Spatiotemporal Asymmetries of Longitudinal Screening Mammograms for Breast Cancer Risk Prediction

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Accurate breast cancer risk prediction is crucial for early detection and prevention.
• Existing risk models often do not fully leverage longitudinal mammographic data.
• Identifying subtle changes over time can significantly improve risk assessment.

Purpose of the Study:

• To develop and evaluate a novel deep learning model, STA-Risk, for breast cancer risk prediction.
• The model explicitly captures bilateral and longitudinal asymmetries in sequential mammograms.
• To compare STA-Risk performance against existing risk models.

Main Methods:

• A retrospective case-control study using two datasets (CSAW-CC and an independent cohort).
• Development of STA-Risk, a deep learning model incorporating side encoding, temporal encoding, and asymmetry loss.
• Model evaluation using concordance index (C-index) and time-dependent area under the curve (AUC) via cross-validation and joint training.

Main Results:

• STA-Risk achieved higher C-indexes (0.72-0.73) compared to existing models (0.66-0.70) on individual datasets.
• The model demonstrated consistently higher AUCs for 1-to 5-year risk predictions.
• Ablation studies confirmed the contribution of all three key components of STA-Risk to performance improvement.
• Joint training mitigated domain shifts, achieving cross-cohort C-indexes of 0.75 and 0.67.

Conclusions:

• The STA-Risk deep learning model significantly improves breast cancer risk prediction.
• Utilizing spatiotemporal asymmetries from longitudinal mammograms is key to enhanced prediction.
• STA-Risk offers a promising advancement over current breast cancer risk assessment tools.