New evidence on concordant AI and radiologist readings in mammography
A recent retrospective study published in Radiology: Artificial Intelligence examined how the agreement between radiologist interpretations and standalone AI assessments of screening mammography relates to future breast cancer risk. Using data from 82,899 women with a median follow-up of about five years, researchers compared four groups based on initial screening results: concordant-positive readings by both radiologist and AI, concordant-negative readings by both, radiologist-positive and AI-negative readings, and radiologist-negative and AI-positive readings.
Key findings: higher risk with concordant-positive results
The analysis showed that women with concordant-positive findings—where both the radiologist and the AI system flagged a potential abnormality—faced more than a 4.5-fold higher risk of developing incidental breast cancer within five years compared with the concordant-negative group. The five-year cumulative incidence in this cohort was 37.4 per 1000 person-years, the highest among all groups, while concordant-negative readings had the lowest incidence at 5.9 per 1000 person-years.
Risk patterns across other groups
Among women with radiologist-negative but AI-positive readings, the study found a 2.3-fold higher risk of future breast cancer, a finding particularly notable in women with dense breast tissue. In contrast, those with an initial radiologist-positive claim and AI-negative assessment had a more modest 15 percent higher risk relative to the concordant-negative group.
Clinical implications and call for enhanced surveillance
Lead author Eun Young Kim, M.D., Ph.D., and colleagues emphasize that a 5-year cancer risk of 3.74% in the concordant-positive group exceeds established high-risk thresholds set by major guidelines. They argue this level of risk may warrant supplemental imaging or even risk-reducing strategies, such as chemoprevention or intensified surveillance, even when the initial work-up is negative according to standard radiology.
Considerations for dense breasts and potential AI advantages
The study notes that a substantial portion of the cohort (about 86.8%) had dense breasts, a condition that can obscure mammographic findings. In this context, AI may be identifying mammographically occult cancers or early precursors missed by human readers, particularly among those with dense breast tissue. The results suggest AI-alone or AI-assisted workflows could play a role in risk stratification and follow-up planning, though they also underscore the need for careful integration into clinical practice to avoid unnecessary alarm or over-testing.
Limitations and scope for future research
As a single-center, retrospective study with a relatively young cohort from Korea (mean age 43.4 years) and a high prevalence of dense breasts, the authors caution against broad generalization. They call for multi-center prospective studies across diverse populations to validate findings and to determine how best to incorporate concordant AI–radiologist results into personalized screening and prevention strategies.
Bottom line
The study highlights a surprising association: concordant AI and radiologist-positive readings on screening mammography may identify women at substantially higher risk for future breast cancer within five years. Whether this concordance should trigger enhanced surveillance remains an open question, but the data support considering risk-adapted follow-up and preventive options for these patients.
