Hologic Genius AI Detection

What is sold has grown larger than the detector. 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. What dilutes the AI's place is the makeup of the offer, not any weakness in the model, and not the size of the company behind it.

The distinction matters at procurement because suite pricing conceals how much of the spend is buying inference.

Findings are shown to the interpreting physician during the read, and the physician confirms or dismisses each one. The 510(k) summaries (K201019, November 2020; K221449, October 2022) indicate Genius AI Detection as a concurrent reading aid for tomosynthesis exams on Hologic Dimensions and 3Dimensions systems, marking soft tissue densities and calcifications with confidence scores and a case score; nothing is triaged away and no exam is routed to a reduced read.

The company publishes a table mapping case score bands to the approximate fraction of cases in each band that prove to be cancer, from roughly one in 2,800 at a score of zero to about one in 27 in the highest band, and the PRO platform presents a color coded score from 1 to 10, so a reader gets a number with its own prior. The physician user guide states that performance has not been evaluated on patients with breast implants or on partial views. What is not published is the rule behind the PRO platform's worklist ordering and automated pre reporting, and what happens after a missed or false mark.

The approach and version history are published; the training data is not. The company states the algorithm class as deep learning and publishes a technical white paper contrasting it with the earlier feature engineered generation; the Genius AI Detection 2.0 summary (K221449) describes a more sophisticated convolutional neural network for calcification cluster classification; versions are tied to clearances (K201019 in 2020, K221449 and K230096 for 2.0) and to the PRO 1.2 release in the product's June 2025 security disclosure; and the company states case level sensitivity of 94 percent and publishes a case score calibration table.

What is missing is the training data: no corpus size, institutional or geographic provenance, development population or count of cancers trained against. The detection model reads only Hologic's own tomosynthesis reconstructions.

The chain is structurally short at the model layer. The detector is Hologic's own deep learning model, cleared by the FDA under Hologic's name (K201019, K221449, K230096), so it cannot be replaced by another supplier's model without going back through clearance.

Hologic's security disclosure for Genius AI Detection PRO 1.2 (MDS2, June 2025) says a software bill of materials identifying third party components, developers and versions is available on request and that the manufacturer evaluates third party components, and that the system connects to the internet for cloud based processing of prior exams and for remote service; the cloud provider and region are not named.

Published research, partly independent, stands behind the detector, all of it retrospective or reader based. The pivotal reader study in the first 510(k) summary (K201019) used 390 cases (106 cancers, 284 negatives) and 17 readers: sensitivity rose from 66.8 to 75.9 percent with the software, AUC from 0.794 to 0.825, recall of negative cases from 23.4 to 25.8 percent, and reading time from 46.3 to 52.0 seconds.

A retrospective study led from Massachusetts General Hospital in the American Journal of Roentgenology examined 7,500 tomosynthesis screening exams from 2016 to 2019 and found the software flagged almost 90 percent of the 500 cancers radiologists had identified and about a third of those initially read as negative.

Two findings cut the other way and are published: a Nottingham reader study against 108 specialists found comparable performance with lower specificity, and a Mass General Brigham analysis of more than 160,000 screenings before and after adoption of SmartSlices found no significant difference in tumors detected. The company footnotes that its pivotal analyses do not control type I error. No prospective screening study reports cancer detection or interval cancers in a live program.

Hologic's security disclosure for Genius AI Detection PRO 1.2 (MDS2, June 2025) describes how patient data moves through the product, and not what happens to it beyond care. The system stores, transmits and displays protected health information, keeps it persistently in internal databases, sends results to the PACS by DICOM, connects to the internet for cloud based processing of prior exams and for remote service, and allows manufacturer initiated remote sessions in which patient data remains accessible; the server component does not encrypt stored data, and TLS is available on channels other than the PACS connection.

It offers de identification to DICOM profiles as a capability. No retention period, and no statement on whether images, regions or scores contribute to model development, is published. The physician user guide states performance was not evaluated for breast implants or partial views, a safety relevant limit. Safety engineering, meaning guardrails and how a safety event is handled, is not published.

No business associate status is stated, and the product's own security disclosure reaches the question. Hologic's MDS2 for Genius AI Detection PRO 1.2 (June 2025) states that the system stores, transmits and displays protected health information, that it connects to the internet for cloud based processing of prior exams, and that the manufacturer initiates remote service sessions in which patient data remains accessible.

That makes Hologic a recipient of protected health information in cloud processing and servicing, and nothing published states whether a business associate agreement covers those paths. The corporate privacy policy linked from the footer is the other published document.

A recognized certification is named in the manufacturer's own material. Hologic's MDS2 for Genius AI Detection PRO 1.2 (June 2025) states that the manufacturer holds ISO 27001 certification and that a software bill of materials is available on request, and describes the product's controls: role based access, external authentication and multi factor support, audit logging, application whitelisting on the Linux server, and TLS on most channels, with no encryption of stored data on the server component.

Hologic publishes MDS2 forms on its breast and skeletal products cybersecurity page. The ISO 27001 certificate's scope and date are not stated, no SOC 2 report or penetration test disclosure is published, and the product page footer carries no security link.

The US position is identifiable by clearance number. The detector is cleared under product code QDQ, 21 CFR 892.2090: Genius AI Detection (K201019, 18 November 2020), a concurrent detection and diagnosis aid for tomosynthesis on compatible Hologic systems, and Genius AI Detection 2.0 (K221449, 6 October 2022; K230096, 23 May 2023; K243341), with a more sophisticated calcification model and lesion correlation across views.

The product page also cites K163623 and K240301 among its clearances, and the tomosynthesis platform itself is approved under premarket approval P080003 (11 February 2011, the first tomosynthesis system approved in the United States), extended by supplements including S008. The indication covers Hologic Selenia Dimensions and 3Dimensions systems at standard and high resolution and 6 mm SmartSlices. Marking and approvals outside the United States are not stated on the product page.

Subgroup evaluations are disclosed with their groups and published as conclusions. The first 510(k) summary (K201019) reports a standalone test on 764 cases (106 cancers, 658 non cancers) stratified by lesion type and breast density, with no significant differences, and a Northwestern analysis reported similar performance across common US racial and ethnic groups while stating that it came from a single academic center with a predominantly white population and limited subgroup numbers.

Per group figures are not published, so where performance is weakest stays out of view, and density is the variable on which a mammography detector most plausibly fails unevenly. No model card, training data composition, governance body or post deployment subgroup monitoring is published.

Scope of use is disclosed better than recourse. The physician user guide names populations the software was not evaluated on (breast implants, partial views), and the company footnotes that its pivotal analyses do not control type I error and cannot be generalized beyond that study.

The 510(k) summaries give no error rate at the operating point: the first (K201019) reports a reader study and a standalone test described as comparable across acquisition modes, and the 2.0 summary (K221449) reports a 764 exam standalone test (106 cancers) in which specificity rose 12 percent against the first version while sensitivity was maintained, without the rates themselves.

No warranty, indemnity, service level or remediation commitment is published; the concurrent read design means a radiologist sees every case and signs the report, which is where liability lands by default. The company became private in April 2026, which narrows what can be verified about the counterparty.

Output travels through open standards and input is closed. Results reach the PACS by DICOM (Hologic's MDS2), packaged as a DICOM Mammography CAD SR object for display on DICOM review workstations, including Hologic's SecurView and third party archives, and the PRO platform's pre reporting feeds the site's reporting software. Input is limited to Hologic's own Dimensions and 3Dimensions tomosynthesis reconstructions (K221449).

Sites using the software appear in published studies (Massachusetts General Hospital, Northwestern, Mass General Brigham). No EHR integration is published, and reprocessing a case later requires the site to keep reconstructed slices or projections.

The product runs partly at the site and partly in the cloud, and the cloud location is not stated. Detection runs with Hologic's acquisition systems at the site, but the MDS2 for Genius AI Detection PRO 1.2 (June 2025) says the system requires internet connectivity for cloud based processing of prior exams and for remote maintenance, and keeps patient data persistently in internal databases. No cloud provider, region, residency commitment or tenancy description is published.

No price, unit of sale or licensing model is published, and the posture is discoverable: the software is sold as an option on Hologic's own mammography systems, and the product pages route to a demonstration request. Whether it is licensed perpetually or by subscription, per system, per reading seat or per exam is not stated. Because it runs only on Hologic acquisition hardware, the software price is negotiated by a buyer already committed to the same vendor's gantry.

The company was taken private by funds managed by Blackstone and TPG on 7 April 2026 and has ended public financial reporting, so segment figures a buyer once read as a check on the line's scale are no longer published.

Coverage is deliberately narrow and validated in part. Detection runs on digital breast tomosynthesis from Hologic Dimensions and 3Dimensions systems for symptomatic and asymptomatic women (K221449), with the PRO platform adding 2D analysis, prior exams and pre reporting, Quantra for density and 3DQuorum for 6 mm SmartSlices.

Screening detection is validated in reader and retrospective studies (K201019; Massachusetts General Hospital; Nottingham), and SmartSlices in a Mass General Brigham before and after analysis; the PRO platform's 2D and prior exam functions have no published validation, and breast ultrasound, MRI and contrast enhanced mammography are outside the product. The US clearance record is documented, and availability and marking outside the United States are not stated on the product page.