Semiquantitative score of breast arterial calcifications on mammography (BAC-SS): intra- and inter-reader reproducibility

Abstract

Background: Breast arterial calcifications (BAC), representing Mönckeberg's sclerosis of the tunica media of breast arteries, are an imaging biomarker for cardiovascular risk stratification in the female population.Our aim was to estimate the intra-and inter-reader reproducibility of a semiquantitative score for BAC assessment (BAC-SS).Methods: Consecutive women who underwent screening mammography at our center from January 1 st to January 31 st , 2018 were retrieved and included according to BAC presence. Two readers (R1 and R2) independently applied the BAC-SS to medio-lateral oblique views, obtaining a BAC score by summing: (I) number of calcified vessels (from 0 to n); (II) vessel opacification, i.e., the degree of artery coverage by calcium bright pixels (0 or 1); and (III) length class of calcified vessels (from 0 to 4). R1 repeated the assessment 2 weeks later. Scoring time was recorded. Cohen's κ statistics and Bland-Altman analysis were used.Results: Among 408 women, 57 (14%) had BAC; 114 medio-lateral oblique views were assessed. Median BAC score was 4 [interquartile range (IQR): 3-6] for R1 and 4 (IQR: 2-6) for R2 (P=0.417) while median scoring time was 156 s (IQR: 99-314 s) for R1 and 191 s (IQR: 137-292 s) for R2 (P=0.743). Bland-Altman analysis showed a 77% intra-reader reproducibility [bias: 0.193, coefficient of repeatability (CoR): 0.955] and a 64% inter-reader reproducibility (bias: 0.211, CoR: 1.516). Cohen's κ for BAC presence was 0.968 for intra-reader agreement and 0.937 for inter-reader agreement.Conclusions: Our BAC-SS has a good intra-and inter-reader reproducibility, within acceptable scoring times. A large-scale study is warranted to test its ability to stratify cardiovascular risk in women.

Deep transfer learning for detection of breast arterial calcifications on mammograms: a comparative study

“…Seventy percent of the exams were allocated to the training subset, 15% to the validation subset, and the remaining 15% to the testing subset. Since BAC incidence was found to be positively associated with women’s age [ 19 ], we conducted a specific data splitting strategy by defining four age classes using the BAC population’s age quartiles as thresholds, stratified splitting within each class separately to preserve BAC age distribution, and then consolidating the sub-splits into the overall corresponding subsets [ 28 ]. The training images were further randomly under-sampled reaching a BAC prevalence of 30%, to alleviate the classification bias toward the majority class of our imbalanced dataset [ 29 , 30 ].…”

“…The technique exploited the last convolutional layer’s rich semantic and spatial information to generate a heatmap that highlighted the most informative pixels contributing to the decision-making process of the network [ 45 , 47 ]. To rank these visual explanations in a somewhat quantitative manner, we assessed the Spearman correlation coefficient of the estimated calcified region delineated through thresholding of the heatmaps [ 28 ], against the corresponding manual measurements of BAC lengths previously measured in a subgroup of BAC exams with MLO views [ 19 ].…”

“…With the increasing use of mammography for breast cancer screening, BAC present an opportunity for CVD risk stratification in asymptomatic women [ 10 , 18 ]. Nevertheless, their assessment is a time-consuming manual task, vulnerable to intra- and inter-observer variability [ 19 , 20 ]; also, the considerable diversity of BAC’s appearance and the lack of a standard reporting guideline limited their adoption as a robust imaging biomarker in clinical practice [ 21 , 22 ].…”