Could artificial intelligence (AI) algorithms improve breast cancer prediction models?

Takeaway: A team of German researchers created a new artificial intelligence (AI) algorithm to improve the sensitivity and specificity of breast cancer mammography screening and help reduce the workload of overwhelmed radiologists. 

October is Breast Cancer Awareness Month, a time to honor the memories of those who’ve lost their battles with breast cancer by encouraging and supporting current patients and survivors. Many organizations, including the Susan G. Komen Foundation, have worked tirelessly for several decades to bring awareness to the importance of early detection and the accessibility of treatment options. Along with the expansion of this key messaging, research efforts have also increased, creating hope around exciting new technologies. Making headlines among these technologies: artificial intelligence.  

The National Breast Cancer Foundation estimates within the United States alone, one in every eight women will develop breast cancer in her lifetime. In 2022, the Foundation predicts 287,500 new cases will be diagnosed, 65% of which will be at a localized stage, meaning the cancer has not yet spread or metastasized outside of the breast, positioning the patient well for a five-year survival rate. Yet, this year, approximately 43,550 women will die from breast cancer—leaving researchers pushing forward to uncover new ways to improve early diagnosis and treatment of this disease. 

Artificial intelligence identified as promising support system for clinicians  

In recent years, the combination of improved medical imaging and AI advances have driven interest around understanding how and where this new technology could be used to more accurately interpret digital health data to improve health outcomes, inclusive of breast cancer patients, which are among the most common cancer diagnoses for women in the US. 

One study highlights how breast cancer screening technologies were among the first to apply an AI-approach with deep neural networks (DNN) imaging, providing the framework for future AI studies. Within this research, the authors also completed a comprehensive analysis on how deep AI algorithms can be used as a successful tool to monitor and categorize legions, including those related to breast cancer, when large data sets can be assessed. However, the expanded use of this technology also created a new challenge for medical teams—increased workloads for radiologists.

Team explores new breast cancer AI algorithm 

Recently, a team of German researchers sought to apply prior findings to determine if breast cancer AI algorithms could serve as a useful, automated tool to reduce the workload of overwhelmed radiologists without reducing the sensitivity of mammography screening. Prior AI models introduced AI-standalone approaches or included combined radiologist and AI approaches that required the radiologist to review all final outputs. Collectively, these methods did not reduce the workload and in some cases were known to deliver high rates of false positives.

In compliance with standard best practices for developing machine learning models, the team created a new DNN-based cancer-classification algorithm for the classification of breast cancer screening based on a negative prediction triaging system. The team compared and contrasted three approaches:

● AI standalone approach—the AI algorithm took over all aspects of the decision-making process and “very confident” assessments are completed automatically. 

● Radiologist approach—an unaided individual interpretation of the imaging was the sole decision-making element. 

● Decision-referral approach—combines both algorithm predictions with individual radiologist interpretation of the image results. In these instances, machine learning automates removal of “confident” positives, sending only “less confident” assessments to the radiologist for confirmation.

In all data subsets, the AI-standalone approach could not maintain the same levels of sensitivity when compared to the radiological reviews, underscoring the findings from previous studies. When comparing AI to decision-referral, performance differed based on group. However, overall, the AI models’ ability to categorize a “triage” group for subsequent review improved the accuracy of the second radiologist read, helping to identify cancers that might otherwise have been missed and increasing the radiologist’s ability to detect malignant in situ and invasive legions while at the same time decreasing strain on radiological teams.

What’s next for AI breast cancer screening? 

While additional algorithms need to be tested, the hybrid AI DNN-based cancer classification algorithm paired with the decision-referral triage approach appears promising—offering radiologists and oncologists a workload management system inclusive of a markedly improved safety net for breast cancer identification, monitoring, and prognosis decisions.