Deep Learning-Based Cardiovascular Risk Prediction Using Routine Mammograms

Abstract

Aims

This thesis aims to develop an innovative deep learning-based framework for the automated detection and grading of breast arterial calcification (BAC) using routine mammograms. The framework is designed to enhance the assessment of BAC severity and its potential role as an independent predictor of cardiovascular disease (CVD) risk. By leveraging data from a woman’s initial mammographic screening, the model categorizes BAC severity into low, moderate, or high-risk groups for adverse cardiovascular outcomes. This personalized approach seeks to provide a more accurate assessment of cardiovascular risk, thereby enhancing the utility of routine mammographic screenings.

Methods

A U-Net-based deep learning model was developed to segment BAC regions, utilizing two datasets: 1,270 women for training the segmentation model and 9,648 women for assessing cardiovascular outcomes. The initial training phase employed a fully supervised learning approach, refined using a semi-supervised technique with progressive pseudo-labelling. BAC severity was categorized using a four-level grading system absent, mild, moderate, and severe based on tertile thresholds of BAC area or intensity. Regression models assessed associations between BAC and age, diabetes, hypertension, and smoking. Cox Proportional Hazards models calculated adjusted hazard ratios (HR) for BAC grades and cardiovascular outcomes.

Results

The model achieved strong performance (Jaccard 0.602, precision 0.77, F1 0.76, accuracy 0.99). BAC prevalence was 26.96%, increasing with age. Age, diabetes, and hypertension were significant predictors; smoking showed an inverse association. Severe BAC was associated with an adjusted HR of 1.65 (95% CI: 1.30 – 2.10).

Conclusion

This research presents a novel deep learning-based BAC framework, supporting its integration into routine mammography as an additional valuable tool for personalized cardiovascular risk stratification in women.