Capital Region of Denmark mammography screening programme
Verified evidenceExceptions onlyDecision rights movedCreator to judge
AI-triaged single versus double reading of screening mammograms (Transpara)
That every screening mammogram must be independently double-read by two radiologists to keep cancer detection and recall rates acceptable.
Healthcare screening · Capital Region of Denmark; women aged 50–69 in the organised biennial programme
Executive brief
The operating-model shift, in one view.
The Capital Region did not give radiologists a faster viewer. It changed who must look. AI became the first sort, senior radiologists became the only readers on the likely-normal majority, and recall stayed human. The measured result is fewer reads, fewer recalls, and more screen-detected cancers in a live national-style programme. That is the MASAI trial's missing piece: standing practice, not a protocol.
AI value · Radiologist screening-read workload
Verified
33.5% fewer radiologist screening reads
Observational sequential cohorts, not randomised. Interval cancers were not yet available. Screening interval lengthened after AI (median 845 to 993 days), which the authors stratified and argue does not fully explain the detection gain.
Before
Acquires four-view full-field digital mammograms under Danish screening guidelines. → Each reads every examination; the second reader is always a senior radiologist. → Resolves recall disagreements.
After
Assigns an examination malignancy score (1–10) and, for higher-risk exams, lesion marks. → Reads likely-normal examinations knowing they are single-read; may still send a recall to consensus. → Independently double-read remaining examinations and consult AI marks after their own evaluation.
Human boundary
The programme sets the score threshold that drops the second human reader. Radiologists retain recall/no-recall authority, including the right to override an AI-normal tag.
Why it matters
Every screening mammogram does not need two human reads.
How the work changed
Before
How the work ran before the change.
Step 1 of 3
Radiographer
Acquires four-view full-field digital mammograms under Danish screening guidelines.
ControlAcquisition protocol; PACS routing
Step 2 of 3
Two independent breast radiologists
Each reads every examination; the second reader is always a senior radiologist.
Assigns an examination malignancy score (1–10) and, for higher-risk exams, lesion marks.
ControlScore threshold (≤5 until 2022-05-03, then ≤7) set by the programme with continuous quality monitoring
Step 2 of 3
Senior breast radiologist (single-read lane)
Reads likely-normal examinations knowing they are single-read; may still send a recall to consensus.
ControlHuman recall authority retained; 0.8% of single-read exams were recalled against the AI-normal tag
Step 3 of 3
Two radiologists with AI decision support.
Independently double-read remaining examinations and consult AI marks after their own evaluation.
Process model built from the published workflow evidence for Capital Region of Denmark mammography screening programme. Every step, actor, and control appears in full below.Every step, actor, and control
Exception path
Unprocessed exams (1.5%, mostly early infrastructure failures or implants) revert to double reading. Single-read recalls go to consensus or arbitration. Continuous quality control justified raising the threshold from 5 to 7.
Decision authority
The programme sets the score threshold that drops the second human reader. Radiologists retain recall/no-recall authority, including the right to override an AI-normal tag.
Before
#
Actor
Action
Control
01
Radiographer
Acquires four-view full-field digital mammograms under Danish screening guidelines.
Acquisition protocol; PACS routing
02
Two independent breast radiologists
Each reads every examination; the second reader is always a senior radiologist.
Assigns an examination malignancy score (1–10) and, for higher-risk exams, lesion marks.
Score threshold (≤5 until 2022-05-03, then ≤7) set by the programme with continuous quality monitoring
02
Senior breast radiologist (single-read lane)
Reads likely-normal examinations knowing they are single-read; may still send a recall to consensus.
Human recall authority retained; 0.8% of single-read exams were recalled against the AI-normal tag
03
Two radiologists with AI decision support (higher-score lane)
Independently double-read remaining examinations and consult AI marks after their own evaluation.
Readers blinded to each other; instructed not to use AI marks as the first look
Work that left the path
Second independent human read on 66.9% of post-implementation screenings (38,977 of 58,246)
Some downstream diagnostic work associated with the 20.5% recall-rate reduction
Human role before
Two radiologists independently interpreted every screening mammogram; senior readers always supplied the second read.
Human role after
AI decides the reading intensity. Senior radiologists single-read the likely-normal majority. Two humans, with AI marks as a second look, still double-read higher-score exams. Recall authority remains human.
AI role
Deep-learning risk stratification and lesion marking. It does not issue the recall. It routes work and, in the double-read lane, supports detection after the human look.
Outcomes
Radiologist screening-read workload
Verified
Two reads per examination (double reading of the entire cohort)→33.5% fewer reads (38,977 of 116,492 possible reads saved) because 66.9% of exams were single-read
After: 2021-11-18 to 2022-10-17 versus before: 2020-10-01 to 2021-11-17 · 60,751 women before AI; 58,246 women after AI; 19 senior high-volume readers
Observational sequential cohorts, not randomised. Interval cancers were not yet available. Screening interval lengthened after AI (median 845 to 993 days), which the authors stratified and argue does not fully explain the detection gain.
Cancer detection rate
Verified
0.70% (423 of 60,751)→0.82% (480 of 58,246); P = .01
Same sequential cohorts · Same populations
Screen-detected cancers only, with ≥180 days follow-up. DCIS share rose (15.1% to 20.4%), so overdiagnosis remains an open question pending interval-cancer follow-up.
Recall rate
Verified
3.09% (1,875 of 60,751)→2.46% (1,430 of 58,246); 20.5% relative reduction; P < .001
Same sequential cohorts · Same populations
False-positive rate also fell (2.39% to 1.63%) and PPV rose (22.6% to 33.6%). Not a randomised comparison.
What leaders can reuse
Anti-pattern
Citing MASAI or ScreenTrustCAD as if they were BAU operations. Those are trials. This case is the programme that actually changed the reading protocol.
Questions
01Where do we still insist on two humans for work a model can already sort by risk?
02Who owns the threshold that drops a reviewer, and how often is it re-validated?
03What override rate would tell us the single-read lane is unsafe?
Portability conditions
An organised double-reading screening programme with a quality-assurance spine
Willingness to let a model drop a human reader on a defined low-risk slice
Senior readers who can own a single-read lane
Interval-cancer follow-up after the early performance paper
Reputation risk
low-medium: observational design, longer post-COVID interval, and higher DCIS share must stay visible. Do not claim mortality reduction.
Evidence and authority
What the public record supports.
Current · updated
2 independent, 1 peer reviewed; publication outcomes are verified.
Bundle 1.0.0 · reviewed 2026-09-06 · stable ID fbf20734fb447a59
