.

An outside penetration test is claimed, and no security attestation is published. The home page lists Penetration Test Certified among the company's credentials, a badge the company has shown with a December 2025 date, which establishes that independent testing was done and roughly when. ISO 13485 certification of the quality system is also listed, but it is a medical device quality standard, not an audit of information security. The general terms describe user IDs, account security duties and hosting on redundant servers with 99.5 percent availability. No ISO 27001 certificate, SOC 2 report, HITRUST certification, trust center, penetration test scope or subprocessor list is published, which matters more because the terms allow images to be sublicensed to unnamed parties.

The controls have an outside check behind them, but none is attested against a recognized framework.

A European regulatory position stated, with everything else research use only. The home page lists 11 CE-IVD products and ISO 13485 certification, and a September 2026 announcement reports that Mindpeak earned IVDR certification; the CE-IVD portfolio covers breast HER2, ER, PR and Ki-67 and lung PD-L1, and BreastIHC reached CE-IVD for primary diagnosis early. The site's regulatory note states that unless otherwise stated all products are intended for research use only and not for use in diagnostic procedures, which covers the prostate and other tumor type assays and use outside the markets where the CE-IVD marks apply. FDA's database lists no clearance for Mindpeak. The company has a US entity in Cambridge, Massachusetts and names US laboratory customers, so a US buyer is buying research use only software.

A non US regime named, with the research use only position stated for the rest; no US clearance.

Bias is addressed by assertion, without results. A published answer states that Mindpeak prioritizes diverse data curation, cross site validation and continuous monitoring to support responsible and generalizable deployment. Cross site validation is the relevant practice in pathology, where staining protocol and scanner variation between laboratories is a real source of model failure, but no method, measurement, threshold or result accompanies any of the three. No performance is published by patient group, scanner or site across a portfolio spanning breast, lung, gastric, esophageal, urothelial, ovarian and prostate indications.

The governance answer also sits awkwardly beside the licensing terms: it refers to continuous monitoring, while the company's image upload policy grants a perpetual right to use uploaded customer images for retraining its models. That retraining practice is disclosed in a licensing document rather than in the account of how the models are governed.

Governance practices named with no evaluation published.

The pathologist can adjust every result, and the terms disclaim its accuracy. The General Terms and Conditions (5 July 2025) state that Mindpeak assumes no responsibility for the accuracy of diagnostic results, that the service works on probabilities calculated by AI, and that the quantitative analysis is only a suggestion the user can influence and change by changing the settings. The region of interest products leave the region to the pathologist. The terms warrant the contractually agreed quality of the service and set error classes with reaction times, which concern service availability, not what a result says.

No error rate with a method and denominator is published on the company's pages, although it cites independent studies of its breast products. No warranty, indemnity or remediation commitment attaches to a result.

Review and adjustment by the pathologist, with accuracy expressly disclaimed.

Reach comes through platform distribution rather than direct integration engineering. Since 2025 the CE-IVD portfolio is accessible inside PathAI's AISight platform, which is itself FDA cleared for primary diagnosis in the US and CE-IVD marked in Europe. For a laboratory already running that platform, adoption is algorithm selection rather than an integration project.

For everyone else, integration depends on which host workspace is in place, and the vendor does not publish a broader connector list. The algorithm reaches the pathologist through someone else's product.

A single hosted service with its location unstated. The General Terms and Conditions (5 July 2025) say Mindpeak makes the service available on a central server or several servers, hosts it on redundant server systems and commits to 99.5 percent annual availability. No hosting provider, country or region, tenant isolation or on premises option is described. The algorithms are also available inside PathAI's AISight platform, where hosting follows that platform. The company says its technology runs in more than 40 laboratories and connects to existing scanners, laboratory information systems and cloud storage through more than 18 image management partner integrations, which is integration rather than deployment detail.

A single hosted option with location implied rather than committed.

No pricing is published. The dual route to market, direct and embedded in PathAI's platform, means commercial terms likely differ by channel, and neither is disclosed. Which route a lab buys through affects both the cost structure and who holds the support and data relationship.

Concentrated on breast cancer immunohistochemistry, the highest volume biomarker workload in European pathology, extending to lung PD-L1 clinically and to prostate and other tumor types for research use only. Depth in a high demand area rather than breadth across pathology, and the clinical versus research use split means actual coverage is narrower than the full product list suggests.