Bifurcated Networks for Breast Density & Cancer Risk: A Technical Framework.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Biomarker Discovery and Validation
• Machine Learning for Healthcare

Background:

• Breast density and cancer risk are crucial imaging biomarkers, but assessment faces challenges with inter-reader variability and reproducibility.
• Current methods for evaluating these biomarkers lack consistent accuracy, impacting clinical decision-making.

Purpose of the Study:

• To evaluate the feasibility of a bifurcated neural network for simultaneously predicting breast density and a composite cancer risk index.
• To establish a methodological foundation for integrating AI-driven multi-biomarker assessment into contrast-enhanced mammography (CEM) workflows.

Main Methods:

• A simulated patient cohort (n=1000) was used to model breast density (Densitanum) and cancer risk (RiskEnum).
• A multi-output bifurcated neural network was developed and compared against linear regression and a single-output MLP.
• Performance was evaluated using R², MSE, and MAE, with trend analysis for physiological consistency.

Main Results:

• Linear regression demonstrated limited predictive power (R² ≈ 0.144).
• A single-output MLP improved cancer risk prediction (R² = 0.436).
• The bifurcated neural network achieved low MAE (2.624 for Densitanum, 3.731 for RiskEnum), proving effective for simultaneous multi-target prediction and reproducing known patterns like age-related density changes.

Conclusions:

• Bifurcated neural networks can effectively model correlated breast imaging biomarkers with high internal consistency.
• This AI architecture offers a reproducible platform for real-world CEM data testing, supporting AI-enhanced risk stratification and personalized screening.