Deciphex

The models sit inside businesses that stand without them. Diagnexia is a diagnostic reporting service delivered by a network of more than 250 pathologists, and Patholytix is an image management platform for preclinical and research pathology, with GLP workflows, audit trails and archiving. CipherX, the company's AI engine, runs inside both: in Diagnexia as tools for image quality, case triage and review after a report is authorized, and in Patholytix as decision support and research analysis. The company states CipherX is not offered as a licensed product. Remove the models and the reporting service and the image platform remain.

Artificial intelligence is a feature layer on products whose value stands without it.

The review structure is described, with the flag threshold and the follow-up unpublished. In Diagnexia every case is signed out by a pathologist, and the company states that the CipherX tools sit outside the direct diagnostic pathway: image quality assessment, complexity-based triage, and a review after authorization that points the pathologist to tissue it flagged. It also states that every CipherX output decomposes into named tissue structures a pathologist can inspect and reject. In Patholytix the AI supports primary evaluation and peer review by a study pathologist. Nothing says what confidence level raises a flag, what happens when the review after authorization disagrees with a signed report, or whether a flag can lead to an amended report.

The oversight structure described, with the threshold and the handling of a disagreement missing.

The architecture is described and the company's own foundation model named, without versioning or an update record. CipherX combines several foundation model encoders, including Deciphex's own DCX-3 fusion model, choosing one per task and per the rights attached to each dataset, and adds a semantic layer that breaks the model's representations into recurring units a subspecialist pathologist has confirmed and named, which compose into signatures such as tumor-infiltrating lymphocytes. The company says DCX-3 was evaluated on the THUNDER and EVA public benchmark suites against named models, and that swapping the foundation model beneath leaves the named signatures unchanged. The other encoders are not named, the training data are described only as the company's clinical, preclinical and research case flow plus partner datasets, and no release history or change notice is published.

Approach named without the version and update discipline.

The architecture is described, with the outside model providers unnamed. CipherX combines several foundation model encoders, one of them Deciphex's own DCX-3, and selects among them per task and per the rights attached to each dataset, so models from other developers may sit between a slide and a result without being identified. The privacy policy lists categories of outside processors, including cloud hosting, pathologists and third party laboratories, without naming any. Training data are described as coming from Diagnexia's clinical casework, Patholytix research work and partner projects including the BigPicture consortium and Novartis.

Architecture described with no model provider named.

Named customers and vendor-reported figures with no method or denominator. The company reports negative predictive values of 99.85 percent for adenocarcinoma and 98.76 percent for melanoma for CipherX tools on routine Diagnexia casework, without case counts, prevalence or a reference standard.

It reports that models developed with Charles River cut pathologist review time in lead optimization and dose range finding studies by over 20 percent, and that pathologists work up to 40 percent faster, a ceiling rather than a result. Named clinical customers include Bedfordshire and East Sussex NHS trusts, with a case study on reducing reporting backlogs, and on the research side Charles River and Novartis. There is no peer reviewed evaluation of the AI tools. The laboratory accreditations the service holds describe the laboratory, not the models.

General assurances, with the training use of clinical cases stated but not governed in public. The group privacy policy says the company processes protected health information as a processor on its clients' instructions, uses pseudonymization and anonymization, keeps a retention schedule, and ties AI processing to the original purpose and legal basis of the data. The company also states that real-world diagnostic data gathered through Diagnexia are integral to training its models, and that more than 147,000 cases a year flow through its clinical and research workflows. Nothing published says how that clinical data is de-identified before training, which client agreements permit it, what retention applies to slides and reports, or whether a client can opt out. No safety engineering is published.

General assurances that do not answer what reaches a model.

Processor status and HIPAA compliance are stated in a policy whose declared scope is website visitors and staff. The group privacy policy, which covers Deciphex, Diagnexia UK, Diagnexia US, Diagnexia Canada and Patholytix, says the company acts as a processor for protected health information when it reports cases for clients, processes medical data in line with HIPAA and UK and EU data protection law, and names a combined data protection and HIPAA privacy officer. The policy describes itself as covering website visitors and staff, and Diagnexia's own site policy covers the website only. No business associate agreement, data processing agreement or statement of the Diagnexia US service's position under HIPAA is published.

A status stated without a document reaching the service.

A recognized certification named without the artifact. The CipherX page lists ISO 27001, and Diagnexia states ISO 27001 certification for information security management issued by BSI. The Diagnexia Oxford laboratory holds UKAS accreditation to ISO 15189, the service reports College of American Pathologists accreditation and Care Quality Commission registration, and Patholytix describes immutable audit trails and single sign-on; those are quality and conduct standards rather than security audits. No certificate, scope or expiry is published, there is no trust center, and it is not stated whether the certification covers Deciphex, Diagnexia and Patholytix alike.

A recognized certification named, with scope to raise.

No device claim; the clinical tools are positioned as outside the diagnostic pathway. The company states that CipherX does not replace pathologist judgment and that its production tools sit outside the direct diagnostic pathway, running image quality checks, triage and a review after the report is authorized, even though it reports detection performance for adenocarcinoma and melanoma on routine casework. Diagnexia is regulated as a laboratory service, with UKAS ISO 15189 accreditation and College of American Pathologists accreditation, and Patholytix is validated for GLP primary and peer review in regulated preclinical studies. FDA's databases list no Deciphex device. Whether detection tools used on every clinical case remain outside device regulation is a question the company does not address.

No device claim, with the tools scoped as quality checks.

Governance is asserted, with no evaluation by group. The CipherX page describes the engine as compliant for high-risk use without saying which regime or what that covered, and the company's privacy policy describes an acceptable use policy for AI under the EU AI Regulation and a steering committee approving generative AI tools used internally. The company stresses that CipherX signatures are validated by pathologists and that each output can be traced to named tissue structures. No results are published by tissue type, species, stain, scanner, laboratory or patient group for any tool, including the adenocarcinoma and melanoma tools in routine clinical use.

Governance committed to in policy language with no evaluation behind it.

Outputs can be traced and challenged, and nothing stands behind them. The company states that every CipherX output decomposes into named tissue structures a pathologist can inspect and reject, Diagnexia reports are signed out by pathologists, and Patholytix keeps immutable, user-linked audit trails for regulated studies. The negative predictive values reported for the clinical tools have no case counts, prevalence or reference standard, so they do not stand as an error rate a buyer can test. No warranty, indemnity or remediation commitment attaches to an AI output, and nothing says what happens when a post-authorization flag was missed or wrong.

Mechanisms to challenge an output, with nothing standing behind it.

Integration is claimed without a system named. Diagnexia says completed reports can be integrated back into a client's laboratory information system and that its platform integrates directly with NHS systems, and Patholytix supports single sign-on and describes itself as the image management system for studies. No laboratory information system, record system, standard or interface is named, so a buyer cannot check how orders, images and reports move between Diagnexia and the hospital.

Integration claimed with no system named.

Options are stated for the research platform, with location open for the hosted routes. Patholytix is available as software as a service, private cloud or on premises, which the company says is designed to meet residency and validation requirements; on premises keeps images at the customer. No host, region or tenant isolation is named for the service or private cloud routes. Diagnexia's slides are scanned on site or shipped to the company's scanning laboratories, including Oxford, Exeter and Toronto, and reported remotely by pathologists across the United Kingdom, United States and Canada, but where images and reports are stored is not stated.

Options stated with residency committed only for the on premises route.

No price is published; how each business sells can be established. Diagnexia sells reporting as a service, with stated turnaround of three to five days for primary reports and seven to ten days for consults, and positions itself against locum pathologists, whose cost it puts at about $3,000 a day without stating its own rate. Patholytix is sold as a platform in three deployment models, with scanning, research services and data licensing alongside, and CipherX reaches customers only through those routes, not as a license. No per case, per study or platform rate is published.

The sales posture is discoverable, with no price.

Two markets named, with evidence behind parts of each. On the clinical side Diagnexia provides subspecialty reporting for NHS trusts and laboratories, with primary reporting in the United Kingdom and Canada and a diagnostics service in the United States, and CipherX tools run on its adenocarcinoma and melanoma casework. On the research side Patholytix supports exploratory, GLP and GCP pathology across species, with models for lead optimization and dose range finding studies developed with Charles River. Validation is stated for Patholytix in GLP primary and peer review and, without a method, for the two clinical tools; the other applications on the CipherX page are described without results.

Coverage named with validation behind part of it.