Description

Breast cancer patients often face overwhelming emotional and practical challenges, from feeling isolated to struggling to find the right information or resources during their treatment and recovery. These barriers can greatly impact both their quality of life and the effectiveness of their care. Among all US women, breast cancer is the second most common cancer and the second most common cause of cancer death. Recent guidelines recommend screening mammography in healthy women starting at 40 years old. The goal of screening mammography is to catch pre-cancerous lesions earlier so there is a high false positive error rate by design to not miss even remotely suspicious lesions on imaging. As evidenced by epidemiological studies, overdiagnosis in breast cancer is now a problem where up to 50% of screened women will have a false positive mammogram interpretation in their lifetime. The psychological impact of the false positivity manifest has significantly increased anxiety among patients that are needlessly recalled for false positive mammogram screening. Moreover, imprecise mammography interpretation may set off a complex cascade of potentially downstream surgical procedures (such as biopsy or mastectomy) for localized disease versus medical interventions such as radiation or chemotherapy for more advanced diseases.

The Health Insurance Portability and Accountability Act (HIPAA) gives patients the right to share their clinical data via informed consent for meaningful use such as research to improve health outcomes. One way of supporting breast cancer patients is by making their mammograms available to them. This allows patients to see and share their images. Some patients may be deterred from obtaining their images from their imaging centers due to cost and as they have no way to view DICOM images. Therefore, this study seeks to provide the mammography images to the participants.

While there are many open-source AI algorithms to improve precision in mammography interpretation, there are widely discrepant outcomes in breast cancer due to a complex and multifactorial disease etiology of different patient populations including social determinants of health. A recent retrospective study found that the integration of AI algorithms performed significantly better than the standard model for predicting breast cancer risk at 0 to 5 years 8. However, the bias in training data used to develop AI has long been recognized as limitations to its widespread application to marginalized populations as recently evidenced by a class action lawsuit against United Healthcare claiming its AI algorithms denied coverage and thus care to black and brown patients at scale 9,10. Moreover, the most cutting-edge algorithms in the current age of generative AI may often make random errors that can be disastrous in a clinical scenario. AI models are not omniscient as there is great variability in humans. AI models may need to be enhanced for different populations (such as different racial groups, ages ranges, or ethnicities) or for different types of breast cancer. Therefore, there is a stark need for accessible patient populations to demonstrate the applicability of robust AI to a diverse US population.

Therefore, using the images that the patients donate, the study aims to build AI models that can better read diagnostic mammograms.