AI in Mammography: How It Works, FDA-Cleared Tools, and What Imaging Centers Need to Deploy It
Artificial intelligence in mammography applies machine learning algorithms, primarily deep neural networks trained on large, annotated mammographic image datasets, to help detect breast cancer, assess breast cancer risk, quantify breast density, triage screening worklists, and structure diagnostic reports. AI in mammography does not replace the radiologist. It operates as a computational layer within the breast imaging workflow: the AI processes the digital mammogram or digital breast tomosynthesis (DBT) study, identifies patterns associated with malignancy or elevated risk, and presents its output to the radiologist for verification and clinical interpretation.
The scope of AI in mammography spans five distinct application categories, each addressing a different part of the breast imaging workflow: computer-aided detection of malignant findings, cancer risk prediction models that estimate future cancer probability from current imaging, breast density assessment and BI-RADS density category assignment, screening workflow triage and worklist prioritization, and reporting assistance with BI-RADS template structuring. These categories differ in what they do, where they sit in the clinical workflow, how the FDA regulates them, and how mature they are in clinical deployment. Understanding these distinctions is essential to evaluating what AI can realistically deliver in a specific breast imaging practice, and what it cannot.
This guide covers the complete picture: how AI is used across the five mammography application categories, what measurable performance has been shown in published studies, where the clinical evidence stands in 2026, where the limitations and regulatory boundaries sit today, how the different application types compare for breast imaging centers evaluating where to invest, and what PACS infrastructure is required to integrate AI mammography tools into existing screening workflows.
How AI Is Used in Mammography: Five Application Categories
Computer-Aided Detection of Malignant Findings (CAD)
Computer-aided detection in mammography refers to AI systems that identify and mark potential cancers on digital mammograms and digital breast tomosynthesis studies, flagging suspicious masses, microcalcification clusters, architectural distortions, and asymmetries for the radiologist’s review. Current-generation AI CAD is fundamentally different from the older rule-based CAD systems deployed in the 2000s. Legacy CAD relied on hand-engineered features and produced high false-positive rates that reduced radiologist confidence; modern AI CAD uses deep convolutional neural networks trained on millions of annotated mammograms and achieves detection performance closer to that of expert breast radiologists.
Named FDA-cleared products in the AI mammography CAD category include iCAD ProFound AI for both 2D digital mammography and 3D digital breast tomosynthesis, Lunit INSIGHT MMG, ScreenPoint Medical’s Transpara, Therapixel MammoScreen, and Hologic Genius AI Detection. Each product operates at the study level, processes the four standard mammographic views (right and left craniocaudal and mediolateral oblique), and produces a per-image overlay highlighting suspicious findings, along with a case-level malignancy score for the overall study.
The FDA maintains a public database of AI/ML-enabled medical devices. As of 2024, radiology has the largest share of FDA-cleared AI devices, and mammography is one of the most active subspecialty areas within that category, with new clearances throughout 2024 and 2025.
Cancer Risk Prediction Models
Cancer risk prediction AI takes a different clinical task from CAD. Where CAD asks “is there cancer in this image now?”, risk prediction asks “what is this woman’s probability of developing breast cancer in the next 1 to 5 years, based on imaging features detectable today?” The clinical value of risk prediction lies in personalizing screening intervals: women at elevated short-term risk benefit from supplemental imaging (breast MRI, contrast-enhanced mammography) or shorter screening intervals, while women at low predicted risk may safely extend their screening interval.
The most recent significant entry in this category is Clairity Breast, which received FDA marketing authorization in 2025 for predicting 5-year breast cancer risk from a standard screening mammogram. The Breast Cancer Research Foundation summary of the Clairity Breast clearance describes the approach: rather than detecting visible cancer, the model identifies subtle imaging features that correlate with elevated future cancer risk, producing a 5-year risk score that supplements traditional risk models like the Tyrer-Cuzick and Gail models.
Other risk prediction approaches in active development include the Mirai model from MIT’s Computer Science and AI Laboratory, which predicts 1-year and 5-year breast cancer risk and has been validated across multiple international datasets. The clinical integration challenge for all risk prediction tools is the same: the risk score is only useful if it changes screening management, which requires both clinician adoption and a downstream pathway for women identified as high-risk to receive supplemental imaging.
Breast Density Assessment
Breast density is the proportion of fibroglandular tissue to fatty tissue in the breast, assessed visually by the radiologist and reported using the four BI-RADS density categories (A almost entirely fatty, B scattered fibroglandular density, C heterogeneously dense, D extremely dense). Density matters clinically for two reasons: dense breast tissue makes mammographic detection of cancer more difficult because fibroglandular tissue and cancer both appear white on mammograms, and dense breast tissue is itself an independent risk factor for breast cancer. As of 2024, the FDA’s updated mammography reporting requirements mandate that breast density be communicated to patients in lay terms, making consistent and accurate density assessment a regulatory as well as clinical priority.
AI density assessment tools automatically assign BI-RADS density categories from the screening mammogram, addressing well-documented interreader variability in visual density assessment. The most established product in this category is Volpara Health’s VolparaDensity, FDA-cleared and widely deployed across US screening practices for objective density measurement. iCAD’s ProFound AI suite includes a density assessment module, and several other AI vendors have integrated automated density assessment as a component of broader mammography AI platforms.
The clinical workflow advantage of AI density assessment is consistency: the same algorithm applied to the same mammogram always produces the same density category, eliminating the visit-to-visit variability that affects manual assessment when different radiologists read the same patient’s serial screenings.
Workflow Triage and Worklist Prioritization
Mammography screening AI can serve as a triage layer that processes incoming screening studies and reorders the radiologist’s reading worklist based on the AI’s suspicion assessment. Studies with high case-level malignancy scores are moved to the top of the worklist for earlier radiologist review. Studies with very low malignancy scores may be flagged for first-read confirmation only, depending on the institution’s protocol and the regulatory pathway under which the AI is deployed.
The most operationally significant triage application is in high-volume screening centers that read thousands of mammograms per week, where AI-driven workflow prioritization can reduce time-to-callback for women with suspected cancer and reduce reading time on clearly normal studies. Clinical evidence for triage workflow deployment is most developed in European screening programs operating under regulatory frameworks different from the US, where AI is being studied as a replacement for one of the two readers in double-read screening programs.
The strongest prospective evidence for triage so far comes from the AITIC trial in Cordoba, Spain, published in Nature Medicine in 2026. Each of 31,301 screening exams was read two ways: standard double reading without AI, and an AI strategy in which exams the AI scored as low risk (about 70% of exams) were classified as normal with no radiologist reading, while the remaining exams were double read with AI support. Radiologists read only about 36% of exams, cutting reading workload by 63.6%, and the cancer detection rate rose from 6.3 to 7.3 per 1,000. The trade-off was a higher recall rate, from 4.8% to 5.5%, which did not meet the trial’s noninferiority margin, although the positive predictive value of recalls stayed essentially unchanged at about 13%.
In US screening, AI triage typically serves as a supplementary layer alongside standard single-reader workflows, with the AI’s worklist prioritization and case-level scores informing, but not replacing, the radiologist’s interpretation. The PRISM trial, a randomized study at six US medical centers, is now testing AI-supported reading in US practice (see “What Comes Next in the US” below).
Reporting Assistance and BI-RADS Structuring
Reporting assistance AI in mammography supports the structured BI-RADS reporting workflow by extracting findings from the radiologist’s dictated text, populating the structured BI-RADS template fields (composition, findings, impression, BI-RADS assessment category, management recommendation), and ensuring the final report conforms to ACR BI-RADS reporting standards. This is the same reporting AI category covered for radiology generally in our guide to AI in radiology, applied specifically to the BI-RADS structured reporting context.
The clinical value of AI-assisted BI-RADS structuring is consistency. BI-RADS classification has documented interreader variability, particularly for the boundary cases between BI-RADS 3 (probably benign) and BI-RADS 4A (low suspicion for malignancy), and between density categories C and D. AI-assisted template population reduces variability by ensuring all required structured fields are completed and that the BI-RADS final assessment category aligns with the findings documented in the report body. The integration architecture for AI-assisted BI-RADS reporting follows the same pattern described in the Medicai structured radiology reporting guide.
What AI Can Do in Mammography: Measurable Benchmarks
Vendor marketing claims about AI mammography performance are abundant. Peer-reviewed clinical validation data is more limited but more useful. The benchmarks below come from published clinical studies, with sources cited.
Accuracy and reading time in a reader study. In a multireader study of iCAD ProFound AI for digital breast tomosynthesis, 24 radiologists read 260 DBT exams with and without AI (Conant et al., Radiology: Artificial Intelligence, 2019).
With AI, sensitivity rose from 77.0% to 85.0%, specificity rose from 62.7% to 69.6%, and the recall rate for non-cancer cases fell from 38.0% to 30.9%. Reading time per exam dropped by 52.7%.
These are reader-study results on an enriched case set, not screening outcomes, but they show the combination that matters clinically: better detection without more false positives.
AI-supported screening versus standard double reading. The Mammography Screening with Artificial Intelligence (MASAI) trial is the first randomized controlled trial of AI in breast cancer screening. It randomized about 106,000 women in Sweden’s national screening program to AI-supported screening or standard double reading by two radiologists.
In the AI arm, the AI score routed lower-risk exams to a single reader and higher-risk exams to two readers. Radiologists made every interpretation.
Earlier analyses reported a 29% higher cancer detection rate without more false positives, and about 44% less screen-reading workload.
The final results, published in The Lancet in 2026, met the trial’s primary goal:
- Interval cancers: non-inferior (1.55 versus 1.76 per 1,000 women). The AI arm had fewer, but the difference was not statistically significant.
- Sensitivity: significantly higher with AI (80.5% versus 73.8%).
- Specificity: the same in both groups (98.5%).
One caveat: sensitivity for in situ cancers was lower in the AI arm.
AI support for single readers. AI-STREAM is a prospective multicenter cohort study of 24,543 women in South Korea’s national screening program. It compared each radiologist’s reading with and without AI support (Lunit INSIGHT MMG).
Breast radiologists using AI detected 13.8% more cancers (5.70 versus 5.01 per 1,000 women), with no significant change in recall rate.
The effect depended on the reader. General radiologists using AI also detected more cancers, but recalled more women (6.89% versus 6.31%).
The published results are a preliminary analysis, with interval cancer data still to come.
Autonomous triage of low-risk exams. The AITIC trial (Nature Medicine, 2026) is the first large prospective test of letting AI clear low-risk screening exams without any human read across both 2D digital mammography and DBT. With AI labeling about 70% of exams as normal, radiologist workload fell 63.6%, and the cancer detection rate increased 15.2% relative to baseline, with more small, node-negative invasive cancers and more carcinomas in situ detected. The recall rate also rose, by 14.8% in relative terms, so the result is a workload and detection gain with a recall cost. The trial ran at a single site with one AI system and one mammography vendor, and one author is employed by the AI vendor, so results need confirmation elsewhere.
Cancer detection in dense breasts. Dense breasts are where conventional mammography has its best-known sensitivity limitation, because dense fibroglandular tissue can mask cancers in BI-RADS density categories C and D. That makes dense breasts one of the most important places to test AI. When evaluating a product, ask the vendor for performance results broken down by density category on the same equipment you use, rather than relying on an overall figure.
Interval cancer prediction. The 2025 RSNA report on AI for interval cancer identification describes ongoing work on AI models that identify women at elevated risk of developing breast cancer between screening rounds. Interval cancers, diagnosed after a negative screening mammogram but before the next scheduled screen, are a known limitation of population screening programs. AI identification of women at elevated interval cancer risk could enable supplemental imaging that captures these otherwise-missed cancers. MASAI is the first randomized trial to report interval cancers with AI-supported screening, with a rate that was non-inferior to standard double reading.
Breast density assessment consistency. Automated breast density tools demonstrate near-perfect intra-algorithm consistency (the same mammogram always produces the same density category), addressing the documented interreader variability in visual BI-RADS density assignment. Inter-algorithm consistency across different AI density tools is less perfect, with documented discordance in approximately 10 to 20 percent of cases at the C/D density boundary.
Where the Evidence Stands in 2026
Breast imaging is now the clearest example of AI moving from retrospective reader studies toward prospective clinical evidence. That shift matters because of what happened last time.
Computer-aided detection spread quickly in the early 2000s, and by 2016 more than 90% of US imaging centers were using it, before large studies showed it increased recall rates and reduced diagnostic accuracy in real-world practice. The current generation of AI is being tested far more rigorously before it becomes standard care.
The evidence now spans three levels of strength:
- Real-world implementation. In Germany’s PRAIM study, an observational study of 463,094 women across 12 screening sites, radiologists using AI-supported double reading detected 6.7 cancers per 1,000 women, compared with 5.7 without AI, a 17.6% increase, with no negative effect on recall rate.
- Prospective studies. AI-STREAM (South Korea, 24,543 women) and AITIC (Spain, 31,301 women) tested AI inside routine screening. Both increased cancer detection, and AITIC also showed the trade-offs of letting AI clear low-risk exams on its own.
- Randomized evidence. MASAI, with about 106,000 women in Sweden, is the first randomized controlled trial of AI in breast cancer screening. Its final results were published in The Lancet in 2026.
Key AI mammography studies at a glance
| Study | Where | Design and size | How AI was used | Key result | Status |
|---|---|---|---|---|---|
| PRAIM | Germany, 12 screening sites | Observational, real-world; 463,094 women | AI-supported double reading | 6.7 vs 5.7 cancers per 1,000 women (+17.6%); no negative effect on recall | Published (Nature Medicine, 2025) |
| AI-STREAM | South Korea | Prospective multicenter cohort; 24,543 women | Each radiologist read with and without AI | Breast radiologists: 13.8% more cancers, recall unchanged; general radiologists recalled more | Preliminary (Nature Communications, 2025) |
| AITIC | Spain, single site | Prospective; 31,301 women | AI cleared low-risk exams without a human read | More cancers detected; lower workload; higher recall | Published (Nature Medicine, 2026) |
| MASAI | Sweden | Randomized controlled trial; about 106,000 women | AI score routed lower-risk exams to one reader, higher-risk exams to two | Sensitivity 80.5% vs 73.8%; specificity 98.5% in both arms; interval cancers non-inferior | Final results (The Lancet, 2026) |
| PRISM | US, six centers in five states | Randomized; about 400,000 screening exams | Radiologist reads with or without an FDA-cleared AI tool | Primary outcomes: cancer detection and callback rates | Ongoing; no results yet |
These results are strong but not final. Most measure intermediate outcomes, such as cancer detection, recall, sensitivity, and interval cancers, rather than breast cancer deaths, and a 2026 Lancet correspondence debated how much benefit these endpoints show.
Most of the trials also ran in European double-reading programs, which differ from single-reader practice in the US.
What Comes Next in the US
The PRISM trial is designed to answer the question in US practice. Funded by the Patient-Centered Outcomes Research Institute with $16 million, it aims to include about 400,000 screening mammograms over two years at six medical centers in five states, with UC Davis Health leading the data analysis.
Each exam is randomized to a radiologist reading with or without an FDA-cleared AI tool (ScreenPoint Medical’s Transpara), and a radiologist makes every final decision. Most of the exams will be digital breast tomosynthesis, the standard at the participating centers.
The primary outcomes are cancer detection and callback rates; secondary outcomes include interval cancers, cancer types detected, and patients’ and radiologists’ trust in AI. Results are not yet available.
For a practical way to judge these studies, including why randomized evidence carries more weight than a reader study, see our radiology AI evidence scorecard.
Limitations and Regulatory Status
The capability benchmarks above are real and clinically validated. The limitations that follow are equally real and frequently understated in vendor marketing.
Current Limitations
Validation population versus deployment population mismatch. AI mammography models trained on data from one demographic, geographic, or screening population may underperform when deployed in different contexts. Models trained predominantly on Northern European screening populations may underperform on US Hispanic or African American populations where breast density distributions, body habitus, and disease prevalence patterns differ. The generalizability question must be answered for the specific deployment population, not assumed from the original validation study.
Performance variability across mammography hardware. AI mammography tools are trained on images acquired with specific digital mammography systems from specific manufacturers (GE Healthcare, Hologic, Siemens, Fujifilm, Philips). Performance on images from other manufacturers, older equipment generations, or systems with non-standard acquisition protocols may differ from validation benchmarks. Before deploying AI, the practice must verify that the AI vendor has validated performance on the specific imaging hardware in use.
Tomosynthesis versus 2D performance differences. Most AI mammography tools have separate validation data for 2D full-field digital mammography and 3D digital breast tomosynthesis. Clinical performance characteristics differ between modalities, and validation on one does not automatically transfer to the other. Practices migrating from 2D screening to DBT screening must verify that their AI tools are validated for DBT input.
The AITIC trial shows how large this difference can be. On 2D digital mammography, the AI strategy raised both cancer detection (by 33.7%) and recalls (by 28.2%). On DBT, detection and recall stayed essentially unchanged, and the main benefit was workload, which fell 65.5% because DBT exams take far longer to read.
Integration friction at deployment. An AI tool that performs well in a controlled validation study may underperform in clinical deployment because poor PACS viewer integration, added inference latency, or an output format that doesn’t match the radiologist’s reporting structure can undermine it. Procurement decisions consistently underestimate the performance gap between published validation and clinical deployment reality, in mammography as much as in general radiology AI.
Autonomous reading shifts accountability. When AI clears exams without radiologist review, any cancer it scores as low risk never reaches a human. In AITIC, the AI strategy auto-cleared 11 cancers that standard double reading detected, compared with 54 cancers found only by the AI strategy; most of the 11 had been flagged by just one of the two radiologists, suggesting subtle findings. The authors recommend automated image quality control and continuous post-market monitoring of AI performance before adopting any autonomous rule-out workflow, and they note that the legal and safety questions remain open.
Risk prediction model interpretability. Risk prediction AI tools produce a numerical risk score from a mammogram, but the imaging features the model uses to produce that score are not always interpretable in clinical terms. A woman told her mammogram suggests elevated 5-year cancer risk reasonably wants to know what about her mammogram drives that prediction. Current risk models offer limited interpretability into the features driving the risk score, complicating patient communication and shared decision-making.
Regulatory Status
The FDA AI/ML-Enabled Medical Devices database includes all FDA-cleared AI mammography products. Most mammography CAD products operate under 510(k) clearance, demonstrating substantial equivalence to previously cleared CAD predicate devices. The 2025 marketing authorization of Clairity Breast for breast cancer risk prediction represents a different regulatory pathway, reflecting the novel clinical claim of risk prediction (rather than detection) from screening mammography.
The FDA’s mammography regulation framework includes additional breast-imaging requirements beyond the general AI/ML SaMD framework. The Mammography Quality Standards Act (MQSA) governs mammography facility accreditation, radiologist qualifications, and reporting requirements. AI tools used in MQSA-accredited facilities must operate within the MQSA framework, including the 2024 updated mammography reporting requirements that mandate breast density communication to patients.
In the European Union, AI medical devices fall under the Medical Device Regulation (MDR) 2017/745, with most mammography AI tools classified as Class IIa or Class IIb requiring conformity assessment by a Notified Body. European mammography AI deployment is more advanced than in the US in some respects, with several large population screening programs (Sweden, the Netherlands, Germany) actively studying AI-supported screening as an alternative to traditional double reading. Spain’s Andalusian screening program has now tested automatic normal labeling prospectively in the AITIC trial. US practices should not assume the same model transfers directly: before changing reading workflow, confirm whether any rule-out use falls within the specific product’s FDA-cleared indications and how it fits MQSA interpretation requirements.
AI in Mammography Compared: Five Application Types at a Glance
| AI application type | What it does | Workflow position | Named FDA-cleared products | Clinical maturity and validation |
|---|---|---|---|---|
| Computer-aided detection (CAD) | Identifies and marks suspicious masses, microcalcification clusters, architectural distortions, and asymmetries on 2D digital mammography and digital breast tomosynthesis (DBT) studies for radiologist verification | During image review — AI overlays appear on the PACS viewer alongside the original mammographic images for the radiologist to review and verify | iCAD ProFound AI, Lunit INSIGHT MMG, ScreenPoint Transpara, Therapixel MammoScreen, Hologic Genius AI Detection | Most mature category — iCAD ProFound AI demonstrated 9.6% cancer detection rate increase with 5.6% recall reduction in Radiology validation studies; MASAI trial in The Lancet Oncology demonstrated AI-supported single reading reduced radiologist workload by 44% while increasing detection rate |
| Cancer risk prediction | Predicts 1-year or 5-year breast cancer risk from imaging features detectable in a current screening mammogram, supplementing traditional risk models including Tyrer-Cuzick and Gail | After image review — produces a numerical risk score that informs screening interval personalisation and supplemental imaging decisions (breast MRI, contrast-enhanced mammography) | Clairity Breast (FDA marketing authorisation 2025), Mirai (research-stage, MIT CSAIL) | Emerging — Clairity Breast is the first widely-deployed FDA-cleared risk prediction product; clinical integration pathways still developing as practices establish protocols for routing high-risk patients to supplemental imaging |
| Breast density assessment | Produces automated BI-RADS density category assignment (A almost entirely fatty, B scattered, C heterogeneously dense, D extremely dense) from the screening mammogram, addressing documented interreader variability in visual density assessment | During image review or as a pre-processing step — density category delivered to the radiologist as part of the structured BI-RADS report | VolparaDensity (Volpara Health), iCAD ProFound Density, integrated density modules across most major mammography AI platforms | Established — widely deployed across US screening practices; FDA 2024 updated mammography reporting requirements mandate patient density communication, making consistent density assessment a regulatory as well as clinical priority |
| Workflow triage and worklist prioritisation | Reorders the screening reading worklist based on case-level malignancy scores, prioritising high-suspicion studies for earlier radiologist review and reducing time-to-callback for suspected cancers | Before radiologist review — processes incoming studies within seconds of acquisition and reprioritises the worklist before the radiologist opens any study | iCAD ProFound AI worklist triage, Lunit INSIGHT MMG with workflow integration, ScreenPoint Transpara workflow tools | Operationally significant in high-volume screening centres — clinical evidence strongest in European population screening programmes deploying AI as one of the two readers in double-read screening; US single-reader workflows use triage as supplementary rather than reader replacement. Prospective AITIC trial (Nature Medicine, 2026): AI cleared about 70% of exams as normal, workload down 63.6%, detection up 15.2%, recall up 14.8%. |
| Reporting assistance and BI-RADS structuring | Structures the BI-RADS report, populates template fields from dictated text, ensures the BI-RADS final assessment category (0 through 6) aligns with the documented findings in the report body | After image review — processes the radiologist’s findings and dictation into the structured BI-RADS report delivered to the ordering physician’s EHR via HL7 ORU integration | Nuance PowerScribe (radiology-optimised dictation), AI-enhanced BI-RADS template tools integrated into modern cloud PACS platforms | Mature for dictation, emerging for AI-assisted structured BI-RADS field population — same integration architecture as general radiology structured reporting; addresses documented BI-RADS classification interreader variability particularly at the BI-RADS 3/4A boundary and density C/D boundary |
The PACS Infrastructure Requirement for AI Mammography Deployment
AI mammography tools do not operate in isolation. Each of the five application categories described above requires technical integration with the breast imaging practice’s existing PACS and reporting infrastructure to be usable in clinical workflow. The integration requirements are specific, and practices evaluating AI mammography procurement consistently underestimate the infrastructure work involved.
DICOM Mammography Image Storage
Digital mammography images use the DICOM Mammography (DICOM-MG) Information Object Definition, a DICOM extension with mammography-specific metadata fields, including standard view positions (right craniocaudal, left craniocaudal, right mediolateral oblique, left mediolateral oblique), implant displacement views, and magnification views. Digital breast tomosynthesis uses the DICOM Breast Tomosynthesis Image Storage SOP class and produces significantly larger studies than 2D mammography (a single DBT screening study typically ranges from 800 MB to 2 GB). Both formats must preserve full metadata for AI tools to operate correctly, as AI uses view-position metadata to determine which images to compare for laterality assessment, year-over-year change detection, and structured reporting.
AI Integration via DICOMweb or Vendor APIs
AI mammography tools integrate with the PACS through three primary architectures. DICOM-based integration with the AI vendor’s on-premise or cloud server, where the PACS forwards mammographic studies via DICOM C-STORE and receives the AI results back via DICOM Structured Reports or DICOM Secondary Capture overlays. DICOMweb integration using WADO-RS for study retrieval, QIDO-RS for worklist queries, and STOW-RS for AI result storage is the standard architecture for cloud-native AI tools. Vendor-specific REST API integration for AI tools that operate outside standard DICOM workflows. The PACS must support whichever integration pattern the AI vendor uses, and the integration must be tested for both 2D mammography and DBT inputs.
Routing and Audit for AI Triage
Autonomous or semi-autonomous triage is as much a routing problem as an AI problem. The PACS and worklist must apply the AI score as a rule: route low-risk exams to a normal-result path, route higher-risk exams to the reading worklist with AI overlays, and maintain an audit trail of every score and routing decision for quality assurance and post-market monitoring. Image transfer reliability also matters: in AITIC, exams the AI could not process, including some affected by erroneous PACS image transfer, had to be excluded from the trial. A cloud PACS that stores DICOM mammography and DBT studies with intact metadata, exposes them to the AI through DICOMweb, and logs AI results alongside each study is what makes this kind of workflow auditable.
Hanging Protocol Configuration for AI Output Display
Mammography reading uses standardized hanging protocols that present the four standard views in a specific arrangement (typically right CC and left CC in the top row, right MLO and left MLO in the bottom row, with prior comparison studies displayed on a second monitor). When AI tools produce overlays or annotations, configure the hanging protocol to display the AI output alongside the original images without disrupting the radiologist’s established reading pattern. Configuring the hanging protocol for AI integration is consistently underestimated. The Medicai hanging protocols guide covers the general principles of PACS hanging protocol configuration.
BI-RADS Reporting Integration
The AI output (case-level malignancy score, per-image findings, density category, risk score) must flow into the structured BI-RADS report delivered to the ordering physician. This requires the PACS reporting platform to accept AI-generated findings as structured input and to format them into the standard BI-RADS report fields. Direct integration between AI tools and reporting platforms is increasingly the standard architecture; older deployments that manually transcribe AI results into reports create workflow risks and productivity bottlenecks.
Medicai’s cloud-native PACS on Microsoft Azure provides the infrastructure layer that supports AI mammography integration: DICOM-MG and DICOM Breast Tomosynthesis storage with full metadata preservation; DICOMweb support for cloud-native AI tool integration; hanging protocol configurability for AI output display; and structured reporting integration with BI-RADS template support. For breast imaging centers evaluating AI mammography deployment, the underlying PACS infrastructure determines whether the AI tool delivers its theoretical performance in the actual reading workflow, or whether integration friction erodes the clinical benefit.
Frequently Asked Questions
AI in mammography is the application of machine learning algorithms, primarily deep neural networks, to assist with detecting breast cancer, assessing breast cancer risk, measuring breast density, triaging screening worklists, and structuring diagnostic BI-RADS reports. AI in mammography does not replace the radiologist. It operates as a computational layer within the breast imaging workflow, processing the digital mammogram and producing outputs (detections, risk scores, density categories, structured report fields) that the radiologist verifies and incorporates into the final clinical interpretation.
Yes, AI is in active clinical use across many US and European breast imaging practices, with adoption accelerating since 2020. The most widely deployed AI applications are computer-aided detection of suspicious findings, automated breast density assessment, and workflow triage of screening worklists. The FDA has cleared over a dozen mammography-specific AI products, with new clearances continuing through 2024 and 2025. The clinical deployment model is consistently radiologist-in-the-loop: the AI produces findings, the radiologist reviews and verifies them, and the radiologist remains responsible for the final diagnostic interpretation and BI-RADS assessment category.
The clinical and research consensus is that AI augments rather than replaces breast radiologists. AI excels at high-volume, narrow-task pattern recognition on screening mammography, where consistent application of detection criteria across thousands of studies is exactly the task computers do well. AI underperforms on atypical presentations, complex multi-finding studies, patient communication about screening results, supplemental imaging recommendations for high-risk patients, and the diagnostic workup of suspicious findings identified at screening. The European MASAI trial demonstrated that AI-assisted single reading can match the cancer detection rate of traditional double reading while reducing reader workload, but the AI in that trial operated as one of the two readers, not as a replacement for radiologist judgment.
The 2026 AITIC trial went further, letting AI clear about 70% of screening exams with no human read, but radiologists still double-read every exam the AI flagged and made every recall decision.
For a defined subset of exams, in research settings, yes. In the 2026 AITIC trial, AI classified about 70% of screening exams as low risk, and those exams received no radiologist reading, while radiologists double-read the remaining exams with AI support and made every recall decision. Cancer detection rose and reading workload fell by about two-thirds, but recalls increased, and the approach was tested at a single European site. In routine US practice, a radiologist interprets screening mammograms, and AI functions as detection and triage support.
FDA-cleared AI mammography tools have demonstrated improvements in cancer detection rates of approximately 9 to 10 percent when used alongside radiologist reading, compared with radiologist reading without AI assistance. The iCAD ProFound AI validation in Radiology showed a 9.6 percent increase in cancer detection rate and a simultaneous 5.6 percent reduction in recall rate. The MASAI trial in The Lancet Oncology showed AI-supported single reading produced higher cancer detection rates than traditional double reading. Performance varies by product, patient population, mammography hardware, and the specific clinical workflow in which the AI is deployed. Practices evaluating AI mammography should request validation data specific to their patient demographics and imaging equipment, rather than relying on aggregate marketing claims.
FDA-cleared AI mammography products include iCAD ProFound AI (detection, density, risk on 2D and DBT), Lunit INSIGHT MMG (detection on 2D), ScreenPoint Medical Transpara (detection on 2D and DBT), Therapixel MammoScreen (detection on 2D), Hologic Genius AI Detection (detection on DBT), VolparaDensity (breast density assessment), and Clairity Breast (5-year cancer risk prediction, FDA-cleared 2025). The full list is maintained in the FDA AI/ML-Enabled Medical Devices database and is updated regularly as new clearances are issued. New products receive clearance through both the 510(k) substantial equivalence pathway and the De Novo pathway for novel risk prediction claims.
AI mammography tools integrate with the PACS through three main architectures: DICOM-based integration, where the PACS forwards studies to the AI server via DICOM C-STORE and receives results back as DICOM Structured Reports or Secondary Capture overlays; DICOMweb integration using WADO-RS, QIDO-RS, and STOW-RS for cloud-native AI tools; and vendor-specific REST API integration for AI tools operating outside standard DICOM workflows. The integration must support both 2D digital mammography (DICOM-MG SOP class) and digital breast tomosynthesis (DICOM Breast Tomosynthesis Image Storage SOP class), with full preservation of mammography-specific metadata fields including view positions, laterality, and acquisition parameters. Modern cloud-native PACS platforms support all three integration patterns and provide the hanging protocol configuration and BI-RADS reporting integration needed to deliver AI outputs into the radiologist’s reading workflow.
Sources
Related Articles
Lets get in touch!
Learn more about how Medicai can help you strengthen your practice and improve your patients’ experience. Ready to start your Journey?
Book A Free Demo