Fig. 1: Overview design of the study.

A Schematic architecture of the proposed Mammo-AGE model and illustration of the occlusion analysis. The model utilizes four-view mammograms (CC and MLO views of both breasts) as input to predict breast age. An instance-bag transformer, inspired by the global-local transformer framework, integrates self-attention and cross-attention mechanisms to fuse information across views. The model incorporates multi-task learning for breast density prediction and uses a combination of cross-entropy (CE) loss, mean-variance (MVL) loss, and probabilistic ordinal embedding (POE) loss to optimize learning. The model outputs age predictions, with saliency maps highlighting age-relevant regions. Five different backbone-based (ResNet-18, ResNet-50, ConvNeXt-Tiny, EfficientNet-B0, and DenseNet-121) models were ensembled by weighted averaging of predicted ages. The detailed description of the network architecture is provided in Supplementary Fig. 1. B Age distributions of the internal and external datasets. The combined dataset was split patient-wise for five-fold cross-validation. External validation was conducted on additional datasets. C Performance evaluation of the Mammo-AGE model on age prediction. D Association analysis between the breast age gap (predicted age minus chronological age) and breast cancer after bias correction. Specifically, Kaplan–Meier (KM) curves analysis of the inhouse cohort (n = 10,392) demonstrates that individuals with a higher breast age gap have a significantly (two-sided log-rank test) lower healthy probability over time compared to those with a lower breast age gap. Shaded areas represent the 95% confidence intervals. E Evaluation of the Mammo-AGE model on downstream clinical tasks, including breast cancer diagnosis and risk prediction.