Abstract
Purpose: Mammographically detected breast arterial calcification (BAC) has been reported as a surrogate marker for coronary vascular disease (CVD). BAC is often not reported, reducing the power of this risk marker. Here we assess a deep-learning algorithm for BAC quantification, and in cases where the software identified BAC but the readers did not, assess correlation with clinical CVD. Methods: Screening digital mammograms (n=285, 1232 images) were selected from a tertiary Australian hospital from a sample of patients who underwent CVD screening (137 confirmed CVD). Two readers identified binary presence/absence of BAC. A deep learning software (cmAngio™ , CureMetrix, USA) quantified BAC score (scale 0-100). Reader-software discordance was adjudicated by a third observer. Area under the curve (AUC) was used to assess performance with 95% confidence intervals (CI) presented. Interobserver agreement was assessed by Cohen’s kappa (k). Results: Interobserver agreement for visual BAC was good (k=0.78, CI:0.71-0.85, p<0.01). BAC scores ranged from 0-100, with 12% scoring 0. The algorithm had high performance, AUC 0.92 (CI:0.88-0.96, p<0.001) vs initial reader assessment, increasing to AUC 0.98 (CI:0.97-0.99, p<0.001) after adjudicated assessment. At an empirical threshold of BAC score ≥5 to denote BAC presence, BAC prevalence was 38% (109/285) by software, and 31% (87/285) visually. In cases where only the software identified BAC (n=35), BAC scores were all <25, with 37% (n=13) having confirmed CVD and 9% (n=3) myocardial infarction. Conclusion: cmAngio demonstrates excellent performance for detecting BAC on screening mammograms. Faint BAC may be missed by human readers, is better identified by the software, and in these cases there is a high rate of CVD suggesting use in a screening population may improve risk prediction for CVD.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call