Pramana

The models are what make the scanner autonomous, and autonomy is the entire commercial proposition, so this grades well above the hardware records elsewhere in the index.

The distinction that matters is against the handheld ultrasound cluster, where a capable instrument remains if the models are removed. Remove the models here and what remains is a conventional slide scanner requiring a trained operator to set parameters per slide and a human to inspect every output for focus, stitching and artifact errors. That operator cost is precisely what the service business is sold to eliminate, and the Mayo result, 23,916 slides in 30 days with a single human operator, is a direct measure of how much labour the models displace.

The models are doing several distinct jobs. Scan parameter selection replaces human judgement about focus and z plane, with in line decisions taken on tissue type, thickness and artifacts such as annotations and debris. Quality models assess every slide in real time for named error classes and trigger autonomous rescans. Edge models execute feature detection, quantification and classification during the scan itself.

What holds this at B rather than A is that the product is genuinely also an optical instrument. The company describes hardware and software co design, holds granted patents on a volumetric imaging method replicating microscope fine focus, and a customer is buying optics, mechanics and throughput alongside the intelligence. The models make the hardware autonomous rather than constituting the product on their own.

The autonomy is placed upstream of clinical judgement, which is the right place for it, and one consequence of that placement goes unexamined.

What the models decide is how to image, not what the image means. Scan parameters, focal plane, rescan decisions and quality annotations all occur before a pathologist opens the case, and the diagnosis remains entirely human. Compared with the analysis vendors elsewhere in this lane, whose models score tumour grade or detect malignancy, the autonomy here carries far less direct clinical weight.

The oversight surface is also unusually good. Quality is not merely judged, it is exposed: tile based feature overlays show focus and stitching quality across the slide at a glance, and the assessment can be reduced to a numeric metric feeding in line or downstream quality control. A pathologist can see where the image is weak rather than trusting that it is uniformly adequate. Publishing quality as a visible artefact rather than a pass or fail is the strongest oversight design encountered in this session.

The unexamined consequence is what happens when the quality model is wrong in the permissive direction. A slide passed with an undetected artifact becomes an image a pathologist diagnoses from, believing the scan was verified, and the failure is silent. That is the inverse of the usual concern about automation and it is specific to putting models at the acquisition step.

A second point deserves recording. The company states that novel visualisation and image enrichment can reduce inter observer variability between pathologists. Changing how tissue appears in order to change how readers interpret it edges from acquisition into interpretation, and nothing describes what enrichment does to the image.

Graded B.

The best mechanism disclosure of any vendor built in this session, sitting alongside no performance disclosure at all.

What is described is genuinely technical rather than promotional. Volumetric imaging is explained as replicating the fine focus function of an optical microscope, with fast focus sampling and dynamic calculation of the tissue z plane rather than a fixed focal plane. In line decisions are attributed to specific inputs: tissue type, thickness, and slide artifacts including annotations and debris. Quality assessment names its error classes, focus, stitching, bubbles and folds, and describes the output form, tile based feature overlays abstracted to a numeric quality metric that can drive in line or downstream quality control. Edge computation is described as executing feature description algorithms in real time during scanning so results are available at scan completion. Granted United States patents are cited for the underlying method.

A reader finishes that with an accurate mental model of how the system works, which is rare anywhere in this index and unheard of on a hardware record.

What is entirely absent is how well any of it works. No accuracy, sensitivity or specificity for any error class, no training data description or provenance, no evaluation method, no versioning and no update cadence. On a scanner deployed for archival projects spanning months, whether the quality models changed partway through a digitisation run is a reasonable question and nothing addresses it.

Graded B: mechanism explained unusually well, performance not disclosed at all.

Corporate provenance is unusually traceable and technical provenance is not, which is the reverse of the usual pattern in this index.

The corporate chain is fully documented and checkable. The company was founded in 2021 by a named health data analytics business, took venture funding with an investor named, and has since been acquired by an established life science imaging manufacturer under whose catalogue the scanners are now sold with the original brand intact. A buyer can follow the ownership, and the acquisition matters technically as well as commercially, since scanner optics and manufacture now sit inside a company with decades of microscope engineering behind it.

The technical claim is in house and is made through patents rather than assertion. Granted United States patents are cited for the volumetric imaging method, hardware and software are described as co designed, and the algorithms are characterised as proprietary. A patent is a genuine provenance artefact because it is public, examined and attributable, and it distinguishes this from vendors who simply decline to say.

What is missing is everything about the models as models. No training data description, no statement of what corpus the quality and feature models were built on, no acknowledgement of any open source framework or pretrained component, and no bill of materials for the scanner's embedded software. The training corpus question has weight here given the company's origin in a health data business and its access to a large academic tissue archive: whether the quality models were trained on that archive is unstated and would be worth knowing.

Graded C.

One operational study of genuine quality, from a named institution, with method detail, and no clinical performance evidence at all.

The Mayo Clinic engagement is the substance. The institution first evaluated quality, ease of use and throughput, then ran a competitive process and awarded a multi million slide archival contract, which is procurement evidence rather than a testimonial. The published result reports 23,916 slides digitised over 30 days by a single operator, approximately 800 a day, on a four head system, with slides drawn from the pathology tissue archive spanning the 1950s to the present and no cleaning or preparatory steps applied. Each slide was annotated in real time for focus, stitching, bubbles, folds and other errors.

The method detail is what earns the grade. Stating the operator count, the date range of the material, and the absence of slide preparation makes the throughput figure interpretable rather than promotional, and archival slides from the 1950s are the hardest case, so a result obtained on them is a strong result. The company also situates its work against the existing peer reviewed whole slide imaging literature rather than ignoring it.

What is absent is any evidence about diagnosis. No study establishes that images produced by these scanners support diagnostic accuracy equivalent to glass or to other scanners, no concordance study was located, and the quality models have no published sensitivity or specificity for the error classes they detect. The United States research use only status means no such evidence has been required there yet.

Graded B: exceptional operational evidence, no clinical evidence.

The secondary use question is more concrete on this record than on any other in this session, because the company markets the secondary use itself and the corporate history points the same way.

The service material describes the output of digitisation as a whole slide image repository and states its value in terms a buyer should read carefully: enabling new research revenue streams, supporting artificial intelligence diagnostic development, and streamlining educational opportunities. That is an explicit statement that digitised archives have downstream value beyond the diagnostic use they were created for. It is framed as value accruing to the institution, and nothing published states whether the vendor retains any interest in the images it produces, any right to use them, or any role in the research access it describes.

The corporate provenance sharpens the question rather than answering it. The company was founded by a health data analytics business whose model is deriving insight from clinical data held by health systems, and this company's flagship engagement digitised a major academic centre's tissue archive spanning seventy years. The material at stake is decades of patient tissue collected long before anyone contemplated computational analysis, and the consent basis for its digitisation and any subsequent use is exactly the issue that has generated litigation and regulatory attention elsewhere in health data.

Nothing published addresses de identification standards, label identifier handling, retention, whether images inform the vendor's own model development, or who may access a repository once built.

Graded D on the absence. Publishing a plain statement that the vendor claims no rights in customer images, if that is the case, would be the single most valuable disclosure available here.

No published position was located, and the service model makes that a larger gap than it would be for a software vendor.

Under Digital Pathology as a Service an institution physically ships glass slides to the vendor, which digitises them and returns images. Glass slides carry accession numbers and often patient identifiers on the label, the label is typically captured in the scan, and the resulting whole slide images are protected health information held on vendor infrastructure. So this vendor takes custody of both the physical specimen and the digital record, which is a more complete transfer of patient material than any other record in this session involves.

Nothing published addresses any of it. No business associate agreement is offered or described, no protected data handling summary exists, no chain of custody description for shipped slides, no statement about label capture or identifier handling, no retention position for images held after a project completes, and no deletion or return commitment at contract end.

The scanner only deployment is a narrower question, since images stay in the customer's environment, though the company describes remote viewing and case sharing which implies some vendor mediated path.

One related fact is documented and is a partial answer of a sort: the Mayo relationship was substantial, competitively awarded and multi year, and an institution of that kind does not ship its tissue archive without an executed agreement and a privacy review. That is inference about one customer rather than disclosure to the next.

Graded D on published evidence.

No published security posture was located. No trust centre, no service organisation control claim, no HITRUST claim, no independent audit reference, no penetration testing statement, no vulnerability disclosure policy and no incident notification commitment were found. The only security language located is a description of the desktop scanner as designed for information technology friendly deployment with modern connectivity and security practices, which is a design intention rather than an assurance.

The gap is wider than a software vendor's because the deployment has two surfaces. The scanners are networked laboratory instruments sitting inside a hospital environment, connected to image storage and receiving software updates, and networked medical laboratory devices are a recognised attack surface with the same characteristics that make hospital device security hard: long service lives, infrequent patching and operators who are not information technology staff. Nothing addresses update integrity, network requirements or hardening.

The service business is the second surface and it holds protected images on vendor infrastructure with no described control environment.

One mitigating consideration belongs on the record. The European in vitro diagnostic regulation imposes requirements touching software lifecycle and security for devices in this class, so material will have been produced for conformity assessment, and the parent is a large established imaging manufacturer likely to operate a corporate security programme. Neither is published for this product.

Graded D on published evidence rather than on any judgement about the underlying engineering.

An honest and clearly stated position that is incomplete in the largest market, and the honesty is worth as much as the status.

What is disclosed, on product pages rather than in a footnote: CE marked under the European in vitro diagnostic regulation, licensed by Health Canada, research use only in the United States, and a 510(k) submission pending. Every element of that is stated plainly. Vendors routinely blur research use only status or let a European marking imply global clearance, and this company does neither.

The European marking is more substantial than it may look. The in vitro diagnostic regulation replaced a far lighter predecessor regime and imposes conformity assessment through a notified body for devices of this class, so a marking under it reflects external scrutiny rather than self declaration. Health Canada licensing is an independent second regulator reaching the same conclusion.

The United States position is the limit and it is material. Research use only means a United States laboratory cannot use these images for primary diagnosis, so the domestic market for the scanners is research, education and archival digitisation rather than clinical sign out. The Mayo work is archival, which fits that constraint exactly. A prospective United States clinical buyer should understand they are buying into a pending submission with no stated timeline.

One question is unaddressed and follows from the acquisition. Regulatory registrations attach to a legal manufacturer, and nothing published describes how the submission and the existing markings are affected by the change of ownership.

Graded C.

Nothing was located, and the position of these models in the pipeline makes the absence propagate further than it would anywhere else in this lane.

Every analysis vendor in the pathology category consumes whole slide images produced by scanners like these. If the acquisition and quality models perform unevenly, the variation is baked into the images before any diagnostic model sees them, and a downstream vendor evaluating its own bias has no visibility into it. Acquisition bias is upstream of every other kind in digital pathology and it is the least examined.

The mechanisms are concrete. Slide preparation varies systematically between laboratories, in stain protocol, section thickness and coverslipping, and laboratories serving different populations differ in exactly those ways. A quality model trained predominantly on material from well resourced academic laboratories may flag more errors, trigger more rescans or pass fewer slides from a community or international laboratory whose staining differs, and the effect appears as a throughput or quality problem attributed to that laboratory rather than to the model. Tissue type and stain are a second axis: performance on routine haematoxylin and eosin sections says little about immunohistochemistry or cytology smears.

The archival case makes it sharpest. Slides from the 1950s through the present differ enormously in preparation and condition, and nothing published describes how quality model performance varies across that range even though the company's flagship engagement spanned it.

No model card, training data description, per error class accuracy figure or subgroup analysis was located.

Graded D on the absence, with the propagation point recorded because it is not obvious.

Nothing published addresses responsibility for an automated outcome, and this record carries a category of exposure no other vendor in this session does.

The distinctive one is physical custody. Under the service model an institution ships glass slides to the vendor, and archival slides are unique specimens: a block may be exhaustible or gone, the patient may have died decades ago, and a slide lost or damaged in transit or handling is unrecoverable evidence of a diagnosis. The company's flagship engagement moved nearly 24,000 slides spanning seventy years. Nothing published describes chain of custody, insurance, loss or damage liability, or what recourse an institution has for material that cannot be replaced.

The model specific exposure is quieter and follows from where the intelligence sits. A quality model that passes a slide with an undetected artifact produces an image a pathologist diagnoses from, in the belief that quality was verified. If the diagnosis is wrong because the image was inadequate, the causal chain runs through an automated quality judgement, and the pathologist had a false assurance rather than no assurance. Nothing published describes accuracy commitments for quality detection or where responsibility sits.

A third follows from format. An institution digitising an irreplaceable archive is making a long term bet on retrievability, and while the open format commitment mitigates this substantially, no continuity or escrow position is published.

No indemnity, limitation, performance warranty or service level was located, and the research use only status in the United States shifts responsibility to the customer there without addressing it elsewhere.

Graded D.

Open formats are stated explicitly, which matters more in digital pathology than almost anywhere else, and the specifics are thin.

The substantive commitment is to standardised and open data formats for the images produced. Proprietary whole slide image formats are the defining lock in problem of this field: an institution that digitises an archive into a vendor specific format has bound decades of irreplaceable material to one supplier's continued existence and goodwill. A vendor whose business is large scale archival digitisation, and which commits to open formats for the output, is addressing the single largest risk its own customers carry. That commitment earns most of this grade.

Around it the platform is positioned as a component rather than a destination: compatible with standard digital image storage and management solutions, supporting downstream integration with image analysis and artificial intelligence tools, enabling remote viewing and case sharing for consultations and tumour boards, and described as designed for information technology friendly deployment.

What is missing is specification. No standard is named, so whether the open format commitment means the recognised medical imaging standard for digital pathology or a documented proprietary format with published structure is unclear, and those are materially different promises. No laboratory information system integration is named, which is the connection that matters for routine clinical work, since a scanner that cannot reconcile a slide to an accession record creates manual reconciliation. No image management vendor is named as a tested partner.

Graded B on the strength of the format commitment.

Two genuinely different deployment models exist, one is described well and the other is not described at all.

The scanner model is the better documented and the architecture is good. Models run on the scanner itself, executing feature detection and quality assessment in real time during the scan rather than sending images away for processing, so the compute is local and the images stay in the customer's environment by default. Edge execution is a strong answer to residency questions because it removes the need for them, and the company describes deployment as designed for existing pathology information technology environments with modern connectivity and security practices.

The service model inverts that entirely and is undescribed. Under Digital Pathology as a Service the institution ships glass slides to the vendor and receives images back, so both physical specimens and digital records leave the customer's control. Nothing published states where scanning is performed, where images are stored during and after a project, in which jurisdiction, for how long, under what retention or deletion terms, or how the physical slides are tracked, secured and returned. For archival material the physical custody question is the sharper one, because those slides are unique and irreplaceable.

A third question follows from the acquisition. The company is now part of a larger imaging business with its own global infrastructure, and nothing describes whether service operations, storage or support have moved as a result.

Graded C on the strength of the edge architecture disclosure, with the service path entirely open.

The commercial model is explained clearly and no figure attaches to it, which places this above the pricing floor common in this index.

What is disclosed is the structure and the argument behind it. Digital Pathology as a Service is described as letting a centre pay for high quality images without capital budgets or hiring skilled staff, explicitly with no upfront capital expenditure, and the company names the costs it is displacing: scanner capital, the human capital to operate them, parameter selection and post scan qualification of every image. That is a genuine total cost of ownership argument stated in terms a laboratory director can test against their own budget, and the implied unit is per image or per slide.

The hardware alternative is also structured visibly, with a high throughput configuration and a desktop scanner positioned for different laboratory sizes, so a buyer can see there is an entry point.

What is missing is every number. No per slide rate, no scanner price for either configuration, no service minimum, no contract term and no implementation cost. The one figure in public circulation is the Mayo archival contract described as multi million, which indicates the scale of a large archival project and nothing about unit economics.

One question is specific to the service model and unaddressed. Archival digitisation is a finite project while routine clinical scanning is ongoing, and nothing describes whether pricing differs between them or what happens to stored images at contract end.

Graded C.

Broad sample coverage within pathology, two deployment sizes, and two real limits.

The sample range is the strongest element and it is wider than most scanner vendors claim: anatomic pathology, cytopathology, hematopathology, clinical pathology and microbiology. Cytology is the hard case, because smears are thick and irregular and defeat fixed focal plane scanning, and the company's volumetric approach with dynamic z plane calculation is specifically the technique that addresses it. The scanners are described as handling most slide types without user intervention, and the Mayo archival work demonstrates tolerance of degraded, uncleaned material from as far back as the 1950s, which is a coverage claim most vendors could not make.

Institutional coverage spans high throughput hospital laboratories, satellite and outreach sites needing local scanning, academic and teaching institutions building digital slide sets, and tumour boards requiring shared images, with the desktop configuration positioned for labs beginning digital adoption.

Two limits hold this at B. The scanners are brightfield, so fluorescence work sits outside this product and is served by a separate line in the parent's catalogue, which excludes a substantial share of research and some clinical immunofluorescence. And geographic coverage is uneven by regulatory status: full diagnostic use in Europe under the in vitro diagnostic regulation and in Canada, research use only in the United States. A United States laboratory can buy the coverage described here and cannot use it for primary diagnosis.

Graded B.