PACS-integrated machine learning breast density classifier: clinical validation.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Machine Learning Applications

Background:

• Accurate breast density assessment is crucial for mammography interpretation and breast cancer risk stratification.
• Current methods rely on subjective radiologist interpretation, leading to inter-observer variability.
• Novel AI tools offer potential for objective and consistent breast density classification.

Purpose of the Study:

• To evaluate the performance of a novel machine learning-based tool for predicting Breast Imaging Reporting and Data System (BI-RADS) breast density.
• To assess the accuracy of the AI tool against radiologist consensus in diverse clinical settings.

Main Methods:

• A convolutional neural network was trained on 33,000 mammographic examinations from one academic medical center (Site A).
• The AI tool's performance was validated on separate datasets of 500 studies from Site A and 700 studies from a second academic medical center (Site B).
• Radiologist consensus served as the ground truth for performance evaluation at both sites.

Main Results:

• The AI classifier achieved an accuracy of 84.6% (Site A) and 89.7% (Site B) for four-category BI-RADS density classification.
• For binary classification (dense vs. non-dense), accuracies were 94.4% (Site A) and 97.4% (Site B).
• The AI classifier never disagreed with the consensus reading by more than one density category.

Conclusions:

• The automated breast density tool demonstrates high agreement with expert radiologists' assessments.
• This AI-powered tool shows significant potential for improving the consistency and accuracy of breast density evaluations in mammography.