How Concordant AI and Radiologist Readings Elevate Risk Signals
A new retrospective study published in Radiology: Artificial Intelligence has found that when a screening mammogram yields concordant positive results from both the radiologist and an AI tool, the five-year risk of developing breast cancer is markedly higher than when both interpretations are negative. The study analyzed a large cohort and tracked cancer outcomes over a median period of 5.3 years, offering important insights into how AI-assisted readings may refine risk stratification in population-based screening.
In this study, researchers evaluated 82,899 women who underwent screening mammography. They categorized participants into four groups: (1) concordant-positive radiologist and standalone AI results, (2) concordant-negative results, (3) radiologist-positive but AI-negative findings, and (4) radiologist-negative but AI-positive assessments. The investigators then compared five-year cumulative breast cancer incidence across these categories.
Key Findings: Risk Magnitudes Across the Four Groups
The concordant-positive group—where both the radiologist and the AI tool (Lunit Insight MMG) flagged positives—had the highest five-year cancer incidence, at 37.4 per 1,000 person-years. This group showed more than a 4.5-fold higher risk of future incidental breast cancer than the concordant-negative group, which had the lowest incidence at 5.9 per 1,000 person-years.
For women with radiologist-positive and AI-negative readings, the study noted a 15% higher five-year risk compared with the concordant-negative group. Those with radiologist-negative but AI-positive results faced a 2.3-fold higher risk, highlighting that AI may detect early or subtle changes not seen by human readers, especially in dense breasts.
Overall, the five-year cumulative cancer incidence among the concordant-positive group reached 3.74%, already surpassing high-risk thresholds set by major guidelines such as the National Comprehensive Cancer Network and the U.S. Preventive Services Task Force. This finding underscores the potential need for enhanced surveillance or preventive options for women with concordant-positive results, even when initial evaluations are negative for immediate clinical concern.
Implications for Clinical Practice and Surveillance
Lead author Eun Young Kim, MD, PhD, and colleagues suggest that concordant-positive readings may identify women at substantially elevated risk who could benefit from supplemental imaging, closer watch, or risk-reducing strategies such as chemoprevention. The authors emphasize that current diagnostic work-ups may not fully reflect an individual’s future cancer risk in cases of concordant-positive interpretations.
Among the new cancers detected during follow-up, the majority were invasive and localized. Notably, women in the concordant-positive group tended to receive their diagnosis sooner after the baseline screening, reinforcing the principle that concordant-positive readings may indicate biologically meaningful, albeit early-stage, disease processes.
Context, Limitations, and Future Directions
The study’s authors caution that the cohort consisted entirely of Korean women with a high prevalence of dense breasts and a relatively young average age (mean about 43 years). While the results illuminate important dynamics between radiologist and AI interpretations, generalizing to broader populations requires further multi-center research with diverse demographics and screening technologies.
These findings contribute to the evolving understanding of how AI can augment radiology by not only flagging suspicious lesions but also shaping risk-based management pathways. Clinicians should balance AI-assisted risk signals with clinical judgment, patient preferences, and individual risk profiles when considering intensified follow-up or preventive strategies.
Bottom Line
Concordant positive readings from radiologists and AI on screening mammography may indicate a higher near-term cancer risk than previously recognized. As AI tools become more integrated into screening programs, risk stratification could become more nuanced, potentially guiding personalized surveillance and preventive decisions.
