Overview
A new retrospective study published in Radiology: Artificial Intelligence examines how agreement between radiologist interpretation and standalone AI assessment of screening mammography correlates with the risk of developing breast cancer within five years. The findings suggest that concordant positive results between radiologist and AI are associated with markedly higher future cancer risk than either discordant results or concordant negative readings.
Study design and groups
The analysis drew on 82,899 women who underwent screening mammography, with a median follow-up of 5.3 years. Researchers categorized participants into four groups based on the initial readings: (1) concordant-positive radiologist and AI assessments, (2) concordant-negative results, (3) radiologist-positive and AI-negative findings, and (4) radiologist-negative and AI-positive evaluations. The AI model used was Lunit Insight MMG v1.1.7.3 (Lunit).
Key findings
Concordant-positive results (radiologist and AI both positive) carried the highest five-year breast cancer risk and the highest incidence. Specifically, this group had a 5-year cumulative incidence of 3.74%, equating to 37.4 per 1,000 person-years, and a more than fourfold increase in risk compared with the concordant-negative group (5.9 per 1,000 person-years).
Additionally, the study found:
– Radiologist-positive and AI-negative readings were associated with a 15% higher five-year risk than concordant-negative findings.
– Radiologist-negative and AI-positive readings were linked to a 2.3-fold higher five-year risk, with particularly notable risk among women with dense breasts.
Clinical implications
These patterns imply that concordant-positive findings should trigger enhanced surveillance, supplemental imaging, or even risk-reduction considerations, even if the initial diagnostic work-up does not yield a cancer diagnosis. Lead author Eun Young Kim, MD, PhD, notes that a 5-year cancer risk of 3.74% exceeds the high-risk thresholds set by major guidelines, underscoring the potential value of integrating AI findings into follow-up strategies.
Why AI may detect additional risk
For women with radiologist-negative but AI-positive assessments, the increased risk suggests that AI may reveal mammographically occult or pre-malignant changes that escape human readers, especially among those with dense breast tissue. The study observed that the 5-year incidence in this group exceeded rates seen in concordant-negative cases, highlighting AI’s potential role in identifying early or subtle abnormalities.
Limitations and context
Authors acknowledge several constraints. The study is retrospective and single-center, focusing on a relatively young cohort with a mean age of 43.4 years. The site-specific population—Korean women in a dense-breast screening program where density prevalence is high—limits the generalizability of results to broader populations. The authors also emphasize that follow-up and management implications require prospective validation across diverse groups.
Takeaways for practice
Three key takeaways emerge from the study:
1) Concordant AI–radiologist positive findings indicate a notably higher future breast cancer risk, warranting consideration of supplemental imaging or risk-reducing discussions.
2) AI may uncover cancers not visible to radiologists at screening, especially in dense breasts, reinforcing the potential value of AI-assisted screening as part of risk stratification.
3) Clinicians should consider tailored surveillance strategies for patients with concordant-positive results, even when the initial work-up is negative.
Bottom line
The study adds to a growing body of evidence that AI can complement radiologist interpretation in mammography by flagging patients at elevated short-term risk. If validated in broader populations, these insights could shape recommendations for additional imaging and preventive interventions, aiming to reduce interval cancers and improve long-term outcomes.
