AI Triage Cuts Radiologist Workload by 63% — A Partially Autonomous Workflow Demonstrated by AITIC | Reading Breast Cancer Diagnosis with AI, Vol. 2

Key Takeaways

• In April 2026, Nature Medicine (vol. 32, pp. 1296-1305) published the AITIC trial from Córdoba, Spain. This prospective paired study of 31,301 women evaluated a more aggressive partially autonomous workflow: mammograms classified as low-risk by AI were treated as normal without any radiologist reading.
• In the AI strategy, radiologist workload dropped 63.6% while cancer detection rose 15.2%. However, the recall rate increased 14.8% and failed to meet the prespecified noninferiority margin (−5%).
• Results diverged sharply between 2D digital mammography (DM) and 3D digital breast tomosynthesis (DBT). DM saw cancer detection up 33.7% and recalls up 28.2%; DBT saw essentially no change in either.
• This is Volume 2 of “Reading Breast Cancer Diagnosis with AI.” Building on MASAI from Volume 1, we examine the operational core question — “How far can we trust AI?” — through the numbers and limitations AITIC laid bare.

Introduction — The Next Question AITIC Took On

The MASAI trial from Volume 1 demonstrated that an “AI triage plus reading support” framework could lift screening quality while reducing radiologist workload. It was a meaningful step forward, but it preserved a human safety valve: even AI-flagged low-risk exams were still read by at least one radiologist.

In April 2026, a study tested the next step head-on. Published in Nature Medicine and centered at Reina Sofía University Hospital in Córdoba, Spain, the AITIC trial (Artificial Intelligence in Breast Cancer Screening Program in Córdoba, ClinicalTrials.gov NCT04949776) asked whether the safety valve could be removed for a large fraction of exams.

The core design: 31,301 women had their mammograms read under two strategies in parallel.

• Standard strategy: Independent double reading by two radiologists, no AI input.
• AI strategy: AI analyzed each exam and assigned a risk score from 1 to 10. Scores 1-7 (low risk, ~64% of exams) were treated as normal without any radiologist reading. Scores 8-10 received double reading with AI markings.

Because the same woman’s exam was read under both strategies — a paired design — the AI-driven differential could be measured directly. Equally important, AITIC evaluated both digital mammography (DM) and digital breast tomosynthesis (DBT), the increasingly adopted 3D modality.

This piece reads AITIC across three axes: (1) workload reduction; (2) cancer detection improvement; (3) the recall rate increase. The latter half explores what the DM/DBT divergence implies, what the 24 missed cancers actually were, and where the ethical line sits when AI is given autonomy.

Main

1. Trial Frame — A Paired Read of 31,301 Women

AITIC was conducted between March 15, 2022 and January 11, 2024 within the Andalusian population-based screening program in Córdoba, Spain. Of 33,171 women presenting for routine screening, 31,856 consented, and after standard exclusions, 31,301 women entered the analysis. Median age was 59 (IQR 54-64), with 11.3% on their first round. Breast density distribution by BI-RADS: A 20.6%, B 46.5%, C 27.6%, D 5.3%.

The AI system was Transpara version 1.7 (ScreenPoint Medical), the same family used in MASAI. It is built on deep convolutional neural networks trained on more than 15 million images from over 15 sites across 10 countries. It outputs scores 1-10, with 1-7 classified as low risk, 8-9 intermediate, and 10 elevated.

2. A 63.6% Workload Cut — What Was Removed

The operational result was striking.

Of the 31,301 exams, 19,917 (~64%) received scores 1-7 from AI and were labeled normal without any human reading. Only the remaining 11,384 (36.4%) received double reading with AI annotations. Total radiologist readings dropped from 62,602 in the standard strategy to 22,768 in the AI strategy — a savings of nearly 40,000 reading sessions.

That translates to a substantial reduction in radiologist time. The authors note that DBT exams take roughly twice as long to read as DM, so the savings would be even more material in DBT-heavy programs.

3. Detection Up 15.2% — Cancers AI Caught and Humans Missed

The intuitive worry — “if humans don’t read, won’t more cancers slip through?” — was reversed by the data.

The cancer detection rate (CDR) rose from 6.3 to 7.3 per 1,000, an absolute increase of 1.0 and a relative increase of 15.2% (95% CI 6.6-24.4%, P<0.001). The increase comfortably exceeded the prespecified noninferiority margin and crossed into statistically superior territory.

What changed concretely. Of 252 cancers detected by either strategy:

• Detected only by standard strategy: 24 cancers
• Detected only by AI strategy: 54 cancers
• Detected by both: 174 cancers

The AI strategy picked up 30 more cancers than two-radiologist double reading alone. More importantly, the cancers preferentially detected by the AI strategy had favorable characteristics:

• Invasive carcinomas: +10.1%
• Ductal carcinoma in situ (DCIS): +35.0%
• Grade I (low-grade) invasive: +30.2%
• T1 (small tumors, ≤2 cm): +13.5%
• N0 (lymph node-negative): +15.6%

In other words, AI tilted detection toward smaller, earlier, more treatable cancers. That is a clinically meaningful direction with real implications for patient survival and treatment burden.

4. Recall Up 14.8% — A Failure of Noninferiority That Demands Examination

The recall rate (RR) is where AITIC told a more difficult story. RR rose from 4.8% to 5.5%, a relative increase of 14.8%. That exceeded the prespecified noninferiority margin (within +5% of standard), so noninferiority was not demonstrated.

“Recall” means a woman is called back for additional workup (special views, ultrasound, biopsy if needed) because something on the image looked off. The positive predictive value of those recalls (PPV — fraction that actually turned out to be cancer) was essentially identical at ~13% in both strategies. So AI raised both the cancer-finding rate and the false-positive rate to a similar degree.

Why did recalls rise? The authors point to (a) AI markings on score 8-10 exams making suspicious findings more visible to readers, (b) the Córdoba program’s lack of a “consensus meeting” before recall — when in doubt, recall, (c) a wider AI threshold (scores 8-10 trigger double reading) compared with other trials such as MASAI (only score 10).

The lesson: operational design choices, not the AI itself, can dominate downstream outcomes. This is one of AITIC’s most important contributions.

5. DM vs. DBT — How 2D and 3D Behave Differently

AITIC’s other distinctive contribution is its head-to-head evaluation of digital mammography (DM) and digital breast tomosynthesis (DBT), the latter being increasingly adopted across Europe.

In DM, AI lifted detection by a striking 33.7% but raised recalls by 28.2%. In DBT, both detection and recall were essentially unchanged. Workload reductions were comparable (mid-60s%) for both modalities.

How to read this. The authors argue (a) DBT already operates at high baseline accuracy, leaving less headroom for AI to add detection; (b) Córdoba’s radiologists had ≥5 years of DBT reading experience, narrowing AI’s window; (c) the AI system was trained primarily on DM data, so DBT performance may lag.

The implication is operationally important. In DBT-heavy programs, AI’s principal value is workload reduction, not detection lift. In DM-dominant settings, AI can move detection meaningfully. The case for AI adoption depends on modality and existing reader expertise.

6. The 24 Missed Cancers — What Did “Low Risk” Actually Look Like

Of the 24 cancers missed by the AI strategy, what were they? This is the most consequential information for any “trust AI” debate.

Eleven of the 24 received AI scores 1-7 (low risk) and were therefore labeled normal without any human reading. Of these, 9 were DBT exams and 2 were DM. The other 13 had scores 8-10 and received double reading, but the radiologists on duty did not recall them.

The 11 missed-because-low-risk cancers (Extended Data Table 5) included:

• Histology: 1 invasive lobular carcinoma, 6 invasive ductal carcinomas, 4 ductal carcinomas in situ.
• Imaging features: calcifications, asymmetry, architectural distortion, masses.
• Grade: a mix of I-III.
• In the standard strategy, 23 of 24 missed cancers (96%) were recalled by only one of the two radiologists — i.e., these were “subtle cases” by any reading.

The cancers AI missed were largely cancers that humans found borderline as well. Conversely, two-thirds (34 of 54) of the cancers caught only by the AI strategy had been recalled by both radiologists in standard reading — meaning the standard pathway sent them for workup, but the workup did not produce a final diagnosis. AI’s prompts on these cases plausibly nudged the diagnostic process to completion.

The arithmetic: 24 missed by AI vs. 54 newly caught by AI. Net +30 in favor of AI.

7. AITIC vs. Other Prospective Trials

Several prospective AI-mammography trials have been published over the past two to three years. A comparative view:

• MASAI (Sweden, Lancet 2026): AI plus single/double reading. CDR +20%, recalls flat. Interval cancer −12%.
• ScreenTrustCAD (Sweden, Lancet Digital Health 2023): AI fully replaces the second reader. CDR +4%, workload −50%, recalls +21%.
• PRAIM (Germany, Nature Medicine 2025): Voluntary AI assistance by radiologists. CDR +17.6%.
• AITIC (Spain, Nature Medicine 2026): Low-risk exams not read by humans. CDR +15.2%, workload −63.6%, recalls +14.8%.

AITIC stakes the strongest position on AI autonomy: low-risk-classified exams receive no human reading at all. That maximizes workload reduction and yields a robust detection lift, but it leaves recall control as an open challenge.

8. Where Does AI Authority End — The Ethics

The AITIC authors confront the ethics directly. “Reading most screening mammograms only by AI and automatically labeling them normal” raises real concerns, they acknowledge.

The 11 missed cancers are a fact. But, they note, “in standard reading the radiologists missed 54 cancers — more than AI did.” Among the 11 missed-because-low-risk cancers, 91% were also recalled by only one of two radiologists in the standard strategy — meaning these were difficult by any reading.

That said, the authors agree clearly that AI autonomy must be paired with rigorous quality assurance: (a) automated mammography image quality control; (b) continuous post-market surveillance of AI performance; (c) certification and audit under each jurisdiction’s regulatory framework.

“AI is more accurate than a human” and “AI should be entrusted with the decision” are different propositions. The latter requires social processes — accountability, informed consent, public trust — that no benchmark can deliver alone.

9. Messages for Screening Program Designers

AITIC carries practical messages for governments and municipalities designing programs.

• Operational design dictates outcomes. The presence or absence of consensus meetings, recall thresholds, and scoring cutoffs may matter more than AI accuracy itself.
• Modality reshapes the value equation. In DBT-heavy regions, AI’s main payoff is reading volume; in DM regions, detection lift is the larger prize.
• Vendor diversity matters. AITIC’s results bind to a single AI system. Performance differs across commercial systems; local third-party evaluation is essential.
• Transparency and feedback loops. Continuous post-market monitoring and feedback from missed cases are prerequisites for sustained trust.

10. The Patient’s Perspective — When Can We Be at Ease?

The natural question for the woman receiving the screen is: “If my mammogram was read only by AI, am I really going to be okay?”

The honest summary based on AITIC: at the program level, the chance of cancer being detected is higher under AI than under humans alone. AI is not perfect, however. If you notice symptoms — a lump, nipple discharge, skin changes — see a doctor regardless of your screening result.

Screening is fundamentally a probabilistic risk-management tool for asymptomatic populations. Symptoms warrant a clinical visit independently of any screening outcome. AI does not change that principle.

One operational implication: as recall rates rise under AI, the screening system needs to upgrade pre-screening explanation, recall-time support, and the speed and clarity of result communication. The patient experience must be designed alongside the algorithm.

Conclusion

• AITIC, published in Nature Medicine, prospectively tested a partially autonomous workflow in 31,301 women: AI-flagged low-risk exams were treated as normal without radiologist reading.
• Headline results: workload −63.6%, detection +15.2% (superior), recall +14.8% (not noninferior), PPV equivalent.
• The detection lift was concentrated in early-stage cancers (Grade I, T1, N0) — a clinically favorable direction.
• DM saw large effects on detection and recall; DBT saw essentially none. AI’s value depends on modality and existing expertise.
• Operational choices — consensus meetings, thresholds, recall criteria — shape outcomes more than AI accuracy alone. This is now a system design question more than a technical one.

My Perspective & Outlook

AITIC is one of the first prospective trials to start answering — with data — the global question “how much should we trust AI?” What I find most useful is the trial’s frank disclosure of complexity: superiority and a noninferiority miss coexist; DM and DBT diverge. That complexity is itself the picture of a maturing clinical AI. The next phase, in any healthcare system considering AI integration, will be operational learning: optimizing the trade-off between detection and recall to local contexts. The value structure of AI adoption shifts across geographies — DM-dominant regions stand to gain more on detection lift, DBT-mature programs stand to gain more on workload reduction — and any country’s mix of these modalities will shape its expected return. Equally critical, regardless of jurisdiction, is the mechanism for feeding missed cases back into model training, and the institutional design of patient explanation and consent. The biggest risk in medical AI globally is technology outrunning governance, and AITIC is both a warning and an encouragement on that front. For readers tracking Japan and other Asian markets specifically, the DM/DBT split observed in AITIC matters because Japan still has a substantial DM-screened population alongside expanding DBT adoption in metropolitan centers — meaning Japan may benefit on both axes if implementation is well-designed, but only if Asian-cohort validation work catches up with European data.

Next Up

Volume 3 shifts the lens from detection to prognosis. Presented at the San Antonio Breast Cancer Symposium (SABCS 2025) and reported in AACR Cancer Discovery News (December 2025), Joseph Sparano and colleagues at Mount Sinai built a multimodal AI model for recurrence prediction. Their integrated clinical-imaging-molecular ICM+ model outperformed the current standard 21-gene Oncotype DX test for 15-year distant recurrence — a reanalysis of TAILORx data we will unpack carefully. AI is now stepping into treatment-decision territory; we examine what that means.

Edited by the Morningglorysciences team.

Related Articles

• Reading Breast Cancer Diagnosis with AI Vol. 1: What MASAI Answered
• Reading Breast Cancer Diagnosis with AI Vol. 3: Multimodal AI for Recurrence Prediction (forthcoming)
• DNA’s 98%: A Genomic AI Primer Through DeepMind’s AlphaGenome