Artificial Intelligence Model Created by Northwestern Medicine, Google Health Improves Breast Cancer Detection


A collaborative team of engineers and clinicians trained a breast cancer artificial intelligence tool to read mammograms more accurately than a panel of expert radiologists.


Breast cancer is one of the leading causes of death in women, with one in eight facing a diagnosis in her lifetime.1 While mammograms can help identify cancer early, they are not a perfect science. In fact, half of all women experience a false positive over the course of a decade, leading to invasive biopsies and unnecessary stress. Mammograms can also miss up to 20% of breast cancers present at the time of screening, resulting in delayed diagnosis and care2.


Understanding the significant costs of inaccurate readings — coupled with the implications of the increased time it takes to get results from radiologists due to staff shortages — a large, interdisciplinary team at Northwestern Medicine partnered with Google Health to develop a better way to detect breast cancer in mammograms using artificial intelligence (AI).

Training the AI Model

When Google Health engineers realized they had the necessary AI framework to potentially detect cancer, they looked to Northwestern Medicine and Northwestern University Feinberg School of Medicine to develop and deploy the AI in a clinical setting. Using high volumes of custom mammogram data, scientists spent more than a year developing an AI model that could identify breast cancer rapidly. Under an Institutional Review Boards-approved clinical trial at Northwestern Medicine, in less than two minutes — before the patient leaves the room following her mammogram — the model can deidentify the image, read it using leading-edge AI, and send it back to the radiologist to confirm the result.
If the result is positive, patients can move into diagnostic imaging that same day, speeding up treatment time and decreasing the stress they otherwise would face in waiting up to 30 days for results. Additionally, because in the U.S. mammograms are typically only read by one radiologist, the AI model could someday offer an important secondary screening touchpoint to find cancer that is invisible to the human eye.

Demonstrating Success

Once the AI model showed it could recognize cancer, the team needed to gather more data to analyze its effectiveness. In an international retrospective study, the model was tested against new sets of mammograms from screening centers in the U.K. and the U.S. Compared to predictions made in clinical practice and from six radiologists in an independent study, the model’s predictions resulted in:
  • An absolute reduction of 9.4% in false negatives (U.S.)
  • An absolute reduction of 5.7% in false positives (U.S.)
  • Evidence of the tool’s ability to generalize from what it learned on the U.K. sites to the U.S. sites, showing its applicability to different populations

Moving Toward Clinical Use

Following the success of the retrospective study, a prospective trial was launched in 2021 to study the model in clinical use. Eligible patients undergoing mammograms at Northwestern Medicine can opt to have the model read their results in conjunction with radiology. While the longer-term goal is Food and Drug Administration approval for widespread use, the AI model is already working to improve breast cancer detection rates at Northwestern Medicine, a significant step toward achieving better outcomes for breast cancer patients in the U.S.



1  American Cancer Society®. Breast Cancer Facts & Figures 2019-2020. American Cancer Society; 2019.

2  Limitations of mammograms. American Cancer Society Web site. Updated October 3, 2019. Accessed August 24, 2021.