DeepHealth

Two businesses sit under one brand, with different amounts of model in them. The population health suites are model led: SmartMammo Dx (Saige-Dx), the density, risk and breast arterial calcification tools, and the chest, prostate, neuro and thyroid suites exist only because the models exist.

The enterprise imaging half is different: Diagnostic Suite and Operations Suite are a picture archiving system and a radiology information system, largely inherited from eRAD, with AI layered onto them, and scheduling, registration, billing and image management would still be sold if every model were switched off. A buyer purchasing Operations Suite is buying workflow software with AI features; a buyer purchasing the breast suite is buying a model.

The autonomy runs in the conservative direction. The published workflow, described in the ASSURE study (Nature Health, 2025), is a concurrent reading aid that localizes 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 worklist prioritization changes reading order rather than reading requirement. What is not published is the rule governing that routing: the score scale, its calibration, the threshold at which a case enters safeguard review, and the proportion of examinations it selects. The model's part in deciding which cases a second radiologist reads is visible; where it draws the line is not.

The approach, the training corpus and version numbers are published, and the operating characteristics are not. The mammography model is described as trained on more than a million images, 100,000 cases and 8,000 biopsy proven cancers across diverse populations, and published work names a specific version of the triage model and its mechanism: each image evaluated independently and regions of interest aggregated into a single study level score from 0 to 1.

The FDA database lists versioned clearances (Saige-Dx 3.1.0, K243688, December 2024; Saige-Density 2.5.0, K243705), so a result can be matched to a build. The company publishes neither the architecture nor the calibration and thresholds of the delivered score, and says nothing of the populations behind the models beyond the word diverse. The portfolio holds several model lineages from acquisitions (iCAD, Aidence, Kheiron, Quantib, See-Mode, Gleamer) with no statement of which module comes from which.

Some of the chain is identifiable. The clinical AI modules are DeepHealth's own cleared models (Saige-Q, Saige-Dx, Saige-Density and BAC under DeepHealth, Inc. in the FDA database), and models from acquired companies (iCAD, Aidence, Quantib, See-Mode, Gleamer) arrived with their own clearances, so the detection chain is short where a clearance governs it.

The company discloses an AI infrastructure partner for clinical AI performance and safety monitoring and offers the remote scanning platform through a public cloud marketplace. No subprocessor list, platform hosting provider or region is published, and the generated reporting and impression features name no model or model provider.

The headline claim has a prospective, multi site study behind it. The ASSURE study (Louis and colleagues, Nature Health, November 2025, doi 10.1038/s44360-025-00001-0) analyzed 579,583 women at more than 100 community imaging sites in four states, comparing an AI driven workflow (208,891 examinations) with the prior standard of care (370,692), and reported 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 general radiologists using the workflow reaching detection rates comparable to fellowship trained breast imagers.

The charge for the enhanced workflow was waived during the study to remove self selection, and external academic investigators are coauthors alongside RadNet's chief science officer. The control is historical rather than concurrent or randomized, and the intervention is a bundle, the model plus a safeguard review by a second radiologist, so the gain belongs to both. The sites are also substantially RadNet's own, so transfer to another network's equipment, staff and recall culture is untested. The company has said it hopes to run randomized trials.

General assurances are published, and the questions AI raises are not answered. The platform is described as cloud native with a hybrid option, and DeepHealth's subscriber privacy notice states that DeepHealth or its affiliates are ISO 27001 certified; that notice covers marketing contacts, and no retention schedule, de identification standard or statement on whether customer images contribute to model development is published for the products.

The mammography model is described as trained on more than a million images, and RadNet runs more than 400 imaging centers, so it is a fair inference that the parent's own patients contributed; a third party provider sending images into the same platform has no published statement of whether its images are treated the same way. Safety engineering, meaning guardrails and how a safety event is handled, is not published; the safeguard review is an oversight control, described under autonomy.

No business associate status and no privacy document reaching the products are published. The site's legal pages are a privacy policy, terms of service addressing website use, and a subscriber privacy notice that covers newsletter and event contacts and describes DeepHealth as the controller of that marketing data; none addresses patient data or images processed by the software.

DeepHealth is based in the United States and processes protected health information for third party providers, which makes it a business associate by function, and the same group is a covered entity through RadNet's imaging centers: a provider sending images to DeepHealth sends them to a subsidiary of an imaging operator that may compete with it, and the agreement is the only place that boundary gets written down. Gleamer and CIMAR are European and UK operations where European data protection law governs.

A recognized certification is named, with a scope question attached. DeepHealth's subscriber privacy notice states that DeepHealth or its affiliates are ISO 27001 certified. Which entities and which products the certificate covers is not stated, and no certificate, trust center, security page or SOC 2 report is published; the site footer links a privacy policy, terms of service, a carbon reduction plan and a modern slavery statement. The site also carries a meaningful use certification adherence page, which points to federal health IT certification of part of the software.

DeepHealth's clearances are identifiable by product and version. The FDA database lists ten 510(k) clearances under DeepHealth, Inc.: Saige-Q (K203517, April 2021, 21 CFR 892.2080, triage); Saige-Dx, sold as SmartMammo Dx (K220105, May 2022; K241747, November 2024; K243688, version 3.1.0, December 2024; K251873, August 2025; K253825, June 2026; 892.2090, product code QDQ, detection and diagnosis software for mammography lesions suspicious for cancer); Saige-Density (K222275, December 2022; K243705, version 2.5.0, December 2024; product code QIH); TechLive (K243703, June 2025, LLZ); and BAC for breast arterial calcification (K254131, May 2026, QIH).

DeepHealth received 510(k) clearance for Chest XRay on 18 September 2026, the next version of the chest radiograph technology it acquired with Gleamer. The company's count of 26 US clearances and 22 CE markings covers the whole combined portfolio, including acquired lines cleared under their original companies' names (iCAD, Aidence, Quantib, Gleamer).

A subgroup evaluation is published with its groups and a headline result, and the governance around it is not. The ASSURE study (Nature Health, 2025) reported that the gain in cancer detection did not vary significantly across racial, ethnic and breast density groups, and reported a 22.7 percent gain in women with dense breasts, the group in which a cancer is most likely to be missed.

The results for the other groups are summarized as no significant variation rather than published group by group. No model card, no demographic composition of the training data, no post deployment bias monitoring commitment and no governance body is published, and a one time analysis in a study is not a standing control.

Mechanisms exist in the workflow and nothing stands behind the output. The safeguard review routes at risk cases to a second qualified reader, and every study is read by a radiologist. No warranty, indemnity, service level or remediation commitment attaching to a missed cancer is published, and the terms of service address website use.

The SmartMammo Dx (Saige-Dx) 510(k) summaries, the latest in June 2026 (K253825), are the company's own filings where a standalone error rate would sit, and none is cited on the pages that describe the product. One structural feature makes the question pointed: at RadNet's own centers patients have paid about $40 out of pocket for the enhanced detection program, so a patient has bought an expectation of improved detection while sitting outside the provider and vendor contract.

DeepHealth sells the radiology record systems themselves, and being the record is not the same as exchanging with others. Diagnostic Suite (a picture archiving system) and Operations Suite (a radiology information system covering scheduling, registration and billing) are designed to integrate with or replace what is installed.

Exchange with systems outside the product is named in part: clinical AI runs against mammography systems from two major hardware manufacturers, with clearance extended across them and processing latency published (under 5 minutes with one vendor at a stated bandwidth, under 15 with the other); advanced visualization is integrated through a third party partner; and the site carries a meaningful use certification adherence page, pointing to federal health IT certification of part of the software. No integration with a named EHR is published, so how results reach the ordering clinician outside radiology is not shown.

The architecture is stated and the location is not. The platform is described as cloud native with a hybrid option, and the remote scanning platform, TechLive, is offered through a public cloud marketplace. No region list, residency commitment or statement of where images are processed and where they are stored is published, which matters for a portfolio assembled from Dutch, French and UK acquisitions under a US parent.

Adjacent figures are published, and none is the price a provider pays. At RadNet's own centers patients have been charged about $40 out of pocket for the Enhanced Breast Cancer Detection program, with roughly 36 percent electing it, a retail charge that includes the safeguard radiologist read; the breast ultrasound module is positioned against an existing Category III code for quantitative ultrasound tissue characterization, which is the payer's side; and RadNet, a public company, reports more than $16 million of new third party contract value in the first quarter of 2026. No price, unit of sale or rate is published for DeepHealth OS, the enterprise imaging suites or any clinical suite sold to an outside provider, and inquiries go through a demo request.

Coverage is broad and validated in part. Clinical AI spans breast, chest and lung, prostate, brain, thyroid and, through Gleamer, musculoskeletal imaging, across radiography, mammography, tomosynthesis, ultrasound, CT, MRI and PET, and the enterprise imaging layer sits underneath. The breast workflow is validated at scale in community screening (the ASSURE study, more than 100 sites in four states); validation of the other suites in their settings is not described on the company's pages.

The company reports more than 2,700 customer contracts in more than 50 countries including Gleamer, more than 300 screening sites and more than 10 million mammograms a year, and named third party customers include ONRAD, serving more than 120 hospitals and radiology groups, and Wichita Radiological Group. The company states that not all products are available in all countries.