Development and validation of an improved volumetric breast density estimation model using the ResNet technique.

Area of Science:

• Radiology and Medical Imaging
• Machine Learning in Healthcare
• Biomarkers for Cancer Risk Assessment

Background:

• Temporal changes in volumetric breast density (VBD) are potential prognostic biomarkers for breast cancer risk.
• Accurate VBD measurement from archived mammograms is a persistent challenge.
• Previous regression models achieved an R² of 0.868 for VBD estimation.

Purpose of the Study:

• To develop and apply advanced machine learning and deep learning models for VBD estimation from mammograms.
• To compare the performance of Random Forest, XG-Boost, and ResNet models against previous methods.
• To establish a reliable method for retrospective VBD analysis to predict future breast cancer risk.

Main Methods:

• Applied Random Forest, XG-Boost, and Residual Network (ResNet) machine learning models to a dataset of mammography images.
• Utilized imaging parameters (tube voltage, current, exposure time) and patient age as input features.
• Performed five-fold cross-validation to ensure robust model performance assessment using metrics like R², RMSE, and MAE.

Main Results:

• ResNet achieved the highest determination coefficient (R²) of 0.918, outperforming Random Forest (0.895) and XG-Boost (0.907).
• ResNet demonstrated consistently lower error metrics (RMSE, MAE, etc.) across all evaluated folds.
• All developed machine learning models surpassed the performance of the prior multiple regression approach.

Conclusions:

• The ResNet model accurately determines volumetric breast density from historical mammograms, a significant advancement in the field.
• This high-accuracy retrospective VBD analysis capability is crucial for improving future breast cancer risk prediction.
• The findings support the use of advanced AI models for time-series analysis of VBD in breast cancer research.