A Semiautonomous Deep Learning System to Reduce False Positives in Screening Mammography

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Screening mammography is crucial for early breast cancer detection.
• False-positive results lead to unnecessary patient anxiety, procedures, and costs.

Purpose of the Study:

• To evaluate a semiautonomous artificial intelligence (AI) model for identifying negative screening mammograms.
• To assess the AI model's ability to reduce false-positive examinations in breast cancer screening.

Main Methods:

• A deep learning algorithm was trained on over 123,000 mammograms.
• A retrospective study analyzed three independent datasets (U.S. and U.K.) comprising over 14,000 screening mammography examinations.
• AI performance was compared to human readers, and combined human-AI performance was simulated.

Main Results:

• The AI model demonstrated non-inferiority in cancer detection rates across all datasets.
• Significant reductions were observed in screening examinations requiring radiologist interpretation (up to 41.6%), callbacks (up to 31.1%), and benign biopsies (up to 7.4%).
• These reductions varied across the different U.S. and U.K. datasets.

Conclusions:

• A semiautonomous AI system shows potential for improving breast cancer screening efficiency.
• The AI model can reduce false positives, leading to fewer unnecessary procedures, decreased patient anxiety, and lower medical expenses.
• This technology supports radiologists by streamlining workflow and potentially enhancing screening outcomes.