PathAI

PathAI sells two things, and its models are the engine of one of them. AISight is an image management system for case management, slide viewing and collaboration; its cleared edition, AISight Dx, is cleared for primary diagnosis as an image management system, and laboratories such as Labcorp and MedStar Health deploy it for that purpose. Beside it sits PathAI's own algorithm family: AIM-MASH AI Assist for scoring liver biopsies in MASH trials, AIM biomarker quantification (PD-L1, HER2, ER, PR, Ki-67), TumorDetect for case prioritization, ArtifactDetect for slide quality control, AIM-TumorCellularity and PathAssist Derm. The company says its algorithms are trained on more than 32.5 million annotations from a network of more than 450 board certified pathologists, and the platform also hosts partners' algorithms.

The image management business stands without the models; the AIM products and the biopharma services built on them do not.

The model is the engine of a core module on a platform that carries other value.

For the lead algorithm the review workflow is published in detail; for the laboratory algorithms it is not. In AIM-MASH AI Assist, as the company's Nature Medicine 2025 validation describes it, the primary pathologist reviews every algorithm score and performs a quality control review of the slides, deciding whether restaining or rescanning is needed. A disagreement of two points or more on any component sends the case to a second pathologist, and continued disagreement goes to a consensus call; the workflow deliberately limits when the pathologist can change the algorithm's score, to standardize reads. The tool is qualified by FDA and EMA for MASH trial enrollment and endpoint assessment, and the company states it is not for use in diagnostic procedures.

The laboratory tools publish no equivalent. TumorDetect prioritizes cases with likely tumor to expedite IHC orders and second reviews, and ArtifactDetect flags slides for rescanning or restaining, with no threshold, override path or error handling described on the product pages. The company calls AIM-MASH 100 percent repeatable, which describes precision, not correctness.

The oversight structure is described for the lead algorithm, with the laboratory tools' part missing.

The approach and training data are described; version and update discipline are not published beyond the lead algorithm being locked. The company describes its models as trained on more than 32.5 million annotations from a network of more than 450 board certified pathologists, and its AIM-MASH validation paper (Nature Medicine 2025) describes a locked algorithm scoring the MASH CRN components from H&E and trichrome whole slide images, developed and verified beforehand against a panel of manual readers. The product pages list the algorithms by name and indication, including the stains and antibody clones each biomarker algorithm supports (AIM-PD-L1, for example, with Dako 28-8 and 22C3 and Ventana SP142 and SP263). The FDA and EMA qualifications fix the version of AIM-MASH used in trials, and the paper says performance monitoring will indicate when further training may help.

No version history, model release notes or update practice is published for any algorithm.

The approach is named without the version and update discipline behind it.

A short chain for PathAI's own models, and a partner layer named but not described. PathAI builds its own algorithms, trained on annotations from its own pathologist network, so no outside model provider sits behind them. The platform also embeds partner algorithms through AISight Link, and the algorithm menu names the partners: Deep Bio, DoMore Diagnostics, Mindpeak, Paige, Primaa, Stratipath and Visiopharm. A slide sent to one of them reaches that company's model under terms the buyer has to obtain separately. The privacy notice lists categories of service providers (hosting, data analytics, communications, customer service) without naming them, and states the platforms are hosted from the United States; the hosting provider is not named on the product pages.

No subprocessor list is published, and nothing says whether partner algorithms process images inside AISight or send them elsewhere, or what retention and training terms bind them.

Substantial partial disclosure, with the host and the partners' terms missing.

Peer reviewed, multi site validation of the lead algorithm, accepted by two regulators. Pulaski and colleagues (Nature Medicine 2025, 31:315 to 322, open access, PathAI led) validated AIM-MASH on 1,481 liver biopsies from four completed global MASH trials with about 13,000 independent reads, testing the algorithm alone and as an aid against a consensus ground truth set by panels of hepatopathologists. AI assisted reads were superior to unassisted reads for ballooning, inflammation, the MAS of 4 or more enrollment criterion (weighted kappa 0.63 against 0.51) and MASH resolution (0.54 against 0.37), and non-inferior for steatosis and fibrosis. Reads were retrospective, on stored trial biopsies. EMA's CHMP issued a qualification opinion in 2025, and FDA qualified the tool as a drug development tool on 9 December 2025 for MASH trial use.

The evidence covers AIM-MASH. The laboratory algorithms are research use only in the United States, and no validation for them appears on the product pages. Deployments at Labcorp, MedStar Health, Moffitt Cancer Center, Northwestern Medicine and University Hospital Zurich are scale of use for the image management system, not outcome evidence.

Peer reviewed, multi site and independently evaluated performance for the lead algorithm, with its method readable.

General privacy assurances that do not answer what reaches a model. The privacy notice (effective 13 February 2025), which covers PathAI's platforms, says PathAI requests only the minimum information needed and retains it only as long as necessary to provide the services or for compatible purposes, including providing additional services. It describes encryption in transit, firewalls, access control lists, logging and monitoring, network isolation and host hardening, routes protected health information to customer agreements and business associate agreements, and states candidly that complete deletion is not always possible and residual data may remain on backup media.

Three things are not published: a retention period for slide images, any statement on whether customer images or derived data train PathAI's models, and a limit on de-identified data, which the notice says PathAI may share with third parties for any purpose. The company describes its training data as annotations from its own pathologist network. Safety engineering, meaning how a wrong model output is handled as a safety event, is not published.

General assurances that leave retention, training use and de identification unanswered.

Business associate status is stated through the governing instrument, and the agreement is not published. The privacy notice (effective 13 February 2025) covers PathAI's platforms and carries a Protected Health Information section: use of the platforms may give PathAI access to protected health information as HIPAA defines it, certain PathAI entities are required to comply with HIPAA and equivalent rules elsewhere, and in those cases processing of PHI is governed by customer agreements, terms of use and business associate agreements as applicable. Naming the business associate agreement as the instrument tells a buyer where the obligations live. The notice gives a privacy officer contact and appoints DataRep as EU data protection representative, with locations in all 27 member states, the UK, Norway and Iceland.

What is missing: no business associate agreement or template is published, the tier it applies at is not stated, nothing says whether it binds subcontractors, and the hedged wording (certain affiliated entities, as applicable) leaves unclear which PathAI entity signs.

Status stated and supported by a substantive privacy document, with the agreement unpublished.

One recognized certification is named, beside a published vulnerability disclosure policy, with no artifacts. The AISight image management page states ISO 27001 certification and that the platform follows industry recognized security and privacy frameworks. The footer links a vulnerability disclosure policy that applies to all systems PathAI owns or operates, accepts reports from any source, commits to respond, work with researchers to remediate, protect PHI during the process and acknowledge researchers, and promises no legal action against good faith reporting; it treats machine learning models and training datasets as protected assets. The privacy notice describes encryption in transit, firewalls, access control lists, logging and monitoring, network isolation and host hardening.

No SOC 2 report, certificate with its scope and expiry, trust center or penetration testing statement is published.

A recognized certification named without the artifact.

The regulatory position is identifiable and matches what is sold. FDA's database lists AISight Dx (K243391, 26 June 2025, product code QKQ) under PathAI, Inc., and an earlier 510(k), K212361 (11 August 2022, QKQ), under the device name Novo. The company states that AISight Dx is cleared for primary diagnosis in the United States with the Hamamatsu NanoZoomer S360MD, Leica Aperio GT 450 DX and Roche VENTANA DP200 and DP600 scanners, and is CE-IVD marked for primary diagnosis in the EEA, the UK and Switzerland. The clearance is for the image management system, not for an algorithm.

The AI is positioned consistently with that. The product pages state that AISight, the research edition, and PathAI's algorithms, including AIM-PD-L1, AIM-HER2, AIM-ER, AIM-PR, AIM-Ki-67, AIM-MASH, TumorDetect and PathAssist Derm, are research use only in the United States, and describe ArtifactDetect, case prioritization and AIM-TumorCellularity as workflow tools; no algorithm is called cleared. AIM-MASH AI Assist is qualified by FDA as a drug development tool (9 December 2025) and by EMA for MASH trial enrollment and endpoint assessment, and the company states it is not for use in diagnostic procedures. Partner algorithms on the menu carry their makers' labeling.

Behavior across scanners and sites is tested and published with numbers for the lead algorithm, AIM-MASH. The company's validation paper (Pulaski and colleagues, Nature Medicine 2025, led by PathAI staff and cited in the company's qualification announcement) reports how consistently the locked algorithm scores the same slides scanned on different days and at three different sites. Interday agreement ran from 0.93 to 0.96 across steatosis, lobular inflammation, hepatocellular ballooning and fibrosis. Intersite agreement was 0.91 for ballooning, while steatosis (0.86), lobular inflammation (0.85) and fibrosis (0.87) had confidence intervals falling slightly below the study's 0.85 acceptance criterion. Scanner and site are the conditions that move the input in pathology, and the paper states where the algorithm fell short on them.

What is not published is accuracy compared across patient groups, scanners or stain protocols for any product, and no governance framework or AI management certification is named. The image management clearance speaks to regulatory status, not to model behavior across groups.

The company stands behind a published accuracy figure for its lead algorithm, and nothing further. PathAI's validation paper for AIM-MASH (Pulaski and colleagues, Nature Medicine 2025, PathAI led and cited in the company's FDA qualification announcement of 9 December 2025) compares the locked algorithm, without pathologist review, against a consensus ground truth set by panels of hepatopathologists on 1,481 liver biopsies from four completed global MASH trials. Measured as linearly weighted kappa against that ground truth, the algorithm's agreement exceeded the mean of individual pathologists by 0.15 for hepatocellular ballooning (95% CI 0.11 to 0.18) and 0.12 for lobular inflammation (0.08 to 0.17), and matched it within the non-inferiority margin for steatosis (0.01, -0.02 to 0.03) and fibrosis (-0.01, -0.04 to 0.02). These are results on retrospective reads of stored trial biopsies, not performance in use, and they cover AIM-MASH only; no error rate is published for the laboratory algorithms, which are research use only in the United States.

No warranty, indemnity or remediation commitment attaches to any output. In the AI Assist workflow a pathologist reviews every score and a second pathologist reviews disagreements of two points or more.

An error figure with its method and denominator, adopted by the vendor, with nothing contractual behind it.

Integration with the laboratory information system is claimed through PathAI's own connector, with no system named. The image management pages say AISight is interoperable with most major laboratory information systems and with Hamamatsu, Leica, Roche and Philips scanners, and that AISight Link, PathAI's open API, powers LIS integration and embeds partner algorithms in the viewer. The pages name no LIS product, no standard such as HL7 or DICOM, and no integration documentation. Named deployments exist (Labcorp, MedStar Health, Moffitt Cancer Center, University Hospital Zurich), but none is described as a named LIS connection. For a pathology platform the laboratory information system, not the patient chart, is the correct surface.

Integration claimed through a middleware layer with no system named.

One hosted service with its location stated, and the processing path open. AISight and AISight Dx are described as cloud-native and browser accessible, with cloud-based image and data storage that needs no additional infrastructure. The privacy notice (effective 13 February 2025) states that PathAI's platforms are operated and hosted from the United States, and that information collected from people outside the United States is transferred to, processed and stored there, with standard contractual clauses for transfers from the EEA. The hosting provider is not named on the product pages, no tenant isolation model is published, and no on premises option is described. For customers of the CE-IVD edition in the EEA, the UK and Switzerland, that statement places data in the United States unless a contract says otherwise, and the product pages do not address it.

A hosted option with its location committed and isolation left open.

No pricing is published for any part of the portfolio. The product pages carry no license fee, per case or per slide rate, tier structure or worked example for AISight, AISight Dx, PathAI's own algorithms or the partner algorithms offered through the platform, and commercial contact runs through a demo request or a dedicated sales address.

The sales structure, by contrast, is laid out: the image management system for laboratories (a research edition and the cleared AISight Dx), PathAI's algorithms on the same menu as partner algorithms from Deep Bio, DoMore Diagnostics, Mindpeak, Paige, Primaa, Stratipath and Visiopharm, and separate biopharma services covering clinical trial operations, real world data and companion diagnostic development.

Which items are clinically deployable is published: AISight Dx is cleared, and PathAI's algorithms are research use only in the United States. Nothing says whether platform access is a condition of algorithm access, or who bills partner algorithms. Roche announced a definitive agreement to acquire PathAI on 7 May 2026, which PathAI's home page still describes as pending.

Two customer bases, with validation behind part of each. AISight Dx is cleared for primary diagnosis in the United States with four named scanners (Hamamatsu NanoZoomer S360MD, Leica Aperio GT 450 DX, Roche VENTANA DP200 and DP600) and CE-IVD marked in the EEA, the UK and Switzerland; the research edition is described as scanner agnostic across Hamamatsu, Leica, Roche and Philips. More than 100 institutions use AISight, across academic medical centers, reference laboratories, health systems and independent laboratories. AIM-MASH is validated for MASH trial histology on biopsies from four completed global trials, with scanner and site repeatability reported. The biomarker algorithms name their indications and supported antibody clones (PD-L1 in non-small cell lung, head and neck, melanoma and urothelial cancer; HER2, ER, PR and Ki-67 in breast), and the menu extends to ulcerative colitis and dermatopathology, but these are research use only in the United States with no validation named on the product pages.

Coverage named with validation behind part of it.