AI-enabled detection of breast arterial calcifications in mammography: diagnostic performance and impact of image quality in a large real-world screening population

Abstract Objective To evaluate the diagnostic performance of an AI-based breast arterial calcification (BAC) detection system and to assess the influence of technical and image-quality parameters on automated BAC detection in routine mammography. Materials and methods In this retrospective single-center study, 4692 women who underwent 6061 digital full-field mammographic examinations (24,519 images) between 2011 and 2017 were included. A deep-learning model was applied to standard craniocaudal (CC) and mediolateral oblique (MLO) views to detect BAC and differentiate vascular from non-vascular calcifications. Technical, acquisition-related, and image-quality parameters were automatically extracted. Manual BAC assessment in a randomly selected subset of 1993 images served as the reference standard. Mixed-effects regression models accounting for within-patient clustering were used to identify independent predictors of AI-detected BAC. Results AI detected BAC in 13.9% of women (95% CI: 12.9–14.9), 13.2% of examinations, and 6.7% of images. In the manually reviewed subset, the model showed very high specificity (0.99) and overall accuracy (0.88), with moderate sensitivity (0.35). Most false negatives were associated with subtle or low-intensity vascular calcifications and dense breast tissue. In multivariable mixed-effects models, higher compression force and post-surgical status were independently associated with AI-detected BAC, while breast density was not. Most acquisition-related image-quality parameters showed no significant independent association. Conclusions AI enables scalable automated BAC detection with very high specificity and moderate sensitivity. Detection performance remained largely stable across variations in image-quality parameters, supporting integration of AI-based BAC assessment into opportunistic cardiovascular risk screening workflows. Key Points Question Reliable, scalable identification of breast arterial calcifications (BAC) on routine mammography remains limited, and the influence of technical and image-quality parameters on AI-based detection is insufficiently understood in real-world screening populations. Findings AI-based BAC detection achieved very high specificity and good overall accuracy, but only moderate sensitivity, while most technical and image-quality parameters showed no independent association with detection performance. Relevance statement AI-based BAC detection may enable opportunistic cardiovascular risk assessment from routine mammography without additional imaging or radiation exposure. Its robustness across variations in image quality and acquisition parameters supports integration into clinical workflows, allowing scalable identification of women with definite vascular calcifications who may benefit from targeted cardiovascular evaluation and preventive care.