Artificial Intelligence in Breast US Diagnosis and Report Generation

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Informatics

Background:

• Breast ultrasound (US) reporting is crucial for accurate diagnosis.
• Standardizing report quality and improving radiologist efficiency are ongoing challenges.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI) system for automated breast US report generation.
• To assess the performance of AI-generated reports compared to human radiologists.

Main Methods:

• A retrospective study utilized over 100,000 breast US cases for AI system training, validation, and testing.
• AI-generated reports were compared against reports from midlevel and junior radiologists (with and without AI assistance) by senior radiologists.
• Primary outcomes included acceptance rates of Breast Imaging Reporting and Data System (BI-RADS) categories and lesion characteristics.

Main Results:

• In external testing, the AI system achieved a BI-RADS acceptance rate comparable to a midlevel radiologist (92.33% vs 95.00%).
• Junior radiologists using the AI system showed improved BI-RADS acceptance rates compared to their unassisted performance (e.g., 90.25% vs 84.75%).

Conclusions:

• The developed AI system shows comparable performance to a midlevel radiologist in generating breast US reports.
• The AI system effectively aids junior radiologists in achieving accurate BI-RADS classification, enhancing diagnostic consistency.