News

AI Model Shows Promise in Reducing False Positives in Breast Cancer Screening

Researchers develop system that explains its decisions to radiologists, potentially cutting unnecessary procedures

• By John K. Waters
• 07/16/2025

Researchers at Microsoft's AI for Good Lab and the University of Washington have developed an artificial intelligence model that could reduce false positives in breast cancer screening while providing clearer explanations for its decisions.

The model, called FCDD (Fully Convolutional Data Description), uses anomaly detection to identify potential cancers in breast MRI scans. Unlike traditional AI systems that learn to recognize cancer patterns, this approach learns what normal breast tissue looks like and flags deviations from that baseline.

The research, published in the journal Radiology ("Cancer Detection in Breast MRI Screening via Explainable AI Anomaly Detection"), tested the model on more than 9,700 breast MRI examinations across multiple datasets, including realistic screening populations where cancer rates were as low as 1.85%.

Addressing a persistent challenge
Breast cancer screening faces a significant trade-off between sensitivity and specificity. While MRI is highly sensitive for detecting cancer, particularly in women with dense breast tissue, it generates high rates of false positives that lead to unnecessary anxiety, additional imaging, and biopsies.

Current screening methods result in less than 5% of women undergoing breast MRI ultimately being diagnosed with cancer, according to the study. The remainder face follow-up procedures that prove unnecessary.

"The problem is especially acute for the nearly 50% of women who have dense breast tissue—a condition that not only increases the risk of breast cancer but also makes it harder to detect abnormalities through traditional imaging methods," the researchers noted.

Key findings
In testing, the FCDD model demonstrated several advantages over conventional AI approaches:

• Improved accuracy in low-prevalence settings: The model achieved double the positive predictive value of standard models and reduced false alarms by more than 25% in screening-like environments.
• Enhanced explainability: The system generates visual heatmaps showing suspected tumor locations, which matched expert radiologist annotations with 92% accuracy at the pixel level.
• Cross-institutional performance: The model maintained effectiveness when tested on external datasets without additional training, suggesting potential for broader clinical adoption.

Clinical implications
The model's ability to maintain high sensitivity (95-97%) while improving specificity could help reduce unnecessary callbacks and biopsies, potentially easing both emotional and financial burdens for patients.

Savannah Partridge, Professor of Radiology at the University of Washington and senior author of the study, said the model's efficiency with minimal image data could help shorten both scan times and interpretation times.

"We are very optimistic about the potential of this new AI model, not only for its increased accuracy over other models in identifying cancerous regions but its ability to do so using only minimal image data from each exam," Partridge said.

Next steps
The researchers acknowledge that prospective testing in larger, diverse clinical populations will be necessary before the technology can be widely implemented. The team has made their code and methodology available to the research community to encourage further development.

The work represents part of a broader effort to develop AI tools that complement rather than replace radiologists, providing what the researchers describe as "sharper tools and clearer signals" for evaluating difficult cases.

Breast cancer affects approximately one in eight women in the United States during their lifetime. Early detection through screening has been shown to reduce mortality by 20-40% for women aged 50-69, making improvements in screening accuracy a significant public health priority.