DeepHealth

Two businesses sit under one brand and they have different amounts of model in them. The population health suites are model led in the way this index treats as central: SmartMammo Dx, the chest, prostate, neuro and thyroid tools and the breast ultrasound module exist only because the model exists, and nothing in them functions without it. The enterprise imaging half is different. Diagnostic Suite and Operations Suite are a picture archiving system and a radiology information system, inherited largely from eRAD, with AI layered onto them. Scheduling, registration, billing and image management are conventional software that this company would still sell if every model were switched off.

That mix is the honest reason this is not the top band. A buyer purchasing Operations Suite is buying workflow software with AI features; a buyer purchasing Breast Suite is buying a model. The published material presents both under one portfolio narrative, so the distinction has to be made by the buyer rather than read off the site. Ask which line items on your quote would still be delivered if the clinical AI modules were removed, and what each of those line items costs on its own.

The autonomy runs in the conservative direction, which is unusual in this lane and deserves credit. The published workflow is a concurrent reading aid that localises lesions and assigns suspicion levels at finding and case level, and its most consequential automated act is routing at risk cases into a safeguard review where an additional breast imaging radiologist reads them. The model adds a human reader rather than removing one. Nothing is reported, recalled or cleared without a radiologist, and case prioritisation on the worklist changes reading order rather than reading requirement.

What holds this below the top band is precision of disclosure rather than design. The routing exists but the rule governing it is not published: no score scale, no calibration, no stated threshold at which a case enters safeguard review, and no statement of what proportion of examinations that is. ScreenPoint publishes exactly those figures, which is why it sits a band higher on an arguably more aggressive workflow. A buyer here can see that automation decides who gets a second read, and cannot see where the line sits or how many cases fall on each side of it.

Ask for the numeric threshold that triggers safeguard review, the proportion of examinations it selects at your case mix, and whether that threshold is configurable by the customer or fixed by the vendor.

The training corpus is quantified, which is more than most vendors here disclose: more than a million images, 100,000 cases and 8,000 biopsy proven cancers, described as spanning diverse populations. Peer reviewed work has also named a specific software version of the triage model and described its mechanism, evaluating each image independently and aggregating regions of interest into a single study level score from 0 to 1. Naming the version matters and few vendors do it, because a result is only reproducible against a stated build.

The operating characteristics that a clinician needs are not published. No score scale for the clinical product as delivered, no calibration statement, no threshold values, no architecture and no account of the populations the models were developed on beyond the word diverse. A radiologist handed a suspicion level has no published basis for knowing what that level represents in their own population, which is the gap that separates this record from the two graded higher in this lane.

The breadth of the portfolio compounds it, because nine acquired companies means multiple model lineages under one brand with no published statement of which is which. Ask for the model card, the version identifier and the calibration data for each specific module you are buying, and whether the modules share a common architecture or are inherited from separate acquisitions.

Two parties in the chain are identifiable and no more. A named AI infrastructure partner is disclosed as underpinning clinical AI performance and safety monitoring, and a public cloud marketplace is named as a deployment route for one module. Beyond those, nothing: no model or model family behind the generative reporting and impression features, no hosting arrangement for the platform as a whole, no sub processor list, and no statement of which components are built rather than licensed.

The portfolio structure makes this the weakest disclosure surface on the record. Nine acquired companies have been merged into one brand, and each arrived with its own models, its own infrastructure and its own vendor relationships. The product a buyer signs for is therefore a composite whose parts have different provenance, and the published material describes the composite while naming none of the parts. Newer reporting automation adds a further layer, because generated impressions and structured findings typically involve a language model that is not identified anywhere.

This is the axis where the whole lane grades poorly, so the grade is not a singling out. Ask for a current sub processor list, the identity and hosting location of any language model used in report generation, and which modules on your quote were built in house rather than acquired or licensed.

The volume of evidence is exceptional and the design is weaker than the volume suggests. The ASSURE study in Nature Health, November 2025, is prospective across 579,583 women at more than 100 community imaging sites in four states, comparing an AI driven workflow on 208,891 examinations against the prior standard of care on 370,692, and reporting a 21.6 percent increase in cancer detection rate, 5.6 against 4.6 per thousand, in a cohort of more than 2,400 cancers with no rise in the proportion of ductal carcinoma in situ. A companion analysis reported that general radiologists using the workflow reached detection rates comparable to fellowship trained breast imagers. Credit is due for two design choices: the charge for the enhanced workflow was waived during the study period specifically to remove self selection, and external academic investigators are named as co authors.

Two defects in the design hold it below the randomised comparisons this index treats as the top bar. The control is historical rather than concurrent or randomised, so secular change in equipment, staffing and recall behaviour cannot be separated from the intervention. And the intervention is a bundle: the workflow pairs the model with a safeguard review in which a second radiologist reads at risk cases, so the 21.6 percent belongs to model plus extra human read, not to the model alone. A buyer cannot purchase that number without also purchasing the second reader.

A third point is a limit on what the evidence transfers rather than a defect in how it was produced, and the distinction matters. The study sites are substantially the parent's own network. Results generated inside one owner's network, with that owner's equipment, reading staff, case mix and recall culture, do not establish what a different health system will see in its own. That is a question about external validity, not about the integrity of the work, and it applies to any vertically integrated operator studying its own deployment.

The company has itself said it hopes to run randomised trials to eliminate self selection, which is a fair statement of what is missing. Ask what the detection rate gain was for the model alone without safeguard review, and whether any published result comes from sites the parent does not own.

Retrieved in two passes: a privacy policy and terms of service, and nothing beyond them. No retention schedule, no encryption detail, no de identification standard and no stated position on whether customer images contribute to model development was located.

That last omission carries more weight here than at a standalone vendor, because the training corpus is described and the parent owns an enormous source of images. Published material states the mammography model was trained on more than a million images, 100,000 cases and 8,000 biopsy proven cancers across diverse populations, and RadNet operates more than 400 imaging centres performing millions of examinations a year. The obvious inference is that the parent's own patient population contributed, which may be entirely proper, and is not the point. The point is that a third party provider sending its images into the same pipeline has no published statement telling it whether its images are treated the same way.

The safeguard review adds a second clinician on at risk cases, which is a genuine safety property and is described. Ask in writing whether images processed under your contract are retained after the read and whether they may be used for model training, and require the answer as a contractual term rather than an assurance.

No business associate agreement language, no compliance page and no statement of covered entity or business associate posture was located in two retrieval passes, on the company's own pages or elsewhere. The framework applies squarely here, so the ordinary standard governs: the company is United States headquartered in Somerville, Massachusetts, and processes protected health information for third party providers, which makes it a business associate by function. A buyer entering a covered relationship has nothing to review before contracting.

The structure makes the question sharper than usual rather than routine. The same corporate group is both a covered entity, through RadNet's imaging centres, and a business associate to competing providers who buy the software. A provider handing images to DeepHealth is handing them to a subsidiary of an imaging operator that may compete with it in the same market, and the agreement is the only place the boundary between those two roles gets written down.

A second frame applies to part of the estate. Gleamer and CIMAR are European and United Kingdom operations, where European data protection law rather than United States health privacy law governs. Ask which legal entity signs the business associate agreement, and what it commits about access to your images by RadNet's imaging services business.

Two dedicated retrieval passes were run against this axis, one by company name and one by product name paired with certification terms, and neither returned an attestation. The public footer carries a privacy policy, terms of service, a carbon reduction plan and a modern slavery statement, and no trust centre, security page, certification badge or sub processor list sits alongside them.

The absence is more surprising here than at a small vendor, and that is worth stating rather than smoothing over. An organisation holding more than 2,700 customer contracts across more than 50 countries, selling archiving and information systems into hospitals, and publishing a modern slavery statement of the kind United Kingdom law requires of larger firms, has almost certainly been assessed repeatedly by customers and public bodies. Enterprise imaging buyers do not sign without a security review. So attestations very probably exist and are being disclosed privately under agreement rather than published, which is a commercial choice and not a control failure.

The grade records what a buyer can confirm before contacting sales, and that is currently nothing. Ask for the current attestation report and its scope, and confirm specifically whether the scope covers the acquired European platforms or only the United States estate.

The regulatory footprint is the largest in this lane by a wide margin. The company states 26 United States clearances and 22 European conformity markings across the combined portfolio, and several are individually named and dated: SmartMammo Dx, also known as Saige-Dx, for tomosynthesis interpretation, with clearance later extended from one mammography vendor's systems to another's; Saige-Q as a triage product; a prostate clearance alongside European marking; European marking for brain health, brain age and a lumbar magnetic resonance product; European marking for the remote scanning platform; and a breast ultrasound clearance in July 2026 supported by a multi reader multi case study of 16 board certified radiologists.

Held at B rather than A for the same reason the neighbouring record in this lane is held there, and the consistency is deliberate. The device counts are vendor published and were not verified against the regulatory clearance database in this pass, and no individual clearance numbers were retrieved. A published set of cleared devices is exhaustive and cannot be shaped by marketing, so this is a check that can be closed precisely rather than argued about, and closing it is what would move the grade.

Ask for the clearance number and date for each module on your quote, and confirm that the clearance covers the specific imaging hardware in your department rather than a different manufacturer's system.

The subgroup analysis that is missing almost everywhere in this lane is present and published here, which is the single strongest thing on this axis in the radiology records. The ASSURE study reported cancer detection rate across racial, ethnic and breast density groups and found no significant variation, and separately reported a 22.7 percent gain in women with dense breasts. Dense tissue is the central confounder in mammography and the group most likely to have a cancer missed, so a published result showing the workflow does not underperform there answers the exact question this axis exists to ask. It was also conducted in community imaging centres across a diverse population rather than an academic cohort, which is where the generalisation question usually fails.

The governance apparatus around that result is absent. No model card, no published bias monitoring commitment after deployment, no description of the demographic composition of the training data, and no governance structure or review body was located. A one time equity analysis in a study is not a standing control, and model behaviour drifts as populations and equipment change.

Ask what subgroup performance monitoring continues after deployment, at what interval it is reported to customers, and what the vendor commits to do if a disparity appears in your population.

No warranty, indemnity, service level or remediation commitment attaching to a missed cancer was located in two passes, and the terms of service published on the site address website use rather than clinical software performance.

One structural feature makes the question unusually pointed. Under the enhanced detection programme the patient has paid out of pocket, directly, for the AI component of her screening. That creates a consumer who has bought a specific expectation of improved detection and who sits outside the provider to vendor contract entirely, with no published statement of what she is owed if the workflow misses a cancer that the published detection rate implies it should have caught. Most vendors in this index never face that question because no patient ever knowingly buys their product.

The partial mitigation is real and should be credited: the safeguard review means a difficult case gets a second qualified reader, so recourse is partly designed into the workflow rather than left to contract. Ask what the vendor warrants about model performance in your deployment, who carries liability for a false negative on a case the model scored low, and what the customer is told about the AI component if that cost is passed to patients.

Interoperability is not a feature of this product, it is the product. The company sells the picture archiving system and the radiology information system themselves, through Diagnostic Suite and Operations Suite, and states that they are designed to integrate with or replace whatever is currently installed. That places it in a different position from a detection vendor negotiating for worklist access: the worklist, the archive and the reporting layer are all things it ships.

The named integrations are specific rather than generic. Clinical AI runs against mammography systems from two major hardware manufacturers, with clearance extended across them; the remote scanning platform is deployable through a public cloud marketplace; advanced visualisation is integrated through a named third party partner; and a named AI infrastructure partner underpins clinical AI performance monitoring. Processing latency for flagging suspicious cases is published as under 5 minutes with one mammography vendor at a stated bandwidth and under 15 minutes with another, which is a specific operational commitment rather than a claim.

The honest caution is scope: this is imaging infrastructure depth, and no equivalent depth of integration with the wider electronic health record was retrieved. Ask how results reach the ordering clinician outside radiology, and what the migration path and data extraction terms look like if you replace your existing archive and later want to leave.

The deployment architecture is stated clearly, which is more than the neighbouring records manage. The platform is described consistently as cloud native, with a hybrid option available, and one module is offered through a public cloud marketplace, which identifies at least part of the hosting arrangement. Operation across more than 50 countries and more than 300 screening sites means residency has been settled repeatedly in practice.

None of those settlements is published. No region list, no residency commitment, no retention schedule and no statement of where images are processed against where they are stored was located in two passes. That gap is material for a portfolio assembled from European and United Kingdom acquisitions alongside a United States parent, because a French, Dutch or British operation folded into a United States owned cloud platform raises a cross border transfer question that the buyer, not the vendor, currently has to raise.

A hybrid option is also only meaningful if its limits are known, since some modules are typically cloud only. Ask which specific modules can run in your own environment and which require the vendor cloud, and in which country images are processed and stored for each one.

More of a price exists in public here than for almost any vendor in this lane, and none of it is the price a provider would pay. Two figures are established: patients at RadNet centres have been charged about 40 dollars out of pocket for the Enhanced Breast Cancer Detection programme, with roughly 36 percent electing it, and the breast ultrasound module is positioned against an existing Category III procedural code for quantitative ultrasound tissue characterisation, which names a reimbursement route rather than a price. As a public company RadNet also discloses digital health segment figures, including more than 16 million dollars of new third party total contract value in the first quarter of 2026.

What is absent is the business to business number. No price, no unit of sale and no mechanism was located for DeepHealth OS, Diagnostic Suite, Operations Suite or any clinical suite sold to an outside provider. The consumer figure is a poor proxy, because it is the parent's retail charge at its own centres and includes the safeguard radiologist read. The gap matters unusually here because a buyer can calculate the value precisely: detection rate gain per thousand examinations against local screening volume is an afternoon's work once a unit price exists. Ask for the price per examination for the clinical suites and separately per seat or per site for the enterprise imaging suites, and establish whether the quoted rate changes once RadNet is a competitor in your market.

The broadest coverage of any record in this lane, and it is earned on modality and indication rather than on positioning. Clinical AI spans breast, chest and lung, prostate, brain and neuro, thyroid and, through Gleamer, musculoskeletal, across radiography, mammography, tomosynthesis, ultrasound, computed tomography, magnetic resonance and positron emission tomography. The enterprise imaging layer sits underneath all of it, so the same buyer can take detection, reporting, archiving and operations from one supplier.

The institutional spread matches the clinical spread: a combined installed base with Gleamer of more than 2,700 customer contracts across more than 50 countries, more than 300 screening sites, and more than 10 million mammograms a year. Named third party customers run from a national teleradiology provider serving more than 120 hospitals to single market radiology groups, which is a wider range of buyer size than most vendors here reach.

The caution attached to breadth is that clearance is not uniform across geographies, and the company states plainly that not all products are commercially available in all countries. Ask for the list of modules cleared and commercially available in your specific country, and which of them run on the mammography and ultrasound hardware you already own.