Techcyte

Deep learning classification of cells, parasites, bacteria, and cancer on digitized slides is the product, and the company frames the workflow platform as built around the algorithms rather than the reverse. The president's own description is unusually precise about the division of labour: the AI analyzes pixels and proposes high confidence targets, and the workflow adapts those proposals to how the lab already works.

No oversight design is published, and the product category makes the omission specific rather than generic.

The classifiers identify parasitic cells, bacteria and abnormal cells in preparations that a human would otherwise examine field by field. That is where the efficiency comes from, and it is also the shape of the risk: the value of a cell classifier is that the pathologist or technologist no longer looks at everything, so what a human never sees is determined by the model. A blood smear classifier that sorts cells into categories, or a cytology tool that surfaces suspicious fields, is making a selection decision before human review begins rather than after it.

Nothing published states how that selection is presented or bounded. Whether the reviewer sees the full preparation or only flagged fields, whether a proportion of negatives is re examined, what confidence information accompanies a classification, whether the reviewer can tell how many objects were assessed and discarded, and what quality control detects drift in what the model stops surfacing.

The research use only status of the human products in the United States is relevant context and not an answer. It constrains the claims that can be made; it does not describe how the tool behaves in a laboratory using it.

Ask what a reviewer sees and what they do not, whether any negative sampling occurs, and how a laboratory would detect that the classifier had begun missing a category.

Training data provenance is unusually specific: a long standing Mayo Clinic collaboration providing access to the Safe Harbour dataset of more than 17 million de-identified slides paired with pathology reports. Naming the institutional source and the scale lets a buyer reason about representativeness.

Held back from A because no published accuracy, sensitivity, or specificity figures were retrieved for any classifier, which matters more given the Research Use Only status, since a laboratory cannot rely on regulatory review to have examined performance.

The corpus is named at its source and scale, which is a real disclosure, and it concentrates every question this axis asks into one arrangement. The company has access to a dataset of more than seventeen million de identified slides and pathology reports supplied by a major academic medical centre, used to improve the models in its platform, and the arrangement is disclosed rather than hidden. Three things follow. The first is technical and specific.

Whole slide image files routinely embed a photograph of the slide label, which commonly carries the accession number and often the patient's name, and at a scale of seventeen million slides de identification is necessarily automated, so whether the embedded label and macro layers were stripped is a question about the pixels rather than about the policy.

The dataset's name refers to a de identification standard; the method actually applied to the image layers is what matters, and a standard's name in a dataset title is not evidence that every layer was processed. The second is what patients understood.

Records contributed to an institutional research resource and later licensed to a commercial partner for product development is a familiar arrangement and a lawful one, and it is also one most patients have never been told about in those terms. The third concerns customers rather than the dataset: nothing states whether slides processed by laboratories using the platform contribute to model improvement. Ask how image layers were de identified and verified, and for a written customer data position.

Substantial and verifiable adoption, held below the top of the axis because the most impressive evidence is not independent.

What is real: deployment of the anatomic pathology platform across the pathology practices of a major academic medical centre, announced jointly by both parties; a product range spanning parasitology, bacteriology, haematology, cervical cytology and bladder cancer surveillance; and a veterinary segment the company reports as already profitable, with human and environmental segments targeted for profitability the following year. Profitability in one segment is a harder fact than most adoption claims in this index, because it cannot be produced by pilots.

What complicates the reading is that the flagship customer is not only a customer. The same institution is an investor in the company's most recent raise, the licensor of intellectual property the platform is built on, a co-developer of the platform, and the supplier of a dataset of more than seventeen million de identified slides and reports used to improve the models. Each of those relationships is legitimate and none is concealed. Together they mean that the deployment is not an arm's length adoption decision by an independent buyer, and should not be read as one.

What is absent is published performance. No peer reviewed validation of any classifier was located, and the human products are research use only in the United States, which constrains what performance claims can be made at all.

Ask for validation data per product, and for a reference deployment at an institution with no commercial relationship to the company.

No stewardship framework or data governance terms were located, and this record carries the largest training data question in the lane.

The company has access to a dataset of more than seventeen million de identified slides and pathology reports supplied by a major academic medical centre, used to improve the models in its platform. That is a genuinely valuable asset and the arrangement is disclosed rather than hidden. It also concentrates every question this axis asks into one place.

The first is the de identification itself. Whole slide image files routinely embed a photograph of the slide label, which commonly carries the accession number and often the patient's name. At a scale of seventeen million slides, de identification is necessarily automated, and whether the embedded label and macro layers were stripped is a specific technical question rather than a general assurance. The dataset's name refers to a de identification standard; the method actually applied to the image layers is what matters.

The second is what patients understood. Records contributed to an institutional research resource and later licensed to a commercial partner for product development is a familiar arrangement and a lawful one. It is also one most patients have never been told about in those terms.

The third concerns customers rather than the dataset. Nothing states whether slides processed by laboratories using the platform contribute to model improvement, which is the question every customer should ask.

Ask how image layers were de identified and verified, and for a written position on customer data use.

The health privacy rule is named on the company's compliance line. No business associate agreement availability statement, scope description, contracting entity or subprocessor list was located, and naming a law is not describing a posture.

Three relationships need separating here, because this company sits in an unusual number of them at once.

As a platform supplier to human diagnostic laboratories, the ordinary business associate analysis applies, and the questions are the standard ones: which entity signs, what use of customer data the agreement permits, and how subprocessors are handled.

As a partner platform hosting third party algorithms, the chain extends. A laboratory's images reaching a partner's model means the partner is also handling protected health information, and whether that runs through the platform's agreement or requires a separate one with each partner is a question the laboratory should settle rather than assume.

As a recipient of a large de identified research dataset from an academic medical centre, the rule does not apply at all, because de identified data sits outside it. That is the correct analysis and it is worth stating explicitly, because it means the seventeen million slide corpus is governed by the licensing agreement between the two organisations rather than by privacy law, and a customer has no visibility into those terms.

Ask for the business associate agreement and its scope, how partner algorithms are covered, and whether customer slides can enter any research or development corpus.

The company's site carries a compliance line naming a regulator, a European mark, the United States health privacy rule, an unspecified international standard, European data protection law and a service organisation control report. No certificate, scope statement, report type, standard number or trust centre was located behind any of it.

How to read a list like that is the finding. The six items are four different kinds of thing. A service organisation control report is an attestation a company obtains. An international standard is a certification it holds. A European mark is a regulatory conformity route for a product. A privacy rule and a data protection regulation are laws you comply with, not credentials you hold, and a regulator is not a credential at all. Mixing them produces a line that reads as six accreditations and contains at most two. A credential list should be read item by item, asking of each what kind of object it is and who verified it.

One item deserves specific attention. The list names the United States regulator, while the company states plainly elsewhere that its human products are research use only in that market. Those can both be true, since the company operates veterinary and environmental lines and holds a European mark. But a laboratory buyer scanning the line would reasonably infer clearance the human products do not have, and the company's own candour about its regulatory status elsewhere makes the line unfortunate rather than deceptive.

Ask which standard and which report type, for the scope section, and whether the attestation covers the platform including partner algorithms.

The single most important fact for a US buyer, and the company states it clearly rather than burying it: the anatomic and clinical pathology platform and the human products, covering parasitology, bacteriology, hematology, cervical cytology, and bladder cancer surveillance, are Research Use Only in the United States. Fusion carries CE-IVDD marking in Europe.

Research Use Only is a permitted use restriction, not a stage of development: the product is real, sold, and generating revenue, but its outputs may not be used for clinical diagnostic purposes in the US. That also explains the segment economics, since veterinary diagnostics face no equivalent restriction and that segment is reported profitable while human is not. Graded on US clinical availability; the disclosure itself is a credit to the company.

One disclosed design choice deserves scrutiny. The company states that semi quantitative ranges for the peripheral blood smear classifier can be adjusted per customer to match their protocols and populations. Configurability is genuinely useful and honest to disclose, but it moves part of the clinical risk decision onto the laboratory: a threshold set for throughput behaves differently from one set for sensitivity, and no guidance, default rationale, or validation requirement for customer tuning was retrieved. Population adaptation is also raised as a rationale without any accompanying analysis of performance across populations.

Two passes located no published accuracy, sensitivity or specificity figures for any classifier, no evaluation methodology, no published limitations and no warranty, indemnity or remediation commitment. The regulatory status makes that absence weigh more rather than less, and the reasoning is worth stating because it recurs.

Products offered for research use only have not been reviewed by a regulator for clinical performance, so a laboratory cannot fall back on the assumption that someone else examined the numbers before the product reached them. For a cleared device a buyer who finds no vendor published performance can at least read a clearance summary; here there is no such document, and the vendor's own publication is the only possible source of evidence.

That inverts the usual position, where research use status implies lower stakes: it implies less external scrutiny, not less consequence, and a laboratory validating the tool itself needs a starting point the vendor has not given. The provenance disclosure credited on the other axis names the source and scale of the training corpus, which speaks to how the models were built and not at all to how they perform. Ask for sensitivity and specificity per classifier and per specimen type, the validation cohorts, and what a laboratory is expected to establish itself before clinical use.

Openness is an explicit commercial strategy rather than a feature. The platform connects to numerous slide scanners rather than requiring proprietary hardware, integrates with laboratory information systems, and hosts third party algorithms through the Fusion Partner Program. For a laboratory, not being locked to one scanner vendor is a substantive procurement advantage, and hosting external models means the platform can outlive any single algorithm. Buyers should note the corollary: models on the same screen may carry different regulatory status, as the Modella PathChat integration illustrates.

Delivered as a cloud service, with no hosting location, region, tenancy, retention or subprocessor terms located.

The architecture creates one question this index has not had to ask elsewhere in quite this form. The platform is deliberately open, connecting slide scanners, laboratory information systems and third party artificial intelligence developers through a partner programme, and that openness is genuinely the product's strength: a laboratory can reach algorithms from many suppliers without integrating each one separately. One such partner brought a research copilot onto the platform and was subsequently acquired by a pharmaceutical company.

That example shows the issue rather than being the issue. Where third party models run on a laboratory's slides inside a platform, the set of parties with access to that laboratory's images is determined by the platform's partner roster rather than by the laboratory, and it changes over time as partners are added, and as partners are acquired by organisations the laboratory never assessed. A laboratory that assessed a partner at signing may find the counterparty has different ownership a year later.

Ask where the platform and its image storage run, how tenants are separated, and specifically what governs partner algorithms: whether they execute inside the laboratory's tenancy or receive images, whether the laboratory approves each partner individually or accepts the roster, and what notification is given when a partner changes hands.

No public pricing. Contact the vendor. Sold to human, veterinary, and environmental laboratories, with the veterinary segment reported profitable and human and environmental targeted for 2027. US buyers should factor the Research Use Only restriction into any business case, since a product that cannot inform clinical diagnosis has a materially narrower internal use case than the same platform in Europe.

Unusually well bounded for a platform with this much range, and the boundaries are drawn by the company rather than inferred.

The scope spans three distinct laboratory markets, human, veterinary and environmental, and the company says so rather than presenting itself as a healthcare company that happens to have other lines. Within human diagnostics the product areas are named individually: parasitology, bacteriology, haematology with a peripheral blood smear classifier, cervical cytology, and bladder cancer surveillance. The platform's own positioning, unifying anatomic and clinical pathology in one workflow, is a scope claim that is checkable rather than aspirational.

What earns the top grade is that the regulatory boundary is stated alongside the functional one. The company states plainly that its anatomic and clinical pathology platform and its human products are research use only in the United States, while the platform carries a European mark. Very few vendors in this index volunteer the limit on where their product may be used clinically, and fewer still do it in the same breath as describing the product's reach. A buyer can determine what they may actually do with it in their own jurisdiction without asking.

One consequence follows for readers of this record. The veterinary segment is the commercially mature one, and evidence or adoption from that segment does not transfer to human diagnostics, where the regulatory position and the consequences of error are different. Treat the three markets as three products.