Breast Cancer Diagnosis and Risk Prediction in Multi-Incidence Mammograms: Leveraging UK and Brazilian Data and Expertise

This collaborative project aims to develop novel AI (Artificial Intelligence) methods for breast cancer diagnosis and risk prediction using mammograms, by leveraging the combined expertise and diverse mammogram datasets from the UK and Brazil. Breast cancer remains a significant global health burden, with an estimated 2.3 million new cases and 685,000 deaths in 2020 alone. Early detection is crucial for improving patient outcomes. However, AI-based cancer detection and risk prediction models can be biased if the training sample is not representative of the entire population. The UK and Brazil have distinct demographic characteristics. By leveraging mammogram data from both countries, this project aims to reduce bias and improve the generalizability of the diagnosis and risk prediction models, contributing to more equitable and effective breast cancer screening worldwide. The UK-based team has developed a deep learning model called BREST (Breast Risk Evaluation from Screening Test) for three-year risk assessment using mammograms. The Brazilian team has proposed a breast cancer diagnosing algorithm called "Patch to Multi-View" (P2MV) that simultaneously uses the two standard views of the breast to significantly increase the accuracy, compared to other strategies that also use the two views. We will test whether the breast cancer risk prediction provided by BREST can be improved using multiple mammographic views via the P2MV algorithm. When a radiologist finds a suspicious lesion, he/she may request complementary mammogram views, such as cone view, cleavage view, compression view, etc., to better evaluate the detected abnormality. We propose to investigate whether using these complementary views can help to improve breast cancer detection and risk prediction. P2MV algorithm is well-suited for this task, as it can extract information from multiple views. A recent study analyzed 134,870 breast cancer deaths in Brazil in women aged 20 to 69, from 1996 to 2013. Unfortunately, there was a temporal trend of increased breast cancer mortality in young women aged 20 to 49. Therefore, early diagnosis of cancer in young women becomes increasingly important. However, young women have dense breasts, making it difficult to diagnose cancer using X-rays. We want to determine how well AI models perform in detecting and predicting breast cancer risk in young women. This would allow us to propose the best strategies for early cancer diagnosis for this age group.