You get the best of both worlds? Integrating deep learning and traditional machine learning for breast cancer risk prediction

Breast Cancer is the most commonly diagnosed cancer worldwide. While screening mammography diminishes the burden of this disease, it has some flaws related to the presence of false negatives. Adapting screening to each woman's needs could help overcome these challenges. While traditional risk models are valuable tools, we propose an image-based approach. Since artificial intelligence has proven effective in aiding the diagnosis of breast cancer, we aim to translate this technology to risk prediction. A 3-year risk prediction model, with a case-control age-matched approach, was developed based on the analysis of "prior" healthy mammograms. Two classes were defined - "risk" and "control" - based on the assessment done on the most recent examination: if the case was diagnosed with cancer, the prior mammogram was assigned to the "risk" class; otherwise, the prior mammogram was allocated to the "normal" class. In total, we found 3720 available controls and 1471 risk cases. Every mammogram used in this study was taken 3 years before the assessment used for class definition. Risk prediction was aimed through three methodologies: traditional machine learning, deep learning, and a combination of both. The AUCs obtained on the test set were 0.68 for the traditional machine learning, and 0.76 for the other two. No statistically significant differences were found among methods. Our findings suggest that the use of image-based deep learning methods holds promise on the field of Breast Cancer risk prediction, with further validation being needed to confirm their clinical applicability.