What problem does it solve?
Organized breast cancer screening invites large numbers of healthy women every two or three years, and many programs, especially in Europe, have two radiologists read every examination. Almost every examination is normal, so most of the reading effort goes into confirming that nothing is there, while the few cancers have to be found among them. Programs that must keep double reading as a safeguard run short of breast radiologists before they run short of women to screen.
AI changes how that reading time is spent. The system gives each examination a risk score and marks suspicious areas. Programs use the score in two ways: to send examinations that are likely normal to one reader instead of two, and to show the marked areas to the radiologists who read the rest. The evidence on this page comes from programs that kept a radiologist in every reading and compared cancer detection, recall and workload with standard double reading: a randomized trial in Sweden, a before and after evaluation in Denmark and an observational study at 12 German sites.
How does it work?
- Score each examination. After the mammogram is taken, an image analysis model assigns a risk score and marks suspicious lesions and calcifications.
- Route by score. Examinations below a set threshold are read by one senior breast radiologist. Examinations above it go to two radiologists, who see the AI markings as decision support. Some programs keep double reading for all examinations and use the AI only as a second look.
- Read and decide. The radiologists decide whether to recall the woman for assessment. The AI never recalls or discharges anyone on its own.
- Monitor the program. Cancer detection, recall, false positives, interval cancers and the reading workload are tracked against the period before AI and checked on a regular schedule.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Early adopters
- Channels
- Internal tools
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Productivity gain | Too few to pool | 33.5% to 44.2% | 2 | 2 organization |
| Detection improvement | Too few to pool | 17.6% | 1 | 1 organization |
Value drivers: Employee productivity, Risk and loss reduction, Speed and cycle time.
Indicative value
A regional breast screening program with 60,000 screening examinations a year
USD 120,000 to USD 594,000
Radiologist reading cost released per year
How this is calculated
Formula: examinations * readsPerExam * workloadReduction * costPerRead. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Screening examinations per year examinations, examinations | 60,000 | 60,000 | The reference program. |
| Radiologist reads per examination before AI readsPerExam, reads per examination | 2 | 2 | Standard double reading, as in the Danish and Swedish programs on this page. |
| Share of reads saved workloadReduction, fraction of reads | 0.2 | 0.33 | Conservative against the benchmarks on this page (the Capital Region of Denmark reports a 33.5% reduction in reading workload and the Swedish MASAI trial 44.2%), because a first launch often keeps more double reading. |
| Cost of one radiologist read costPerRead, USD per read | 5 | 15 | Editorial assumption, replace with your own fully loaded radiologist cost per examination read. |
What it leaves out: Gross reading cost only. It leaves out the license and integration cost of the AI, the extra assessment work for any additional cancers found, and the benefit of earlier detection.
Who already uses it?
3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Capital Region of Denmark
Denmark · Healthcare · 2021
The Capital Region of Denmark brought an AI system into its biennial mammography screening of women aged 50 to 69 on 2021-11-18. The system scores every screening examination. Examinations the AI rates as likely normal are read once by a senior breast radiologist, and the rest are read by two radiologists with AI decision support. An evaluation by the University of Copenhagen and the region's screening program, published in Radiology, compares 60,751 women screened before AI with 58,246 screened after. Recall fell, cancer detection rose and the reading workload dropped by a third.
- Productivity gain: 33.5%, Saved screening reads after AI implementation, 2021-11-18 to 2022-10-17, versus the double reading period before
"The reading workload was reduced by 33.5% (38 977 of 116 492 reads)."
Claimed by: organization
German mammography screening program (PRAIM study sites)
Germany · Healthcare · 2021
In the PRAIM implementation study, 119 radiologists at 12 sites of the German organized mammography screening program screened 463,094 women aged 50 to 69 between July 2021 and February 2023, 260,739 of them with AI support. Radiologists chose voluntarily whether to use the AI system in a double reading. The AI supported group had a higher breast cancer detection rate and a recall rate that was not worse than standard double reading. The study is observational, not randomized, and was published in Nature Medicine in 2025.
- Detection improvement: 17.6%, Breast cancer detection rate, AI supported double reading versus standard double reading, July 2021 to February 2023
"Radiologists in the AI-supported screening group achieved a breast cancer detection rate of 6.7 per 1,000, which was 17.6% (95% confidence interval: +5.7%, +30.8%) higher than and statistically superior to the rate (5.7 per 1,000) achieved in the control group."
Claimed by: organization
Mammography screening programme in southwest Sweden (Malmö, Lund, Landskrona, Trelleborg)
Sweden · Healthcare · 2021
In the Mammography Screening with Artificial Intelligence trial (MASAI), women eligible for the Swedish national screening programme at four sites in southwest Sweden (Malmö, Lund, Landskrona and Trelleborg) were randomly assigned to AI supported screening or standard double reading. The AI system, Transpara version 1.7.0 from ScreenPoint Medical, triaged each examination to single or double reading and highlighted suspicious findings. Between April 2021 and December 2022, 105,934 women were randomized. AI supported screening detected more cancers without a significant rise in recalls or false positives and cut the screen reading workload. The trial, run by Lund University researchers, is published in The Lancet Digital Health.
- Productivity gain: 44.2%, Screen readings in the intervention group versus the control group, April 12, 2021 to December 7, 2022
"There were 61 248 screen readings in the intervention group and 109 692 in the control group, resulting in a 44·2% reduction in the screen-reading workload."
Claimed by: organization
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Screening examinations in DICOM format with the program's own outcome data
- A written reading protocol that says which examinations get one reader and which get two
- Follow up data on biopsies, surgery and interval cancers to audit the thresholds
Systems to integrate
- PACS and the screening program information system
- Radiologist reading workstation, so the markings show during the read
- Recall and invitation system of the screening program
Complexity: High
The image analysis model is a certified medical device bought from a vendor, so the build is the connection to the PACS and the screening information system, the reading protocol with its thresholds, a clinical validation on the program's own population and ongoing audit of cancers and interval cancers.
- 1
Validate on your own population
Run the AI silently on past examinations with known outcomes and compare its scores with the radiologists' reads and the later cancers, before it influences any reading.
- 2
Write the reading protocol
Decide the score thresholds, who reads which examinations and how a disagreement between the AI and a radiologist is settled. Put the protocol under clinical governance.
- 3
Start with decision support
Show the markings to the radiologists first while every examination is still double read, then move likely normal examinations to single reading once the numbers hold.
- 4
Audit on a schedule
Review cancer detection, recall, false positives and interval cancers every quarter and adjust the threshold when the numbers move.
Guardrails
- A radiologist makes every recall decision, never the AI
- A written protocol with fixed thresholds that only the clinical lead can change
- Interval cancer review so missed cancers in the single read group are found
- Separate monitoring of results by age group and screening round
KPIs to instrument
- Cancer detection rate per 1,000 screens, compared with the period before AI
- Recall rate and false positive rate
- Interval cancer rate and stage at diagnosis
- Screen reading workload in reads per examination
Human in the loop
Breast radiologists read every examination, alone or in pairs, and own the recall decision. The clinical lead owns the threshold and reviews the audit of detected and missed cancers.
Common failure modes
- Radiologists follow the score too closely
- Readers accept a low score and stop looking. Prevent with training, blinded sampling and a review of cancers missed in the single read group.
- Threshold drift
- A threshold set on one population or scanner performs differently on another. Recheck on the program's own data after every change to the equipment or the invited group.
What are the risks and rules?
EU AI Act
High risk
Article 6(1) with Annex I: an AI system that is a medical device under the Medical Device Regulation (EU) 2017/745, or a safety component of one, is high risk when the device needs a third party conformity assessment. Software that analyzes mammograms to support diagnosis is normally class IIa or higher, which needs a notified body, so it carries the AI Act duties for risk management, data governance, logging and human oversight on top of the medical device route. It is not an Annex III use.
Guidance
- Article 6, classification rules for high risk AI systems (European Union, Europe). An AI system that is a medical device, or a safety component of one, is high risk when a third party conformity assessment is required.
- Regulation (EU) 2017/745 on medical devices (European Union, Europe). Software that provides information used to take decisions with diagnostic purposes is classified under Annex VIII Rule 11, usually as class IIa or higher, which requires a notified body conformity assessment before CE marking.
Controls to put in place
- Documented clinical validation of the model on the program's population before use
- Reading protocol under clinical governance with logged threshold changes
- Audit trail that shows the score, the readers and the recall decision per examination
- Regular review of detected and interval cancers
Frequently asked questions
- Does AI replace the second radiologist in breast screening?
In the programs on this page it does not replace a radiologist. In the Capital Region of Denmark, likely normal examinations go to one senior breast radiologist and the others are read by two with AI support, so a person makes every recall decision. The region reports that this reduced the reading workload by 33.5%.
- Does AI find more cancers than standard double reading?
The German PRAIM study reports a breast cancer detection rate of 6.7 per 1,000 with AI supported double reading against 5.7 per 1,000 with standard double reading, with a recall rate that was not worse. The study is observational, so programs should audit interval cancers on their own data.
- Is screening mammography AI a high risk system under the EU AI Act?
Usually yes, because mammography software is a medical device and falls under the high risk rules for products that need a third party conformity assessment. The vendor carries the conformity work, and the screening program carries the duties of a deployer.
How to cite this page
Blits.ai AI Use Case Library, "AI support for reading screening mammograms in breast cancer screening programs", last verified 12 November 2026, https://www.blits.ai/ai-use-cases/breast-cancer-screening-mammography-reading. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 12 November 2026: First published