Hologic Genius AI Detection

What is sold has grown larger than the detector, and that is what sets this grade. The current generation platform wraps the detection model in a reading environment: automated pre reporting, image quality checks, aggregation and display of prior examinations, and the reading interface itself, alongside density assessment and slice generation modules that are separate capabilities rather than the same model applied twice. A site licensing the current platform is buying a reading workflow suite in which detection is one component, and a meaningful share of what appears on the quote is interface and reporting software that would still function if the detector were switched off.

The detector itself is not in doubt. It is a deep learning model that searches every reconstructed tomosynthesis slice, marks lesions and returns a case level score, and no part of the detection product exists without it. Judged on the detector alone this would sit at the top band, and the reason it does not is the composition of the offer rather than any weakness in the model or any judgement about the size of the company behind it.

The distinction matters at procurement because suite pricing conceals how much of the spend is buying inference. Ask which line items on the quote are the detection model rather than the reading and reporting interface, whether the detector can be licensed on its own without the workflow platform, and what each costs separately.

This is the most conservative autonomy model in the lane and it is documented precisely. The regulatory description places the findings in front of the interpreting physician concurrently with their reading of the examination, and the physician confirms or dismisses each one. Nothing is triaged away, no examination is routed to a reduced read, and no case is dismissed without a radiologist seeing it. The automation acts on attention and ordering rather than on whether a human looks.

What lifts this above a bare description is the published case score calibration. The vendor publishes a table mapping score bands to the approximate fraction of cases in each band that prove to be cancer, running from roughly one in 2,800 at a score of zero to about one in 27 in the highest band, with the current platform simplifying presentation to a one to ten scale. A reader given a number that comes with its own prior is in a materially better position than one handed an unanchored confidence value, and very few records here publish anything equivalent.

It sits at B rather than higher because the highest band in this lane is earned by systems that both exercise consequential autonomy and publish the exact rule governing it. Here the autonomy exercised is minimal, which is a defensible design choice rather than a defect, and the operational rule around worklist prioritisation is not published. Ask whether case score thresholds and worklist ordering are configurable by the site or fixed by the vendor, and what proportion of your case mix falls into each score band.

Operating characteristics are published to a standard well above the lane median. The vendor states the algorithm class as deep learning, publishes a technical white paper describing the network approach and contrasting it with the earlier feature engineered generation, names versions explicitly so a result can be tied to a build, states case level sensitivity of 94 percent, and publishes the case score to cancer fraction calibration table that lets a reader convert a score into a prior. It also names what the model was not validated on, specifically implant cases and partial views.

The gap is upstream and it is the same gap that separates this from the top of the lane. Nothing describes the training data: no corpus size, no institutional or geographic provenance, no development population, no statement of how many cancers the model was trained against or how they were confirmed. Architecture detail beyond the general class is not published, and no operating threshold is specified for the routing behaviour of the current platform.

That asymmetry is worth flagging plainly to a buyer, because the published half is the half that helps at the reading workstation while the absent half is the half that predicts whether performance will transfer to a population unlike the development set. Ask for the size, source and demographic composition of the training and validation data, whether any of it came from the sites or gantry generations you operate, and for the version identifier of the build you will be running.

Nothing identifies any component in the chain. No statement of what is built rather than licensed, no third party libraries or frameworks named, no model provenance for the detector or for the density and slice generation modules, and no software bill of materials located in two passes. The predecessor detection line originated with an acquired company, and whether any of that lineage survives in the current models is not stated.

The right question for this product class is different from the sub processor list that suits a hosted platform, and a buyer should ask for the artefact that actually exists. Device software installed on a hospital workstation carries a supply chain in the form of its bundled components, and a software bill of materials is the document that enumerates them. It is also the document that determines how quickly a site can answer the only question that matters when a widespread vulnerability is disclosed, which is whether the affected component is running on the acquisition workstation attached to the mammography suite.

The whole lane grades poorly on this axis, so the bottom band is not a singling out of this vendor. It does carry more operational weight here than for a cloud product, because the software sits on a workstation inside the clinical network rather than in a vendor's environment, which makes the customer rather than the vendor the party exposed. Ask for the software bill of materials for the installed version, the vendor's notification commitment when a bundled component is found vulnerable, and the supported patch path for workstations that cannot be taken out of service.

A real and partly independent body of work exists, and it stops short of the design this index treats as the bar. The pivotal multi reader multi case study reported average observed reader sensitivity for cancer cases of 75.9 percent with the software against 66.8 percent without, a difference of about 9 percent, with significantly increased area under the receiver operating characteristic curve. A retrospective study led from Massachusetts General Hospital and published in the American Journal of Roentgenology examined 7,500 tomosynthesis screening examinations from 2016 to 2019 and found the software flagged almost 90 percent of the 500 cancers radiologists had already identified, correctly localised them, and flagged roughly a third of the cancers initially interpreted as negative. A Northwestern analysis reported similar performance across common United States racial and ethnic groups.

What is missing is prospective screening outcomes. Every retrieved study is retrospective or a reader study. Nothing establishes what happens to cancer detection or interval cancers in a live programme, which is the evidence the three records graded above this one in the lane carry. Two further items belong in front of a buyer, and the vendor deserves credit for publishing both rather than burying them: a Nottingham reader study against 108 specialists found the algorithm performed comparably to radiologists but with lower specificity and therefore more false positives, and a Mass General Brigham analysis of more than 160,000 screenings before and after adoption of the SmartSlice product found no significant difference in tumours detected. The pivotal study also carries the vendor's own footnote that its analyses do not control type I error and cannot be generalised beyond that study.

The predecessor question is treated separately on this record and matters to anyone who has read the older literature. Ask whether any prospective screening study of this software is registered or under way, and what the detection and recall rates were at comparable sites before and after adoption.

One disclosure here is better than the lane norm and deserves to be named first. The physician user guide states that diagnostic performance has not been evaluated for mammograms from patients with breast implants or for partial views. Publishing the populations a detector has not been validated on is a safety practice that most vendors in this index avoid, and it is directly actionable: a reader who knows the model was not tested on implant cases knows not to lean on a negative result in one.

Local processing is the second favourable property, since images that never leave the acquisition workstation cannot be retained, breached or repurposed by the vendor. Neither point is documented far enough to carry a higher grade. No retention position, no statement on whether any images or derived data contribute to model development, no encryption or workstation hardening guidance, and no description of what service access touches was retrieved in two passes.

The model improvement question is the one to press, because this manufacturer has access to the largest tomosynthesis installed base in the country and every one of those workstations runs its software. Ask in writing whether any image, region of interest or score generated under your licence can be collected for model development or quality monitoring, what is captured in diagnostic logs, and require the answer as a contract term rather than an assurance from a sales team.

The governing instrument here is different from most records in this index, and naming it is more accurate than penalising the vendor against a test that does not fit. This is regulated device software that resides on the acquisition workstation inside the customer's own facility and processes images locally. In that architecture the covered entity retains the protected health information and the manufacturer is not necessarily receiving it, so a business associate agreement may not be the operative document at all. The index convention is to name the framework that applies rather than mark down an inapplicable one, and for on premises device software the applicable instruments are the customer's own security rule obligations over its workstations and the manufacturer's device and servicing obligations.

That reasoning holds only as far as the local processing does. What was not established in two passes is whether anything leaves the site: service and remote support access to the workstation, diagnostic logs, telemetry, error reports containing image data, or any component of the next generation platform that may introduce a hosted element. Each of those would create exactly the exposure the local architecture otherwise avoids, and none is described.

So the middle band records a posture that is probably favourable and is not documented. Ask whether the manufacturer or its service organisation can access images or logs containing protected health information during support, whether a business associate agreement is offered for those paths, and whether any module in the current platform transmits data off site.

Two passes were run against this axis, one by company name and one by product name paired with certification terms, and neither returned an attestation, a trust centre, a security page or a sub processor list.

The framework that governs is not the one that governs most records here, and saying so is more useful than recording a bare absence. This is regulated medical device software, not a hosted service, so the instruments a buyer should be asking for are the ones device manufacturers actually produce: the cybersecurity documentation submitted with the premarket clearance, a software bill of materials, the manufacturer's postmarket vulnerability handling and patch commitments, and the manufacturer disclosure statement for medical device security that hospital security teams use for exactly this class of product. A service organisation control attestation is designed for a different architecture and its absence is less meaningful here than it would be for a cloud vendor.

What cannot be resolved from public sources is whether any of those device specific artefacts exist and what they say, and a company of this scale selling into hospital networks has certainly been asked for them repeatedly under agreement. The grade records what a buyer can confirm before contacting sales, which is nothing. Ask for the manufacturer disclosure statement for medical device security, the software bill of materials for the installed version, and the vendor's committed timeline for patching vulnerabilities on the acquisition workstation.

The strongest regulatory record encountered in this lane, and the only one where the clearance register itself was checked rather than a vendor device count taken on trust. The detector is cleared under product code QDQ, 21 CFR 892.2090, as Genius AI Detection under K201019 dated 18 November 2020, and as Genius AI Detection 2.0 under K221449 dated 6 October 2022, with subsequent clearances K230096 dated 23 May 2023 and K243341. The underlying acquisition platform is approved under premarket approval P080003, granted 11 February 2011 as the first tomosynthesis system approved in the United States, extended by supplements covering synthesised two dimensional imaging in 2013, the 3Dimensions system in 2018 and high resolution tomosynthesis in 2019.

Two records in this lane are held one band lower for exactly the gap that is closed here: numbered, dated clearances rather than a count asserted on a product page. The distinction is the whole point of the axis, because a published clearance set is exhaustive and cannot be shaped by marketing, whereas a device count can.

What was not retrieved is the position outside the United States. European conformity marking and other national approvals were not established, and the presence of European operations and patent litigation is not evidence of a specific marking for a specific module. Ask for the conformity marking certificate and the notified body for each module you intend to run in your jurisdiction, and confirm which clearance covers the specific gantry model in your department.

The subgroup question is answered in public, which is more than most of this lane manages. Research presented at a major radiological meeting examined performance of the tomosynthesis detection algorithm across common United States racial and ethnic groups and reported similar measured performance across every cohort evaluated. The peer reviewed retrospective work is also unusually candid about its own limits, stating plainly that it was conducted at a single academic centre with a predominantly white patient population, that subgroup sample sizes limit statistical power, and that results may not generalise to other practice settings or to other algorithms. Limitations disclosed by the vendor in its own announcement rather than extracted by a reader are worth something.

The apparatus around those results is absent. No model card, no description of the demographic or geographic composition of the training data, no governance structure, and no commitment to monitoring subgroup performance after deployment was retrieved. Performance by breast density, which is the central confounder in mammography and the axis on which a detector most plausibly fails unevenly, was not located as a published breakdown for the current product.

A point in time subgroup analysis is not a control, and detector behaviour shifts as populations, gantry generations and reconstruction software change. Ask for performance stratified by breast density category, what subgroup monitoring continues after installation, and what the vendor commits to do if a disparity appears in your population.

Scope of use is disclosed better than recourse is. The vendor states in its own materials that the pivotal analyses do not control type I error and cannot be generalised beyond that study, and the user documentation names populations the software was not evaluated on. Publishing the boundary of a claim is a form of honesty about liability that this index rarely encounters, and it gives a reader a defensible basis for knowing where the product's validated performance stops.

Beyond that boundary nothing was retrieved. No warranty on detection performance, no indemnity, no service level commitment, no remediation position, and no statement of where responsibility sits when a case the model scored low proves to be a cancer. The concurrent read design means a radiologist has seen every case and has signed the report, which is where liability will practically land, and that allocation is the default rather than anything the vendor has committed to in public.

One structural point belongs to a buyer's diligence rather than to a grade. Recourse depends on the counterparty as well as the contract, and the entity behind this product changed hands and ceased public reporting in April 2026, which narrows what can be independently verified about it. Ask what the vendor warrants regarding detection performance in your deployment, who carries liability for a missed cancer marked as low suspicion, and what support and patch commitments survive a change of control.

Interoperability here runs one way and the asymmetry is the finding. On the output side the design is properly open: results are packaged as a DICOM Mammography CAD SR object and distributed for display on DICOM compliant review workstations, so marks and case scores reach the manufacturer's own review software and third party archives and workstations alike. A site does not have to buy the vendor's reading platform to see the output.

On the input side it is closed. The software runs on the acquisition workstation of the manufacturer's mammography system and analyses that system's tomosynthesis reconstructions, so the images it can read are determined by who made the gantry. The user documentation adds an operational consequence that is easy to miss at procurement: reprocessing a case later is not possible unless the site retains reconstructed slices or raw projections, which is a storage decision with real cost implications that has to be made before it is needed.

No integration with the electronic health record, the reporting system or the ordering pathway outside radiology was retrieved, which is normal for this product class but leaves the axis shallow. The wider ecosystem position is unusual and belongs in a buyer's thinking: the acquisition hardware this software depends on is controlled by a company that also competes with every other algorithm vendor in this lane. Ask what storage of reconstructed slices or projections is required to preserve reprocessing, how results reach your reporting system, and what contractual assurance exists that competing detection software will continue to run against this vendor's image output.

The deployment model is stated rather than inferred, which is why this sits above the lane norm. The software resides on the acquisition workstation of the mammography system, inside the customer's facility, and processes images there. Residency follows automatically: the images stay where the site keeps them, under the site's own jurisdiction and controls, and no cross border transfer question arises for the detection step. For a buyer in a jurisdiction with data localisation requirements that is the strongest posture available in this lane, and it is available without negotiating anything.

The limits of that statement are what hold it below the top band. The current generation platform aggregates prior examinations, performs image quality checks and produces automated pre reporting, and whether any of that introduces a hosted or cloud component was not established. Nor was any position on telemetry, diagnostic log egress, remote servicing connections or update distribution. A local architecture with an undocumented service channel is only as local as that channel allows.

Sizing and infrastructure also went undescribed, which matters because reprocessing depends on retaining reconstruction data. Ask whether every module you are licensing runs entirely on premises, what network connections the workstation requires for updates and support, and what storage the site must provide to retain the data the software needs.

Two passes over the vendor site, the product pages and the trade press established no price, no unit of sale and no licensing model. Whether the software is licensed perpetually, per system, per reading seat or per examination is not stated anywhere retrieved, and the product pages route to a demand a demo form.

The grade is at the bottom band rather than the middle because disclosure moved backwards during the period under review. Until April 2026 this was a listed company filing periodic reports, and a buyer could read segment level Breast Health revenue, research spending and guidance as a proxy for the vendor's stability and commitment to the line. On 7 April 2026 the take private completed and the company filed to suspend trading, delist and terminate its registration. That reporting has ended. A comparable record in this lane holds the middle band partly because exchange requirements give a buyer visibility no private competitor offers, and that is precisely the visibility this vendor no longer provides.

The practical effect is narrow and worth stating without overreach: nothing about the product changed, but the amount a counterparty can independently verify about the seller did. Ask for the licensing model in writing, whether the licence is tied to the mammography system or to the site, what happens to pricing at renewal under the new ownership, and what the vendor will commit to on support and product roadmap for the line now that segment performance is no longer publicly reported.

The scope is one organ and one modality family. Detection runs on digital breast tomosynthesis, with density assessment through Quantra, slice reduction through 3DQuorum, a legacy two dimensional detector, and in the current generation combined two dimensional and tomosynthesis analysis incorporating prior examinations. Within mammography that is a reasonably complete set: detection, density and reading efficiency are all covered.

Outside mammography there is nothing. Breast ultrasound, breast magnetic resonance and contrast enhanced mammography are absent, which matters because those are the modalities a screening finding actually proceeds into. A breast centre buying this covers the screening read and then leaves the software behind at the point the diagnostic workup begins, so the product sits at one station of the pathway rather than across it. Several records in this lane now extend into ultrasound or risk assessment, and this one does not.

Geographic reach is the second limit and it is a retrieval gap rather than a known absence. The United States position is documented in detail through the clearance record. Availability, marking and commercial release outside the United States were not established in two passes, and the records graded a band higher here carry documented deployment across dozens of countries and national screening programmes. Ask which modules are cleared and commercially released in your country, and what the vendor offers for the ultrasound and magnetic resonance steps that follow a suspicious screening read.