Incorporating global-local tissue changes to predict future breast cancer from longitudinal screening mammograms

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Early breast cancer (BC) detection via mammography is vital but population-based screening may not be optimal for all women.
• Current deep learning models for mammogram analysis face challenges in interpretability, temporal change modeling, and precise time-to-event prediction.

Purpose of the Study:

• To develop an interpretable deep learning framework, TA-BreaCR, for personalized breast cancer risk prediction and time-to-onset estimation.
• To integrate multiscale longitudinal tissue changes and model temporal relationships for enhanced clinical utility.

Main Methods:

• Proposed the Tracking-Aware Breast Cancer Risk (TA-BreaCR) model, a novel framework for mammogram analysis.
• Integrated local-to-global multiscale features and explicitly modeled the ordinal relationship of time to BC events.
• Evaluated the model on two independent datasets (In-house and EMBED) for risk classification and time-to-event prediction.

Main Results:

• TA-BreaCR outperformed existing and state-of-the-art methods in both breast cancer risk classification and time-to-event prediction.
• Visualization analysis demonstrated consistent attention to high-risk regions over time, improving model interpretability.
• The model achieved joint prediction of future BC risk and estimated time to onset.

Conclusions:

• TA-BreaCR offers a promising approach for individualized breast cancer screening and prevention strategies.
• The model's ability to integrate temporal dynamics and provide interpretable risk assessments enhances clinical decision-making.
• This framework has the potential to optimize mammography screening protocols and improve patient outcomes.