Longmambattn: A Novel Architecture for Enhanced Breast Cancer Risk Prediction Using Variable-Length Longitudinal Mammograms
Zhengbo Zhou, Dooman Arefan, Margarita L. Zuley, Jules H. Sumkin, Shandong Wu · 2025
Recent advancements in deep learning have shown significant potential in predicting breast cancer risk from mammograms. While leveraging longitudinal changes in mammograms is crucial for accurate risk prediction, existing models often face limitations in capturing these temporal relationships effectively, and they tend to be computationally intensive. To address these challenges, we introduce LongMambAttn, a novel architecture designed to handle variable-length, multitemporal mammograms, thereby enhancing breast cancer risk prediction accuracy. LongMambAttn models the temporal changes of mammograms taken over periods ranging from 1 to 8 years. In an internal case-control dataset of 590 patients, our model demonstrates superior performance in predicting future cancer incidence, surpassing methods that rely only on the most recent prior mammogram as well as other models that incorporate mammograms taken at varying time intervals. Our results show that incorporating longitudinal mammograms via LongMambAttn leads to improved predictive accuracy for breast cancer risk.