An interpretable AI system reduces false-positive MRI diagnoses by stratifying high-risk breast lesions

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Breast cancer diagnosis via MRI faces challenges with high false-positive rates and inter-reader variability, particularly for BI-RADS category 4 lesions.
• This often results in unnecessary biopsies, impacting patient management and healthcare costs.

Purpose of the Study:

• To evaluate the efficacy of the BI-RADS 4 Lesions Analysis System (BL4AS), an AI tool, in improving the diagnostic accuracy of breast cancer detection using dynamic contrast-enhanced MRI.
• To assess BL4AS's impact on reducing false-positive rates and inter-reader variability in classifying BI-RADS 4 lesions.

Main Methods:

• Development and validation of the BL4AS artificial intelligence system, utilizing foundation models and spatiotemporal information from dynamic contrast-enhanced MRI.
• Multicenter study involving 2,803 lesions from 2,686 female patients.
• Comparison of BL4AS performance against radiologists, including assessment of diagnostic accuracy, specificity, false-positive rates, and inter-reader variability.

Main Results:

• BL4AS demonstrated strong performance with areas under the curve ranging from 0.892 to 0.930.
• The AI system significantly outperformed radiologists in specificity (0.889 vs. 0.491).
• BL4AS-assisted interpretation improved diagnostic accuracy, reduced inter-reader variability by 24.5%, and decreased false-positive rates by 27.3%.

Conclusions:

• The BL4AS artificial intelligence system effectively addresses diagnostic challenges in breast MRI for BI-RADS 4 lesions.
• BL4AS offers a practical tool for precision breast cancer management by stratifying lesions into subcategories (4A, 4B, 4C) for refined risk assessment.