Purpose: Mammographically detected breast arterial calcification (BAC) has been reported as a surrogate marker for cardiovascular disease (CVD). However, BAC is under-reported and limited to human visual assessment and lack of a quantitative scoring system. Subtle BAC may be missed, reducing the potential power of this risk marker. Artificial intelligence (AI) may allow rapid and accurate quantification of BAC and streamline assessment on routine mammography to provide an estimation of CVD risk in addition to cancer screening. We evaluate the efficiency and performance of an AI-based BAC algorithm and compare it to human visual assessment.
Materials and Methods: Digital 2-d screening mammograms (n=1567 images) from 306 women (age range 40-70 years, date range 2014-2019) were randomly selected, obtained at a large tertiary hospital. A subset of patients with known CVD (based on either presence of plaque on coronary CT or symptomatic myocardial infarction) were also analysed (n=104). Images were scored by 2 readers by visual analysis. Then, BAC segmentation and quantitation (BradleyTM score) was obtained using cmAngioTM (CureMetrix, La Jolla, CA). Inter-observer agreement was assessed by Cohen’s kappa (k) with confidence interval (CI) of 95%. Area under the curve (AUC) was used to assess model performance.
Results: Excellent agreement for visual BAC between the two readers was demonstrated, k=0.78 [0.71-0.85]. Excellent correlation was seen with the visual human assessment and Bradley score, r=0.84 [0.80-0.87, p<0.001]. There was a high diagnostic performance of cmAngio with AUC=0.87 [0.83-0.92]. A Bradley score of 286 corresponded with an 85% sensitivity and 77% specificity for BAC presence. Using this threshold, BAC prevalence was 68%, compared with 37% with visual BAC. Of these discrepant patients without visual BAC, 36% had CVD and 4% had a myocardial infarction. Average time for quantitative Bradley score calculation was 2.5±0.3 seconds per image.
Conclusion: AI based cmAngio which produces the quantitative Bradley score demonstrates efficient processing of mammograms and excellent performance for detection of BAC in screening mammograms. Discordant cases with negative visual BAC and a positive Bradley score were associated with a high prevalence of CVD, suggesting higher sensitivity of the AI model in significant BAC detection.
Clinical Relevance Statement: AI-based BAC scoring of mammograms may provide a useful pathway for heart disease screening in women, without additional radiation or exam time.
