Artificial Intelligence as triage in breast screening: simulated impact on workload and detection

Background

Artificial intelligence(AI) is increasingly proposed as a triage tool in mammography screening to reduce double reading workload. However, its impact on cancer detection and screening performance remains uncertain. Within a project funded by the Italian Ministry of Health on the evaluation of AI in mammography screening, this study evaluated, through simulation, the impact of AI-based triage in a population-based screening programme.

Methods

A retrospective simulation was conducted on the 2025 cohort of the largest Italian Local Health Unit. AI scores (Lunit INSIGHT) were analysed using a predefined threshold(2.9).Exams below threshold were considered for single reading. Outcomes included workload reduction and potential missed cancers compared with standard double reading.

Results

Among 38,075 mammograms, 4,542(11.9%) had AI scores below threshold. Among those,96% were also negative at double reading, while 201 women were referred to second-level assessment, resulting in 18 cancers. Cancers not identified by AI were predominantly associated with microcalcifications(61%), followed by distortions(22%),opacities(9%) and asymmetries(8%). In the simulated scenario, 4,542 exams would have undergone single reading. Of the 201 recalls,46 cases(concordant positive or symptomatic) would certainly have been recalled, identifying 17 cancers and 2 lesions of uncertain malignant potential. The outcome of discordant cases cannot be predicted; therefore, up to one additional cancer could potentially be missed. Workload reduction was estimated at approximately 5% of annual reading volume.

Conclusions

AI-based triage may reduce workload and support sustainability of screening programmes. However, current performance shows limitations in detecting cancers presenting as microcalcifications, with possible implications for diagnostic sensitivity. These findings highlight the need for cautious implementation and careful integration with radiologist interpretation and quality assurance processes.

Key messages

• AI triage can reduce screening workload but introduces uncertainty in cancer detection.

• Careful AI evaluation is needed to ensure safe and sustainable integration in breast screening workflow.