Detection and quantification of breast arterial calcifications on mammograms: a deep learning approach

• DOI: 10.1007/s00330-023-09668-z
• Publication Date: 2023-05-11T01:21:44Z
• Authors: Nazanin Mobini; Marina Codari; Francesca Riva; Maria Giovanna Ienco; Davide Capra; Andrea Cozzi; Serena Carriero; Diana Spinelli; Rubina Manuela Tr...; Giuseppe Baselli; Francesco Sardanelli
• Keywords: Interquartile range; Neuroradiology

Breast arterial calcifications (BAC) are a sex-specific cardiovascular disease biomarker that might improve risk stratification in women. We implemented deep convolutional neural network for automatic BAC detection and quantification.In this retrospective study, four readers labelled four-view mammograms as positive (BAC+) or negative (BAC-) at image level. Starting from pretrained VGG16 model, we trained to discriminate BAC+ BAC- mammograms. Accuracy, F1 score, area under the receiver operating characteristic curve (AUC-ROC) were used assess diagnostic performance. Predictions of calcified areas generated using generalized gradient-weighted class activation mapping (Grad-CAM++) method, their correlation with manual measurement length subset cases was assessed Spearman ρ.A total 1493 women (198 BAC+) median age 59 years (interquartile range 52-68) included partitioned training set 410 (1640 views, 398 BAC+), validation 222 (888 89 test 229 (916 94 BAC+). The accuracy, AUC-ROC 0.94, 0.86, 0.98 set; 0.96, 0.74, 0.96 0.97, 0.80, 0.95 set, respectively. In 112 analyzed Grad-CAM++ predictions displayed strong measured (ρ = 0.88, p < 0.001).Our model showed promising performances quantification burden, showing measurements.Integrating our clinical practice could reporting without increasing workload, facilitating large-scale studies on impact risk, raising awareness women's health, leveraging mammographic screening.• (CNN) quantification. • Our CNN had an receiving operator composed 916 which . Furthermore, measurements 0.88) views.