A Multisite Study of a Breast Density Deep Learning Model for Full-Field Digital Mammography and Synthetic...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology and Diagnostic Imaging
• Machine Learning for Healthcare

Background:

• Accurate breast density assessment is crucial for mammographic interpretation and cancer risk stratification.
• Deep learning (DL) models offer potential for automating breast density classification.
• Multi-site validation is essential for generalizability of AI models in healthcare.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for Breast Imaging Reporting and Data System (BI-RADS) breast density classification.
• To evaluate the model's performance on synthetic mammography (SM) images derived from digital breast tomosynthesis.
• To assess the impact of adaptation methods using limited SM data on model performance across different institutions.

Main Methods:

• A DL model was trained on full-field digital mammographic (FFDM) images from a large cohort (site 1).
• The trained FFDM model was retrospectively evaluated on SM datasets from two independent institutions (site 1 and site 2).
• Adaptation techniques were investigated to optimize performance on SM datasets, considering the effect of dataset size.

Main Results:

• The DL model showed substantial agreement with radiologists for BI-RADS breast density on both FFDM and SM images without adaptation.
• Performance on SM images from site 2 significantly improved after adaptation using only 500 SM images (κw=0.79 vs 0.72, P<.001).
• Adaptation methods demonstrated effectiveness in enhancing model performance on unseen SM data with minimal training examples.

Conclusions:

• A BI-RADS breast density DL model exhibits strong performance on both FFDM and SM imaging data.
• The model's performance on SM images can be significantly improved with limited site-specific data adaptation.
• This study supports the potential of DL models for consistent breast density assessment across different mammographic techniques and institutions.