Artificial Intelligence in Mammography Screening in Norway (AIMS Norway): Protocol for a randomized controlled trial

This paper outlines the protocol for the AIMS Norway randomized controlled trial, which aims to determine if an AI-stratified mammography reading strategy—where low-risk cases are read by a single radiologist and high-risk cases by two—is non-inferior to standard double reading in terms of screen-detected breast cancer rates.

The Gist

Imagine you are running a massive, highly efficient factory that inspects millions of packages every year. The goal is to find the few packages that contain something dangerous (in this case, breast cancer) while letting the millions of safe packages pass through quickly.

This is exactly what the AIMS Norway study is about. It's a big experiment to see if we can use a super-smart computer assistant (Artificial Intelligence) to help human experts (radiologists) do their job better, faster, and without missing any dangerous packages.

Here is the breakdown of the study using simple analogies:

1. The Problem: The "Needle in a Haystack" is Getting Harder

Currently, Norway has a program where women aged 50–69 get a mammogram (an X-ray of the breast) every two years.

• The Current System: Every single X-ray is looked at by two different human doctors independently. If they disagree, a third doctor helps them decide.
• The Bottleneck: There are too many X-rays (about 99% are perfectly normal) and not enough doctors. It's like having two security guards checking every single person walking through an airport, even though 99% of them are just regular travelers with no weapons. This is slow, expensive, and the doctors are getting tired.

2. The Solution: The "Smart Sorter" Robot

The researchers want to test a new system using an AI tool called Transpara. Think of this AI as a super-fast robot that can look at an X-ray and give it a "Risk Score" from 1 to 10.

• Score 1–7: "This looks very safe." (Low risk)
• Score 8–10: "This looks suspicious." (High risk)

3. The Experiment: The "Tug-of-War"

The study is a Randomized Controlled Trial. This means they are splitting the women into two equal groups to see which method works best. It's like a race between two different ways of sorting mail.

• Group A (The Control Group - The Old Way):

  • Two human doctors look at every single X-ray, just like they always have.
  • They don't know the AI's score while they are looking.
  • Goal: This is the "Gold Standard" to compare against.

• Group B (The Study Group - The New Way):

  • The AI looks at the X-ray first and gives it a score.
  • If the score is 1–7 (Low Risk): Only one human doctor looks at it. The second doctor is saved for other work.
  • If the score is 8–10 (High Risk): Two human doctors look at it, just like the old way.
  • Crucial Rule: The doctors don't see the AI score or the robot's "highlights" while they are making their first decision. They have to trust their own eyes. They only see the AI's opinion later if they need to discuss a case with a colleague.

4. The Big Question: "Is One Doctor Enough?"

The main goal is to prove Non-Inferiority.

• Translation: They aren't trying to prove the new way is better. They just want to prove it is not worse.
• The Analogy: Imagine you are testing a new, faster route to work. You don't need to prove you get there in 10 minutes (the old way takes 20). You just need to prove you still get there in 25 minutes or less, without getting lost.
• The Metric: The most important thing is: Did they find the same number of cancers in both groups? If the new way finds just as many cancers as the old way, but uses fewer doctor-hours, then the new way wins.

5. Why Do This?

• Safety First: They want to make sure that by letting one doctor handle the "safe" cases, they don't accidentally miss a cancer.
• Efficiency: If the AI is right, the doctors can spend their time on the tricky cases (the high-risk ones) instead of staring at thousands of normal X-rays. This saves money and prevents doctor burnout.
• The Future: If this works, it could change how mammograms are done all over the world, making screening programs sustainable even when there is a shortage of doctors.

Summary

Think of this study as a test drive for a self-driving car feature in a busy city.

• The Old Way: Two drivers are always in the car, one steering, one navigating, for every single trip.
• The New Way: A smart computer checks the road. If the road is empty and clear (Low Risk), one driver takes the wheel. If the road is chaotic and dangerous (High Risk), both drivers take the wheel.
• The Goal: To prove that the car arrives at the destination safely (finding all the cancers) even when one driver is resting, ensuring the system can keep running smoothly for years to come.

The study involves about 165,000 women in Norway and will take two years to complete. The results will tell us if we can safely trust AI to help us sort the "safe" from the "suspicious" in breast cancer screening.

Technical Summary

Based on the provided protocol paper, here is a detailed technical summary of the Artificial Intelligence in Mammography Screening in Norway (AIMS Norway) trial.

1. Problem Statement

Breast cancer screening programs face a critical sustainability crisis driven by three converging factors:

• Workforce Shortage: A global deficit of radiologists limits the capacity to interpret increasing screening volumes.
• Inefficiency: Approximately 99% of mammograms are negative, yet the current standard of care in Norway (and Europe) requires independent double reading by two radiologists for every exam, followed by consensus meetings for flagged cases.
• Diagnostic Limitations: Retrospective data indicates that nearly 25% of screen-detected cancers are initially missed by one of the two radiologists, and 20–25% of interval cancers were visible on prior mammograms, highlighting the subjectivity and limitations of human interpretation alone.

The core challenge is how to maintain high diagnostic performance (cancer detection rates) while significantly improving workflow efficiency to address resource constraints.

2. Methodology

The study is a randomized, controlled, parallel-group, non-inferiority, single-blinded trial conducted within the BreastScreen Norway program.

• Study Population: Women aged 50–69 invited to screening in Western, Central, and Northern Norway (approx. 100,000 annual screenings).
• Intervention (AI-Supported Strategy):
  • Mammograms are processed by Transpara® (ScreenPoint Medical), a CE-marked Class IIb AI device using convolutional neural networks.
  • The AI generates a malignancy risk score (1–10).
  • Risk-Stratified Reading Protocol:
    • Scores 1–7 (Low Risk): Interpreted by a single radiologist.
    • Scores 8–10 (Intermediate/High Risk): Interpreted by independent double reading (two radiologists).
  • Blinding: Radiologists are blinded to AI scores and markings during the initial reading to prevent automation bias. AI data is only disclosed during consensus meetings.
• Control Group:
  • Receives the standard of care: Independent double reading by two radiologists for all exams, regardless of AI analysis (which is performed retrospectively for research only).
• Randomization: 1:1 allocation via a computer-generated algorithm based on personal identification numbers immediately after the pre-screening interview.
• Sample Size:
  • Target: 132,184 exams (66,092 per arm).
  • Assumptions: Control cancer detection rate of 6.1 per 1,000; non-inferiority margin of 0.0012 (allowing a drop to 4.9 per 1,000); 80% power; $\alpha$ = 0.025.
  • Recruitment goal: ~165,230 women invited (assuming 80% participation).

3. Key Contributions

• First Large-Scale Prospective RCT: Unlike previous retrospective studies, this is a prospective trial evaluating AI integration in a real-world, population-based screening environment.
• Novel Workflow Integration: It tests a specific risk-stratified triage model where AI determines the number of human readers required, rather than just acting as a second reader for all cases.
• Blinded Design: The protocol ensures internal validity by blinding radiologists to AI outputs during the primary interpretation, isolating the effect of the reading strategy itself.
• Comprehensive Outcomes: Beyond cancer detection, the study tracks operational metrics (reading time, consensus rates) and safety metrics (interval cancers, recall rates).

4. Results

• Status: No results are reported in this document.
• Reason: This is a study protocol published on medRxiv (March 15, 2026) describing the design, methodology, and ethical framework of the trial. The trial is currently ongoing or in the planning/execution phase.
• Planned Analysis: Non-inferiority will be assessed by calculating the confidence interval for the difference in screen-detected cancer rates between the AI-supported group and the control group. Non-inferiority will be concluded if the upper bound of the confidence interval does not exceed the predefined margin.

5. Significance

• Policy Impact: The findings will provide robust evidence to determine if AI can safely replace the second radiologist for low-risk cases, potentially transforming national screening guidelines.
• Sustainability: If successful, this strategy could alleviate radiologist shortages, reduce workload, and lower the cost of screening programs without compromising patient safety.
• Clinical Safety: By setting a strict non-inferiority margin, the trial prioritizes patient safety, ensuring that any efficiency gains do not come at the cost of missing cancers.
• Global Relevance: As a model for integrating AI into public health infrastructure, the results could influence screening protocols worldwide facing similar workforce and volume challenges.

Trial Registration: ClinicalTrials.gov (NCT06032390).
Funding: Norwegian Cancer Society and Regional Health Authorities.