Exact FDA record: K220105 · Agent source review · . This is not clinician review or local acceptance.
Within this record's scope
Consider for concurrent screening DBT review in women aged 35 or older, using the labeled DBT stack and accompanying 2D images.
Limits to check
Full physician review remains required. This 2022 submission does not establish the current commercial release or support for another input configuration.
Questions for the vendor and your team
Which installed version and licensed functions match the quoted Saige-Dx release, and what is its exact regulatory record?
How will readers evaluate finding marks, suspicion levels and false findings on the department's representative screening workflow?
These questions are prompts for evaluation, not claims that licensing, compatibility, performance or price has been confirmed.
A tag identifies a documented role; it does not establish compatibility or local acceptance. Methods and sources
What does this software do?It analyzes a single screening DBT study, localizes suspected soft-tissue lesions and calcifications, and provides finding- and case-level suspicion information for physician review.
Which examinations and patients are in scope?It is intended for women aged 35 years or older undergoing screening mammography with compatible DBT hardware, using DBT stacks with accompanying full-field digital mammography and/or synthetic 2D images.
Who uses it?Interpreting physicians qualified to read screening mammography examinations use the output during concurrent review.
What does the software produce?It produces finding-level boxes and suspicion levels, an aggregate case suspicion level, a DICOM Structured Report for workstation overlay, and a DICOM Secondary Capture summary report.
What is its most important limitation?It is an adjunct, not a replacement for full physician review or clinical judgment, and decisions must not be made solely from its analysis.
Clinical takeaway. Saige-Dx is a concurrent-reading aid for screening DBT mammography. Its marks and suspicion information support, but do not replace, physician interpretation.
Evidence takeaway. The FDA submission describes separate reader and standalone studies. In the 240-case reader study, aided reader AUC was 0.925 versus 0.865 unaided; this reader-aided result is not a standalone sensitivity or specificity measure.
Additional limitations
The exact-submission materials do not establish a current commercial version, local compatibility, or site-specific workflow configuration.
Evidence context: questions, not a score
Agent transcription of regulatory evidence context, not independent validation, clinical acceptance or a product recommendation. Separate studies and outcomes are not a leaderboard.
Existing study anchor: Saige-DX regulatory performance study summary. Tested version label: Exact product version not reported; regulatory study summarized for FDA submission K220105.
What was tested?
Two different retrospective tests: a 240-case, 18-reader aided-versus-unaided MRMC study and a 1,304-case standalone test.
Source section: 510(k) summary, section 7: Performance Data, Reader Study and Standalone Study
On which cases and versions?
Reader cases used Hologic DBT, women 35 or older, 100 cancer and 140 non-cancer. The standalone set came from nine US sites with 136 cancer and 1,168 non-cancer. Do not add these into a single study sample.
Source section: 510(k) summary, section 7: Performance Data, Reader Study and Standalone Study
What changed?
The FDA summary reports increased reader ROC area with assistance. The reported reader-specificity difference has a confidence interval spanning zero, so a specificity improvement is not established. Standalone results remain separate from reader effects.
Source section: 510(k) summary, section 7: Performance Data, Reader Study and Standalone Study
What does this not establish?
Reader sensitivity changes are not standalone sensitivity, and previously withdrawn metrics remain withdrawn. The study does not establish reduced recalls, less rework, patient outcomes, performance on all vendors or local throughput.
Source section: 510(k) summary, section 7: Performance Data, Reader Study and Standalone Study
Study design, population and missing context
evidence Role
Retrospective MRMC reader study and separate retrospective standalone evaluation.
reader study · 510(k) summary, section 7: Performance Data, Reader Study and Standalone Study
design
Reader study: 18 radiologists read 240 cases twice, aided and unaided, with at least four weeks washout. Standalone: blinded multicenter retrospective study of 1,304 cases.
reader study · 510(k) summary, section 7: Performance Data, Reader Study and Standalone Study
setting
Not established: The reviewed regulatory summary does not establish this field for this study context.
training Population
Training used six datasets from US and UK locations. Validation data came from different sites and did not overlap training.
training · 510(k) summary, section 7: Performance Data, Reader Study and Standalone Study
evaluation Population
Reader set: 240 women age 35 or older, 100 cancer and 140 non-cancer. Standalone set: 1,304 cases, 136 cancer and 1,168 non-cancer. The sets are not combined into one denominator.
reader study · 510(k) summary, section 7: Performance Data, Reader Study and Standalone Study
reference Standard
Pathology and expert reference review; two breast-imaging specialists and a third adjudicator for disagreements in the reader-study truthing process.
reader study · 510(k) summary, section 7: Performance Data, Reader Study and Standalone Study
comparator
Reader interpretation with versus without Saige-Dx. Standalone performance is a separate endpoint, not the aided-reader effect.
reader study · 510(k) summary, section 7: Performance Data, Reader Study and Standalone Study
sites
The standalone study used nine US clinical sites. Reader-study site count is not provided in the reviewed text.
validation · 510(k) summary, section 7: Performance Data, Reader Study and Standalone Study
scanner Protocol
Reader cases were Hologic DBT with accompanying 2D mammograms. Do not extend the study result to every DBT manufacturer.
reader study · 510(k) summary, section 7: Performance Data, Reader Study and Standalone Study
sample Composition
Reader cases: 100/240 cancer, including 67 cancers recalled and 33 not recalled in routine practice; 140/240 non-cancer. Standalone: 136/1,304 cancer and 1,168/1,304 non-cancer.
reader study · 510(k) summary, section 7: Performance Data, Reader Study and Standalone Study
sponsor Involvement
DeepHealth describes the training/test separation and submitted the manufacturer regulatory study package. Investigator independence is not established by data separation.
regulatory summary · 510(k) summary, section 7: Performance Data, Reader Study and Standalone Study
Public-label constraints, not patient-level eligibility or a medical recommendation. Training and evaluation populations are separate. Read the original labeling and confirm the exact release and protocol.
Concurrent reading aid for interpreting physicians on screening mammograms with compatible DBT hardware; it analyzes DBT stacks and accompanying full-field digital or synthetic 2D images to identify soft-tissue lesions and calcifications that may indicate cancer and assigns finding- and case-level suspicion levels.
DICOM x-ray mammogram files from a single DBT study, including DBT image stacks and accompanying full-field digital mammography and/or synthetic 2D images.
Finding-level bounding boxes and Finding Suspicion Levels, an aggregate Case Suspicion Level, a DICOM Structured Report with markings for workstation overlay, and a DICOM Secondary Capture summary report.
The FDA labels Saige-Dx as an adjunct and not a replacement for full physician review; the patient population is women aged 35 or older undergoing screening. Reader-study and standalone results are separate, and the packet does not establish current commercial version or local compatibility.
The balanced MRMC reader study included 18 MQSA-qualified radiologists reading 240 cases twice, with 100 pathology-proven cancers and 140 confirmed non-cancers, using two expert truthers and adjudication when needed. Reader AUC was 0.865 unaided versus 0.925 aided (difference 0.06, 95% CI 0.041-0.079, p<0.00001); average reader sensitivity increased by 8.8% (95% CI 7.0%-10.6%) and specificity by 0.9% (95% CI -0.9% to 2.7%). A separate standalone study used 1,304 cases from nine U.S. sites and reported AUC 0.930 (95% CI 0.902-0.958). Reader deltas and AUC are not standalone sensitivity or specificity.
Metrics are shown only when they are tied to a source, endpoint, population, and tested product version. Candidate literature without exact product and version linkage is not shown as product evidence.
Evidence summary
The balanced MRMC reader study included 18 MQSA-qualified radiologists reading 240 cases twice, with 100 pathology-proven cancers and 140 confirmed non-cancers, using two expert truthers and adjudication when needed. Reader AUC was 0.865 unaided versus 0.925 aided (difference 0.06, 95% CI 0.041-0.079, p<0.00001); average reader sensitivity increased by 8.8% (95% CI 7.0%-10.6%) and specificity by 0.9% (95% CI -0.9% to 2.7%). A separate standalone study used 1,304 cases from nine U.S. sites and reported AUC 0.930 (95% CI 0.902-0.958). Reader deltas and AUC are not standalone sensitivity or specificity.
Previous value withdrawn: The old value is a change in reader sensitivity, not standalone sensitivity.
Reported specificity
Value withdrawn pending re-verification
Corrected record
Exact FDA submission
Checked 2026-09-12
Previous value withdrawn: The old value is the upper confidence bound for the reader-sensitivity change, not standalone specificity.
No contextualized exact-version metric has completed evidence review. Regulatory-document values, when available, are shown above with their limitations.
Saige-DX regulatory performance study summary
DesignStand-alone performance and reader study (ACR AI Central category)
PopulationPopulation described in the exact-submission ACR model card
Scopeexact submission
Tested versionExact product version not reported; regulatory study summarized for FDA submission K220105
Samplen=1304
Sites9
IndependenceIndependence not established
K220105
ACR AI Central summarizes the regulatory study; consult the exact FDA materials before comparative use.
Recall and adverse-event records are shown only after product matching. Adverse-event reports do not establish incidence or causality.
Postmarket safety review
The exact-identifier FDA recall query returned no native recall record for K220105. This limited result does not establish absence of recalls or other safety information and does not cover later versions, family records, MAUDE reports, corrections, or field notices.
IT implementationDeployment, interfaces, data handling, and support
System and version relationships
Match the AI release, system build, configuration and region. Source documentation does not establish acceptance of an installed system.
No specific system/version pairing has completed review for this FDA record. A broad DICOM or product-family statement does not establish a named viewer build's support.
Public source context, not tested interoperability. Each hop needs its own release/configuration evidence. Missing or unavailable routes are not documented incompatibility. All local tests are planned, not passed; no installed-system acceptance is recorded.
Acquisition/source → routing → AI processing → returned objects → viewer/worklist/report → support owner
No specific route context has been source-reviewed for this record. Obtain both endpoint builds and documents; this is not a finding of incompatibility.
Acquisition/source
Source context, not route validation: DICOM x-ray mammogram files from a single DBT study, including DBT image stacks and accompanying full-field digital mammography and/or synthetic 2D images.
Scope: Exact FDA submission.
Confirm scanner/model, protocol, series completeness and both software builds.
Source context, not route validation: DICOM Structured Report and Secondary Capture outputs are intended for overlay and review on a viewing workstation with compatible DBT hardware; the packet does not establish specific PACS, RIS, EHR, or interface versions.
Scope: Exact FDA submission.
Name gateways, destinations, transfer rules and failure visibility; a DICOM statement does not prove a route.
Unresolved route: Not established for the exact route.
Scope: Exact FDA submission. Previous value withdrawn: The exact FDA decision-summary packet reviewed on 2026-11-12 does not establish this claim; it states no deployment or hosting model.
Confirm processing location, queue ownership, exclusions and exact AI build.
Field checked 2026-11-12 ·
Returned objects
Source context, not route validation: Finding-level bounding boxes and Finding Suspicion Levels, an aggregate Case Suspicion Level, a DICOM Structured Report with markings for workstation overlay, and a DICOM Secondary Capture summary report.
Scope: Exact FDA submission.
Confirm object classes, units, references and whether findings, overlays or measurements reach their intended destination.
Unresolved route: Finding-level bounding boxes and Finding Suspicion Levels, an aggregate Case Suspicion Level, a DICOM Structured Report with markings for workstation overlay, and a DICOM Secondary Capture summary report.
Scope: Exact FDA submission.
Confirm the named viewer/worklist/report build, display behavior and report transfer. Output support alone does not establish this endpoint. The displayed source describes output, not validated end-to-end delivery.
Unresolved route: Not established for the exact route.
Scope: Scope not established.
Name the local, platform and AI supplier support owners and escalation path.
Planned local route tests
No tests below have been performed or accepted.
Patient/study identity (planned). Use approved non-patient test data to verify identifiers, accession and result-to-study linkage at every hop.
Study completeness (planned). Test missing views/series, partial transfer, late arrivals, priors and unsupported acquisition inputs; confirm explicit refusal or fallback.
Ordering and duplicates (planned). Send out-of-order, repeated and corrected studies; verify worklist order, deduplication and no misleading stale results.
Report transfer (planned). Verify output object types, annotations, units, report content, destination and availability of original images.
Access (planned). Verify authorized roles, audit trail, denied access and support ownership at each destination.
Latency and failures (planned). Agree a workload and turnaround target; test queues, timeouts, missing results and fallback reading.
Updates and rollback (planned). Define version change control, revalidation triggers, stop-use criteria and a tested rollback route.
Local implementation checklist
Agree the evidence and acceptance owner with your team and suppliers. Checkmarks apply to this page only and are not acceptance receipts.
DICOM Structured Report and Secondary Capture outputs are intended for overlay and review on a viewing workstation with compatible DBT hardware; the packet does not establish specific PACS, RIS, EHR, or interface versions.
Sources & review detailsFDA identity, research record, source ledger, and history
Regulatory identity
FDA record and catalog context
The FDA listing establishes the regulatory identity. It does not by itself establish local workflow fit, pricing, security, or performance in your environment.
This audit distinguishes completed source review from fields that have not yet been researched.
Exact FDA record reviewedStatus
2026-09-09Last searched
25Fields reviewed
9Source classes checked
1Awaiting review PubMed leads
0Awaiting review trial leads
0Unreviewed FDA recall leads
Exact-submission ACR model-card fields were normalized under the current provenance rules; vendor and independent-study confirmation remain distinct. Exact FDA scope controls clinical claims. ACR provides bounded exact-submission catalog context. Candidate literature metadata remains a discovery queue until full-text identity, endpoint, population, and version review is completed. Automated exact-name discovery found 1 PubMed and 0 ClinicalTrials.gov candidate records. Candidates require human product and version matching; zero candidates is not evidence that no studies exist. Native FDA recall identifiers produced 0 postmarket candidate records; 0 have been reviewed (0 published, 0 rejected) and 0 remain unreviewed.
Candidate leads remain unpublished until a human confirms the exact product and tested version.
Source classes: fda ai list, fda decision summary, acr ai central product, fda device recall, literature index, trial registry, pubmed, clinical trials, openfda device recall
Field review details
Recorded state, verification date, notes and sources for fields without a publishable value. Conflicting records stay explicitly unsettled here; no plain settled value is shown for them.
Public pricing: Not verified - checked 2026-09-09 - Exact FDA submission
No public price or quote for the exact K220105 configuration was established in this review.
Pricing model: Not yet reviewed
Setup and integration fees: Not yet reviewed
Recurring fees: Not yet reviewed
Minimum commitment: Not yet reviewed
Contract duration: Not yet reviewed
Renewal terms: Not yet reviewed
Cancellation terms: Not yet reviewed
Support and service terms: Not yet reviewed
Trial or pilot terms: Not yet reviewed
Reimbursement: Not publicly available - checked 2026-09-09 - Exact FDA submission
No exact-product payer policy, coding instruction, or payment determination was established; eligibility must be evaluated by payer, site of service, and use case.
Deployment: Withdrawn - checked 2026-11-12 - Exact FDA submission
Previous value withdrawn: The exact FDA decision-summary packet reviewed on 2026-11-12 does not establish this claim; it states no deployment or hosting model.
Data hosting location: Not yet reviewed
Data retention: Not yet reviewed
Documented data use: Not yet reviewed
Hardware requirements: Not yet reviewed
Network requirements: Not yet reviewed
Identity and access: Not yet reviewed
Security: Not verified - checked 2026-09-09 - Exact FDA submission
Exact-release security controls, data retention, hosting boundaries, and scoped assurance documentation require vendor confirmation.
Installation timeline: Not yet reviewed
Training: Not verified - checked 2026-09-09 - Exact FDA submission
Current training prerequisites, competency checks, and support commitments for the exact configuration require vendor confirmation.
Monitoring: Not verified - checked 2026-09-09 - Exact FDA submission
Current drift, quality, uptime, alert, escalation, and incident-response commitments for the exact configuration require vendor confirmation.
Authorized purpose: Flags calcifications, masses, asymmetries and architectural distortions on compatible FFDM or DBT screening examinations; management must not be based solely on MammoScreen output.
Authorized purpose: BAC processes screening FFDM and DBT mammograms to detect the presence or absence of breast arterial calcifications at study and breast level and localize detected calcifications for qualified interpreting-physician review.
Authorized purpose: Tags screening cases as suspicious or blank to aid worklist prioritization; radiologists must review every examination and it does not mark original images.
Copy this public context and add the field, proposed wording and original source before sending. Nothing here is submitted automatically. Form-specific prefill has not been verified.
Permission to display a submission is not evidence verification. A submission enters editorial review; it does not approve a public change. Form availability, data entry, receipt, review and publication are separate steps. Data entry and receipt remain unverified.
Accepted repository change history
No entry has been added to this scoped change history for this record. This is not proof that the record never changed.