Generalisable deep learning method for mammographic density prediction across imaging techniques and self-reported...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Breast density is a key breast cancer risk factor and can reduce mammography sensitivity.
• Automated deep learning models can aid risk assessment and identify challenging screening cases.
• Generalizability of these models across imaging techniques and races is not well-established.

Purpose of the Study:

• To develop and evaluate a deep learning model for breast density prediction.
• To assess the model's generalizability across full-field digital mammography (FFDM) and 2D synthetic (2DS) mammography.
• To analyze the model's performance and bias across different racial groups.

Main Methods:

• Utilized a large, diverse dataset of 69,697 mammographic studies from 23,057 participants.
• Developed a deep learning model for four-class BI-RADS breast density classification.
• Evaluated model performance on FFDM and 2DS images, including subgroup analysis by race.

Main Results:

• A model trained on FFDM achieved 80.5% accuracy on FFDM and 79.4% on 2DS.
• Training on both FFDM and 2DS improved accuracy to 82.3% for both techniques.
• The model demonstrated unbiased performance across Black, White, and Asian participants.

Conclusions:

• Deep learning models for breast density prediction are generalizable across imaging techniques and diverse racial groups.
• The developed model shows no substantial performance disparities across racial subgroups.
• This suggests that automated breast density prediction can be applied reliably in diverse clinical settings.