ARTIFICIAL INTELLIGENCE FOR DETECTION AND QUANTIFICATION OF BREAST ARTERIAL CALCIFICATIONS ON MAMMOGRAMS AS A BIOMARKER OF CARDIOVASCULAR DISEASE

Cardiovascular diseases (CVD) represent the leading cause of morbidity and mortality worldwide, imposing a significant healthcare and economic burden. Current risk stratification scores as the main approach for assessing cardiovascular health often underestimate the risk in female population, leading to missed opportunities for primary prevention, early diagnosis, appropriate treatment, and ultimately contributing to the elevated cardiovascular disease burden. This gender-based disparity has prompted the development of innovative sex-specific predictors that could improve women’s CVD risk stratification. In this light, breast arterial calcifications (BAC) have gained traction as one of the most promising women-specific biomarkers: BAC are localized Mönckeberg sclerosis expression involving within the tunica media of breast arteries and detectable as parallel line opacities on about 13% of routine mammograms. They have been shown to be associated with an elevated hazard of cardiovascular adverse events, more accurate than other traditional risk factors in asymptomatic middle-aged women, and also independent of them, indicating the different pathogenesis of BAC from that of atherosclerotic plaques. Considering the widespread diffusion of mammography breast cancer screening programs, systematic BAC assessment could offer a cost-effective cardiovascular risk stratification in women without additional examinations. However, their assessment is a challenging and time-consuming manual task, vulnerable to intra- and inter-observer variability; also, the considerable diversity of BAC’s appearance and the lack of a standard reporting guideline or a reliable quick quantification method have limited their adoption as a robust imaging biomarker in clinical practices. Automated methods using artificial intelligence (AI) and deep learning (DL) algorithms hold promise in addressing the limitations, improving diagnostic reproducibility, reducing radiologists' post-processing workload, and facilitating broader utilization of BAC to improve cardiovascular risk stratification in women and promote awareness of their cardiovascular health, leveraging the large-scale mammographic screening programs. Accordingly, this thesis will present an overview of the current state of knowledge on the automatic BAC assessment using AI-based algorithms (section I), propose a novel DL-based approach for detection and estimation of BAC burden (section II), and subsequently explore the method by a comparative analysis with other established CNN architectures (section III).