Novel deep learning algorithm demonstrates strong performance for detection and quantification of breast arterial calcification and associates with cardiovascular disease.

Abstract

Background: Breast arterial calcification (BAC) on mammography is associated with cardiovascular disease (CVD).1, 2 Given mammography is widely used as part of national breast cancer screening programs, routine reporting of BAC could address current limitations of CVD detection in women. Fully automated artificial intelligence (AI) software can facilitate seamless integration of BAC assessment alongside standard radiological evaluation of mammograms. and advancement in AI imaging software may result in clinically significant, superior BAC detection. Purpose(s): To evaluate the performance of a deep-learning, fully-automated BAC detection algorithm against gold standard human BAC detection, and to evaluate its association with cardiovascular risk and disease. Method(s): A cohort of 285 women (mean age 61) with 2D digital mammography and CVD assessment with either an anatomic or functional test were evaluated. Cardiovascular risk factors were obtained from medical records. Visual BAC (vBAC) was assessed as present/absent by independent reviewers and BAC was scaled by an algorithm incorporating size, length, density and location (aBAC, 0-100). Diagnostic performance of vBAC and aBAC were compared and expressed as area under the receiver operator characteristic curve (AUC). Cardiovascular risk factor and disease prevalence was compared between vBAC and aBAC with CVD status defined by either abnormal anatomical or functional testing. Result(s): The prevalence of vBAC and aBAC were 30.5% and 59.6%, respectively. The algorithm showed excellent discrimination for BAC detection compared with vBAC (AUC 0.92 [95% CI 0.89-0.98]), with sensitivity 94% and negative predictive value 96%. Interpretation took on average 144 seconds/patient with vBAC and 10 seconds/patient with aBAC (p<0.001). We identified 88 discordant cases (vBAC negative, aBAC positive), which were all represented by 'faint' BAC scores (<25). aBAC detected a greater proportion of patients with traditional cardiac risk factors, particularly in women <65 years (Figure 1). CVD was present in 136 (48%), among whom 37.5% were detected by vBAC and 61.7% by aBAC. CVD was present in 42% of the 88 discordant cases. Increasing BAC score was associated with increased relative risk of CVD presence (Figure 2). Conclusion(s): Algorithmic scoring of BAC demonstrates a high performance and is faster to perform when compared to vBAC. Furthermore, it identifies greater proportion of women with higher cardiovascular risk and those with prevalent CVD. Lastly, aBAC detects 'faint' (visually missed) lesions that may have significant clinical implications. (Figure Presented).