Deep Learning Model Capturing Spatiotemporal Asymmetries on Sequential Mammograms Improves Breast Cancer Risk Prediction

A deep learning model (STA-Risk) capturing bilateral and longitudinal asymmetries on sequential mammograms achieved C-indexes of 0.72 and 0.73 in two cohorts, outperforming other risk models (0.66-0.72).

• Ablation studies confirmed that each model component—side encoding, temporal encoding, and asymmetry loss—contributed to improved performance.
• Domain shifts between datasets were observed; joint training with a mixed dataset mitigated these effects, yielding cross-cohort test C-indexes of 0.75 and 0.67.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.