Closing the Performance Gap between Generalists and Breast Imaging Specialists Using a Nationally Deployed AI...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence in Healthcare

Background:

• Radiologist expertise is crucial for breast cancer screening outcomes.
• Most mammograms are interpreted by general radiologists, not specialists.
• There is a need to improve screening performance across different radiologist expertise levels.

Purpose of the Study:

• To assess the impact of a multistage artificial intelligence (AI) workflow on the clinical performance of general radiologists and breast imaging specialists.
• To compare cancer detection rates (CDR), positive predictive value of recalls (PPV), and recall rates (RR) before and after AI implementation.

Main Methods:

• Prospective study of screening mammogram interpretations from 95 radiologists (60 generalists, 35 specialists) across 109 U.S. facilities.
• Inclusion of digital breast tomosynthesis examinations from women aged 35+ between September 2021 and December 2022.
• Comparison of CDR, PPV, and RR using logistic regression with generalized estimating equations before and after AI workflow integration.

Main Results:

• General radiologists' CDR increased significantly with the AI workflow (3.76 to 4.99 per 1000 exams; P < .001).
• Specialists' CDR remained stable and was comparable to generalists using AI.
• Generalists showed improved PPV of recalls (3.38% to 3.89%; P = .02) with a slight increase in RR.

Conclusions:

• A multistage AI workflow substantially enhanced cancer detection and recall efficiency for general radiologists.
• The AI integration brought generalist performance to a level comparable to breast imaging specialists.
• AI tools show promise in standardizing and improving breast cancer screening quality across diverse radiologist expertise.