Large US real-world study finds higher cancer detection among general radiologists
Recall rates, biopsy waiting times and long-term outcomes remain critical measures
Series Overview
- September 4 | Medical Imaging AI Enters National Infrastructure as HIRA Launches AI-Assisted Claims Review
- September 11 | The Medical Imaging AI Platform Race
- September 18 | Lung Cancer Screening AI and National Follow-Up Care
- September 25 | Breast AI Cuts Interval Cancers and Speeds Diagnosis
- October 2 | Stroke AI Connects Image Analysis to Treatment and Patient Recovery
- October 9 | DICOM Standardization and Performance and Security Management for Medical Imaging AI
- October 16 | Linking Medical Imaging AI Reimbursement to Clinical Outcomes
- October 23 | Global Market Strategies for Korean Medical Imaging AI
The clinical value of artificial intelligence in breast imaging cannot be measured by the number of additional cancers detected alone. If AI increases unnecessary recalls and biopsies, or if patients with suspicious findings still face delays in diagnosis and treatment, the benefits remain limited.
Assessment is expanding beyond cancer detection rates to include interval cancers, aggressive tumours, recalls, diagnostic waiting times and long-term outcomes. Interval cancers are diagnosed after a screening examination but before the next scheduled screen. They may have been obscured or missed on the original mammogram, or may have grown rapidly, and are more likely to be advanced at diagnosis.
Recent studies indicate that AI can classify lower-risk examinations, direct suspicious cases to breast radiologists and help hospitals prioritize additional imaging and biopsy. Its effectiveness therefore needs to be assessed across the screening and diagnostic pathway, not only at the image-reading stage.
Randomized trial reports fewer interval and aggressive cancers
The Swedish Mammography Screening with Artificial Intelligence, or MASAI, trial provides some of the strongest evidence on AI-supported breast screening. Researchers assigned more than 100,000 women aged 40 to 80 to either AI-supported screening or standard double reading.
The AI system classified mammograms by risk and marked suspicious findings. Lower-risk examinations were read by one radiologist, while the highest-risk cases received double reading. Radiologists made the final decision.
Results published in 2026 showed 29% more screen-detected cancers in the AI-supported group. Interval cancers were 12% lower and interval cancers with aggressive biological characteristics were 27% lower. Screening detected 81% of all cancers in the AI-supported group, compared with 74% in the control group.
The reduction in interval cancers suggests that more clinically important cancers may have been identified before they became symptomatic. Longer follow-up is still required to determine whether this improves breast cancer mortality and long-term survival.
Kristina Lång, a radiologist at Lund University and principal investigator of the MASAI trial, said AI-supported mammography could reduce radiologists’ workload while improving the early detection of aggressive cancers. She has also stressed the need to use validated tools and continuously monitor their performance, as results may differ across screening systems and patient populations.
MASAI did not test the replacement of radiologists. AI classified risk and highlighted suspicious areas, while clinicians allocated reading resources and made the final decision. The results therefore reflect both algorithm performance and a redesigned reading process.
A large US real-world study also examined a workflow combining AI with specialist review. Published in Radiology in 2026, it analysed 577,742 digital breast tomosynthesis examinations interpreted by 95 radiologists at 109 imaging centres.
The system sent examinations classified as normal by a general radiologist for additional specialist review when AI detected a high level of risk. This review took place before results were released to the patient.
After implementation, the cancer detection rate among general radiologists increased from 3.76 to 4.99 cancers per 1,000 examinations. Their performance was no longer statistically different from that of breast imaging specialists. The proportion of recalled patients subsequently confirmed to have cancer also increased from 3.38% to 3.89%.
The recall rate, however, rose from 9.06% to 10.40%. More recalls can lead to additional imaging, ultrasound examinations, biopsies, patient anxiety and higher costs. Higher cancer detection and the burden of additional testing must therefore be evaluated together.
The study compared performance before and after implementation rather than randomly assigning AI use. Several researchers were affiliated with DeepHealth, which operates the workflow. Changes in staffing or clinical procedures may have influenced the results, while the effect of higher recall rates on biopsies and costs requires further assessment.
Screening gains must translate into faster diagnosis and treatment
Rapid identification of a suspicious mammogram has limited value if additional imaging and biopsy remain delayed. Researchers at the University of California, San Francisco and the University of California, Berkeley applied AI risk assessment to an expedited diagnostic pathway after breast screening.
The researchers analysed more than 4,100 mammograms and classified 525 patients, or about 12.7%, as high risk. These patients received priority interpretation on the day of screening and, when required, proceeded to additional imaging and clinical assessment.
The waiting time for diagnostic assessment fell from several weeks to about one hour. Among patients ultimately diagnosed with cancer, the average wait for biopsy declined from more than two months to less than 10 days. AI identified patients requiring rapid review and helped the health system reorder its workflow; it did not make the diagnosis.
Maggie Chung of UCSF, the study’s first author, said AI risk assessment could identify women requiring expedited care and help ensure that they receive appropriate tests. The research shows that breast AI needs to connect image analysis with additional imaging and biopsy scheduling.
The study was conducted within a single health system and did not assess long-term survival or mortality. It also remains unclear whether prioritizing high-risk patients increased waiting times for others. Changes to the AI threshold could substantially alter demand for same-day imaging and biopsy.
Korea Needs National Evidence on Interval Cancers and Follow-Up Care
South Korea provides national breast cancer screening every two years for women aged 40 and older. Its single-payer health insurance system, national screening data and widespread picture archiving and communication systems, or PACS, provide a foundation for linking AI with quality control and follow-up care. A stronger national framework is still needed to compare mammography quality, recall rates, interval cancers and reader performance across institutions.
The Korean AI-STREAM study involved about 24,000 women aged 40 and older participating in the national breast cancer screening programme. At six university hospitals, breast radiologists interpreted mammograms before and after receiving AI support.
Preliminary results published in 2025 showed that the cancer detection rate increased from 5.01 to 5.70 cancers per 1,000 examinations without a significant increase in recalls. In a separate simulated analysis involving general radiologists, AI support increased cancer detection by 26.4%.
The study used Lunit INSIGHT MMG and received public funding through Korea’s pan-government medical device research and development programme. The authors declared no competing interests. However, it was an observational study rather than a randomized trial, and follow-up analysis linked to the national cancer registry is needed to assess interval cancers and longer-term outcomes.
South Korean company Lunit has entered overseas screening markets with Lunit INSIGHT MMG for two-dimensional mammography and Lunit INSIGHT DBT for digital breast tomosynthesis. Australia’s BreastScreen NSW introduced Lunit INSIGHT MMG in 2024 to support mammography interpretation.
BreastScreen NSW initially used machine reading for selected examinations of women aged 50 to 59 and later expanded it to all first-time screening participants. The New South Wales government said the system had analysed more than 19,000 images and detected more than 130 cancers while reducing callbacks without compromising cancer detection.
Lunit has also been selected as an AI supplier for the breast screening programme in Spain’s Valencian Community. Its products were included among the options available to French public hospitals through Fujifilm, while Qatar’s Primary Health Care Corporation has used the technology in its Screen for Life programme since 2024. Contract scope, active deployment and examination volumes in several markets remain largely based on company disclosures and require confirmation through public operational data and independent studies.
Regulatory clearances and installation numbers alone cannot establish the competitiveness of breast imaging AI companies. Health systems need evidence on cancer numbers and stages, interval cancers, false positives, recalls, biopsy results and time to diagnosis. Commercial deployment figures and company-reported detection rates should be distinguished from independently verified clinical outcomes.
National screening programmes also need clear operating rules. Healthcare providers must determine who reviews cases flagged by AI, when patients are notified and who arranges follow-up care when screening and diagnostic examinations occur at different institutions. Procurement contracts should cover model version control, performance monitoring, error review, notification procedures, incident response and product replacement.
Detecting more cancers is an important result. Clinical value also requires fewer interval and aggressive cancers, control of unnecessary additional testing and a shorter path to diagnosis and treatment. South Korea’s breast AI policy should measure the full post-screening pathway and long-term outcomes alongside accuracy and deployment numbers.
Part 5 will examine JLK’s domestic and overseas real-world evidence and assess how stroke AI connects image analysis with clinician alerts, inter-hospital transfers, thrombectomy and functional recovery at 90 days.
