Mammography AI product research is most useful when it begins with a specific workflow question: will the software act as a concurrent reader, a triage filter, or a decision-support layer inside an existing double-reading program? The answer changes which evidence matters. A device cleared as radiological computer-assisted diagnostic software must show that it improves reader performance in its intended population, but that regulatory requirement does not by itself establish how the product will perform in a particular practice [1]. Vendor documentation describes supported views, exclusions and processing behavior, while published trials measure program-level outcomes such as cancer detection rate and recall rate [2][4]. A sound evaluation therefore separates device capability, workflow fit and local validation instead of treating them as one question.

Start from the Intended Reading Protocol

The first decision point is whether the AI is meant to assist during interpretation or to remove some exams from human reading. Hologic describes Genius AI Detection software as a concurrent aid in which the interpreting physician confirms or dismisses findings while reading a digital breast tomosynthesis exam [2]. That is a different operational model from the triage approach evaluated in a 2026 Nature Medicine trial, where exams below a predetermined AI risk threshold could be automatically considered normal [4]. These are not interchangeable claims about the same product category.

The reading protocol also determines which performance metrics deserve priority. A concurrent detection aid should be assessed against reader-level measures such as sensitivity, specificity and area under the curve, as specified in the FDA special controls for radiological computer-assisted diagnostic software [1]. A triage configuration is better judged by workload reduction, cancer detection rate and recall rate at the program level [4]. Asking for a single accuracy number before defining the workflow will produce a comparison that cannot support a procurement decision.

Regulatory Controls Define the Evidence Floor

Under 21 CFR 892.2060, radiological computer-assisted diagnostic software for lesions suspicious of cancer is a Class II device with special controls. Design verification must include a detailed description of algorithm inputs, outputs, major components and limitations, and performance testing must demonstrate improved reader performance in the intended use population [1]. The test dataset must contain enough cases from important cohorts, including subsets defined by confounders, effect modifiers, concomitant diseases and image acquisition characteristics, so that performance estimates and confidence intervals can be characterized for those subsets [1].

Labeling requirements are equally specific. The device must describe the indicated patient population, intended reading protocol, intended user and training, compatible imaging hardware and protocols, and situations where the device may fail or operate below expected performance [1]. For product research, these labeling sections are more informative than a summary accuracy figure. A buyer should compare the documented exclusions and supported views across candidates before evaluating any published clinical results, because a favorable study conducted outside those exclusions does not transfer to the labeled use.

Suggested Evaluation Checklist

For each candidate, record the labeled intended reading protocol, the documented exclusions and unsupported views, the compatible imaging hardware, and the DICOM result delivery path [1][2][3]. Then list the published evidence that used the same workflow configuration, noting whether the study was prospective, paired, retrospective or observational [4][5]. Finally, define the local validation dataset and the primary metric before any vendor demonstration, so that the demonstration answers the practice's workflow question rather than the vendor's preferred use case.

Vendor Documentation Shows Where the Product Stops

Hologic's Genius AI Detection user guide states that the software has not been evaluated and may not operate at expected performance for patients with symptomatic or palpable lesions, breast implants, pacemakers in the field of view, skin markers, motion, cut-off anatomy, prior surgery or visible biopsy clips, or studies with missing standard views [2]. The guide also notes that diagnostic performance has not been characterized for mammograms from patients with breast implants or partial views [2]. These are not minor footnotes; they define the population in which the vendor's own validation applies.

The DICOM conformance statement for Genius AI Detection PRO version 3.1.0 adds an interoperability layer to the evaluation. The system can receive SOP instances, query and retrieve prior mammograms from multiple external application entities, process images using AI algorithms, and send results as Mammography CAD SR or secondary capture images [3]. A practice that needs results delivered into a specific PACS or reporting workflow should verify those storage SOP classes and transfer syntaxes against its own environment rather than assuming that a cleared device will integrate without additional configuration.

Published Trials Measure Program Outcomes, Not Product Superiority

The paired noninferiority trial published in Nature Medicine in March 2026 evaluated an AI strategy in a population-based screening program. The standard strategy produced 62,602 radiologist screening readings and 198 screen-detected cancers, with a cancer detection rate of 6.3 per 1,000 and a recall rate of 4.8 percent [4]. The AI strategy used double human reading with AI support only for cases above a predetermined risk threshold, while lower-risk exams could be automatically classified as normal [4]. The coprimary outcomes were workload, cancer detection rate and recall rate, with no formal multiplicity correction applied between them [4].

A separate real-world US deployment study, published in Nature Health, examined a multistage AI-driven workflow using DeepHealth Breast AI version 2.x across five practices and 109 sites. The authors reported that cancer detection rate increased by 0.99 per 1,000 screens, from 4.59 to 5.58, for the population studied [5]. That study is a vendor-affiliated observational deployment analysis, not a randomized comparison, and its authors include employees of DeepHealth [5]. Product researchers should treat such results as evidence about one configured workflow in specific practices, not as a general ranking of one vendor against another.

A Concrete Evaluation Plan for One Practice

A defensible evaluation can be built around three documents and one local test. First, request the 510(k) summary or equivalent clearance documentation and compare the labeled patient population, reading protocol and exclusions with the practice's screening mix [1]. Second, obtain the user guide and DICOM conformance statement to confirm supported views, processing behavior and result delivery [2][3]. Third, identify published studies that used the same workflow configuration the practice intends to adopt, and record the study design, denominator and follow-up period before accepting any outcome [4][5].

The local test should be a retrospective reader study on the practice's own images, using the same protocol that would be used clinically. For a concurrent detection aid, the relevant comparison is reader performance with and without the device on a dataset that includes the practice's expected confounders, such as breast density and acquisition equipment [1]. For a triage configuration, the practice should measure how many exams would be automatically classified as normal and what proportion of subsequently detected cancers would have been missed under that threshold [4]. This kind of local validation is not a substitute for regulatory clearance, but it is the only way to estimate the operational effect in a specific setting.