Using deep learning to estimate breast age from mammogram images

Chronological age is a well-established risk factor for future breast cancer. In addition to accurately predicting breast age, the ability of risk stratification by predicted breast age is further validated. Using data from the inhouse and EMBED datasets, we applied Kaplan-Meier survival analysis and Cox proportional hazards regression models to assess breast cancer risk. The inhouse dataset included a follow-up period of 10 years, and the external EMBED dataset comprised a follow-up period of 5 years. The results from the Kaplan-Meier curves (Supplementary Fig. ) and the Cox proportional hazards regression models (Supplementary Table ) demonstrate that the risk stratification capability of predicted breast age is comparable to that of chronological age. The hazard ratios per year of age are similar in both datasets, with the Inhouse dataset showing HRs of 1.032 for chronological age and 1.030 for predicted breast age, and the EMBED dataset showing HRs of 1.030 for chronological age and 1.038 for predicted breast age. These results show that predicted breast age, similar to chronological age, effectively stratifies breast cancer risk, demonstrating its potential utility in clinical risk assessment. 5 3

Our ensemble Mammo-AGE incorporates various advanced convolutional neural network (CNN) backbones (e.g., EfficientNet³⁵, ConvNeXt³⁶, DenseNet³⁷, and ResNet³⁴). The detailed results for each CNN backbone-based Mammo-AGE are presented in Table 2. To evaluate the effectiveness of our proposed framework, we compared each backbone-based Mammo-AGE model with its corresponding CNN counterpart, including ResNet-18³⁴, ResNet-50³⁴, EfficientNet-b0³⁵, ConvNeXt-tiny³⁶, and DenseNet-121³⁷. Furthermore, we extended the comparison to include two transformer-based variants: Vision Transformer (ViT)³⁸ and Swin Transformer (Swin-ViT)³⁹. All these experiments were conducted using official implementations with pretrained weights. These results show that our proposed architecture consistently outperforms both standard CNNs and VITs.

Besides, we also provided an additional comparison (as shown in Supplementary Fig. 11) with the previous SOTA risk model³⁰, which aims to predict breast cancer risk, with age prediction as an auxiliary task in coarse bins (six age groups: 40-100 years old). Mammo-AGE consistently outperformed the SOTA risk model, achieving a lower MAE and also higher correlation and accuracy, demonstrating its superior ability to capture biologically relevant breast aging features.

Several ablation studies were conducted to optimize the model design. The performance of the Mammo-AGE model based on the different backbones for each fold is summarized in Supplementary Table . The weighted-ensemble Mammo-AGE model, which weighted averages the outputs of the different backbones (ResNet-18, ResNet-50, ConvNeXt-Tiny, DenseNet-121, and EfficientNet-B0), achieved the best performance. 1

Supplementary Table 2 presents the results of the ablation study on the combined dataset for five-fold cross-validation, investigating the impact of the proposed instance-bag transformer module, additional loss functions, and multi-task learning that were introduced to benefit the breast age estimation model. For fair comparison and computational efficiency, during the ablation studies, all methods comparing different modules or image sizes utilized the same backbone (ResNet-18). Our proposed approach, the ResNet18-based Mammo-AGE model, showed superior performance with MAE = 4.736 ± 0.076, CS (α = 5 y) = 66.5% ± 0.7%, and r = 0.855 ± 0.004 (Supplementary Table 2). Figure 3F shows CS curves in different error levels. Additionally, we note that when integrating predicted breast density into the embedding layer, the accuracy of the model could be further improved (Supplementary Table 5). We also provided the results of the breast density prediction of our Mamm-AGE model (Supplementary Table 6).

Furthermore, we investigated the performance of the Mammo-AGE model (using ResNet-18 as the backbone) with different input sizes of mammograms (as training with full-resolution mammograms is very demanding on the computational resources). We tested image sizes of 2048 × 1024, 1536 × 768, 1024 × 512, 512 × 256, and 256 × 128. The comparative analysis revealed that the Mammo-AGE using 1536 × 768 achieved the optimal performance (Supplementary Table ). 2

The breast age gap is defined as the difference between the predicted breast age by the deep learning model and chronological age (predicted breast age—chronological age). The breast age gap, when positive, indicates "older" appearing breasts, while a negative breast age gap suggests "younger" appearing breasts. To ensure a precise assessment of aging through the breast age gap, it is crucial to address the age bias inherent in the model. Age bias commonly occurs in data-driven deep learning age-predicting models due to factors such as regression dilution, model regularization, and the influence of age distribution within the training dataset^19,42.

After bias correction, there is no obvious correlation between the corrected breast age gap and chronological age (Pearson correlation coefficient, r = 0.04). The MAE curve demonstrates improved stability and age independence in the corrected predictions. Consequently, the overall performance of the model, evaluated using five-fold cross-validation, displays enhanced results: MAE = 3.698 ± 0.039, CS (α = 5 y) = 77.6% ± 0.4%, and r = 0.929 ± 0.002 on the combined test set (Table 2).

Figure 5A presents a heatmap illustrating the MAE results of the breast age model on subgroups categorized by different age ranges and breast densities, with MAE values ranging from 2.7 to 5.5. The model's performance demonstrated a higher degree of agreement between internal and external datasets, indicating its generalizability. On internal datasets, the MAE ranges from 3.497 to 4.155, while on external datasets, it ranges from 4.131 to 4.568 (as shown in Table 2). The bias-corrected model results in improved accuracy, minimizing the influence of age-related bias and enhancing its applicability across diverse datasets.

For internal (seen) datasets, the breast age gap of the breast cancer-diagnosed group was significantly higher than that of the healthy women group (inhouse dataset: P = 0.017; RSNA dataset: P < 0.001; VinDr dataset: P = 0.033; ANCOVA, adjusted by age and breast density). The results of the unadjusted p values are reported in Supplementary Table 7, confirming that the observed differences are independent of age and breast density. Scatterplots of the breast age gap and the chronological age are provided in Fig.6. We also calculated the Pearson correlation coefficients between the breast age gap and the chronological age for both groups. Results showed that after the bias-correction by chronological age, there is no obvious correlation between the breast age gap and the chronological age for both healthy and breast cancer groups among these datasets. For external (unseen) datasets (EMBED and CMMD), the breast age gap in the breast cancer-diagnosed group was significantly higher than in the healthy women group (P = 0.039 and P = 0.049, respectively; ANCOVA, adjusted by age and breast density).

To further eliminate potential model bias towards particularly young or older groups, we further explored differences in the breast age gap between normal and breast cancer groups in women aged 40–74 years⁴³ (the target group for breast cancer screening in normal-risk women). Similar group differences persisted among datasets (Inhouse dataset: P = 0.001; RSNA dataset: P < 0.001; VinDr dataset: P = 0.005; ANCOVA, adjusted by age and breast density). Significant differences were also observed in both CMMD and EMBED datasets for the screening target population (P = 0.023 and P = 0.040, respectively; ANCOVA, adjusted by age and breast density).

To explore the robustness and generalizability of the age-related biomarkers from mammograms defined using our approach, we retrained our Mammo-AGE model (only using ResNet-18) separately using each dataset (Inhouse, RSNA, VinDr, EMBED) from scratch. The CMMD dataset was excluded from the model training because it doesn't contain healthy mammogram examinations. Then, we evaluated breast age gap differences between healthy and breast cancer groups for every dataset based on each retrained model. Results based on each specific dataset-trained Mammo-AGE model were consistent with those based on the Mammo-AGE-Combined model, indicating that even with different dataset-trained models, the breast age gap in the breast cancer group tended to be higher than in the healthy group (as shown in Supplementary Fig. ). These results support the robustness and generalizability of the breast age gap as a biomarker for indicating breast cancer. 6

Considering the observed significant difference in the breast age gap between the healthy and cancer groups, we define the women whose breast age gaps are higher than the median (inhouse dataset: 0.36; EMBED dataset: −0.01) of the dataset as a high-risk group and those with lower gaps as a low-risk group. The Kaplan–Meier (KM) curves of women in the high and low-risk groups, which are stratified by breast age gaps of Inhouse and external EMBED datasets, are shown in Fig. 6C. We note that in both datasets, the breast cancer incidence rate in the low-risk group is indeed significantly lower than the high-risk group (both inhouse and EMBED datasets: P < 0.01).

After adjusting for confounding factors such as chronological age and breast density, each 1-year increase in breast age gap was associated with 1.6% and 2.0% increase in breast cancer risk on the inhouse and EMBED datasets, respectively (Inhouse dataset: HR = 1.016 [95% CI: 1.009–1.023], P < 0.001; EMBED dataset: HR = 1.020 [95% CI: 1.008 - 1.033], P = 0.002; Table 3, Model 2). To ensure the validity of the Cox model estimates, the proportional hazards assumption was verified by visual inspection and Schoenfeld residuals tests^15,44 (as shown in Supplementary Fig. 7). Women with a breast age gap higher than the median had significantly increased 10 or 5-year breast cancer risk compared to those with a low gap (Inhouse dataset: HR = 1.157 [95% CI: 1.084–1.235], P < 0.001; EMBED dataset: HR = 1.152 [95% CI: 1.071–1.305], P = 0.026). Further adjusting for additional breast cancer risk factors (Table 3, Model 3) did not substantially alter the association between breast age gap and cancer risk, supporting its robustness and independence. These results suggest that positive breast age gaps (where the breast appears older than the chronological age) are associated with a substantially increased risk of breast cancer.

Allowing for non-linearity, Supplementary Fig. 8 presents the estimated relationship between breast age gap and breast cancer risk. Significant overall and non-linear associations were observed (Inhouse dataset: P_overall < 0.001; P_non-linear = 0.024; EMBED dataset: P_overall < 0.001, P_non-linear = 0.016).

Besides, to better inform the optimal age at which to target mammographic screening, the odds ratios and diagnostic age analysis results (Supplementary Table ) show that extremely accelerated-aging patients were diagnosed earlier (62 years, 95% CI: 47–80), compared to extremely decelerated-aging women, who had a later average diagnostic age (72 years, 95% CI: 47–98). 9

To assess the clinical relevance of the breast age gap compared to current standards, we benchmarked its predictive performance against breast density, a widely accepted imaging-based risk factor and a criterion used in screening practice (e.g., Netherlands Early Breast Neoplasm Screening (DENSE) trial⁴⁵). Under matched high-risk population sizes based on the DENSE trial criteria (i.e., to select the same number of high-risk women), Mammo-AGE consistently identified more breast cancers than breast density across 1-, 5-, and 10-year follow-up periods (Supplementary Fig. 12). Notably, Mammo-AGE identified approximately 30% more cancers than the DENSE trial criteria in the EMBED screening dataset at 1-year follow-up. Additionally, Supplementary Table 8 presents the odds ratio analysis of breast cancer based on breast density and breast age gap.

We evaluated the Mammo-AGE model for risk prediction across three datasets and for breast cancer diagnosis across six datasets. In all tasks, Mammo-AGE outperformed both the baseline model and SOTA risk model⁷ and SOTA pretrained methods⁴⁷. Specifically, Mammo-AGE (with age pretraining) surpassed the risk model across all metrics (AUPRC: 0.40 vs. 0.36; AUROC: 0.75 vs. 0.71 in the inhouse dataset, for long-term risk prediction). Importantly, performance improvements were consistently observed when fine-tuning with age-pretrained weights compared to training from scratch (baseline model). To further investigate this learning efficiency advantage, we fine-tuned Mammo-AGE using different fractions of training data for the breast cancer diagnosis task. As shown in Supplementary Fig. 13, Mammo-AGE consistently outperformed training from scratch (i.e., baseline), particularly under limited data conditions (e.g., using only 5% of the training set, approximately 2000 mammogram examinations).

This study conducted experiments on five datasets from various origins and locations, as summarized in Table 1. These datasets include one large inhouse longitudinal dataset and four publicly available datasets: RSNA dataset^23,63, VinDr-Mammo dataset²⁶, EMory BrEast Imaging (EMBED) dataset²⁴, Chinese Mammography Database (CMMD) dataset⁶⁴. Besides, we also collected the CSAW-CC⁴³ dataset for the external validation during the downstream risk prediction and breast cancer diagnostic tasks.

The overview framework of our Breast-Age model is illustrated in Fig. 1A. Our proposed Breast-Age model is a multi-view-based method, which utilizes four views of mammograms as input (right craniocaudal (CC), right mediolateral oblique (MLO), left CC, and left MLO views), considering bilateral breasts. The baseline convolutional neural network, such as ResNet, DenseNet, etc., is utilized as the shared-weights backbone to extract deep features from each input image. The specific dimension of input data was 1536 × 768 in our study. Experimental tests measuring how different backbones and the input image size affected the model's results were performed separately (see the ablation experiments).

The features before the last pooling and full connection of the backbone are fed into the instance-bag transformer block. The proposed instance-bag transformer is inspired by the previous Global-Local transformer¹⁷ to combine the information of the examination for more accurately predicting breast age. The motivation of the proposed transformer module is to learn fine-grained information from a single view by fusing the global context visual information learned from the whole mammogram examination (four views of the bilateral breast). We define the features of each view as local features and fuse these four local features together to form global features. As shown in Fig. 1A, this transformer module includes self-attention and cross-attention blocks, which are similar to previous studies⁶⁵. Both self-attention and cross-attention blocks are based on multi-head attention, where the global and local features are split into h = 8 parallel parts on the channel dimension. For self-attention, the query, key, and value vectors are from the same features to learn the fine-grained local information. For the cross-attention, we replace the key and value vectors with those from the global features to fuse the global context information. Then, local features and the self-attention and cross-attention generated features are concatenated in the channel dimension and fed into the feed-forward layers to fuse the information and maintain the same size as the local features. The transformer block is repeated N = 1 times to iteratively integrate information.

Following the transformer model, the features on each branch are channeled into the final layer for predicting the biological age of the breast. Different from the direct regression methods^{11,12,14,16,25}, which simply map the feature vector to the age through a fully connected layer, we utilized a Probabilistic Ordinal Embeddings (POE)³² learning strategy in the final layer. This is an uncertainty-aware regression method, which represents each data as a multivariate Gaussian distribution rather than a deterministic point in the latent space and learns a distribution following the ordinal constraint in the embedding space. For instance, it tries to define each mammogram as an age distribution with an uncertain range in the embedding space rather than an accurate point. Like a human, who will have different confidence in their judgments with different judgment bases. A mammogram with less content information or more noise will lead to larger uncertainty. Besides, the ordinal relations of different-aged women's mammograms are also considered in the embedding space. For instance, consider three mammograms from women aged 30, 50, and 70. In the embedding space, the ordinal relations of these samples are constrained (e.g., 30 < 50 < 70). This resulting ordinal distribution constraint in the latent space contributes to ordered probabilistic prediction. Thus, incorporating such embedding learning can yield better predictive performance in the age-estimation task. This module includes several loss functions (referred to in ref.³²); in this study, we collectively refer to them as a composite loss function of POE.

Similar to previous research^32,33, we treated the age estimation task as a classification task with 100 classes corresponding to ages 1–100. We also utilized mean-variance (MV) loss³³ in this study. Specifically, the MV loss consists of a mean loss, which penalizes the difference between the mean of the estimated age distribution and the ground-truth age, and a variance loss, which penalizes the variance of the estimated age distribution to ensure a concentrated distribution. Besides, considering the correlation between breast density and chronological age³¹, we also constrained the model to learn the breast density of the mammograms through multi-task learning. The combined features are leveraged to feed into the density classification layer, and correspondingly, the cross-entropy (CE) was used to calculate the density classification loss.

In this study, we leveraged a variety of well-established neural network architectures, ResNet-18³⁴, ResNet-50³⁴, EfficientNet-b0³⁵, Convnext-tiny³⁶, and DenseNet-121³⁷, initializing them with ImageNet pre-trained weights as the backbone of all our methods. The weight ratios were derived using five-fold cross-validation by first calculating the weight ratios for each backbone model based on its performance (MAE) in each validation fold. These fold-specific weight ratios were then averaged across all five folds to obtain the final set of weight ratios. This ensures a balanced contribution from different architectures while maintaining diversity, reducing the risk of overfitting.

All models are implemented in PyTorch (version 1.12.1). The training strategies were standardized across all methods. Specifically, we opted for the Adam⁶⁶ optimizer with a rate of 0.5 for dropout⁶⁷ after every fully connected layer. The models underwent training for 100 epochs with an initial learning rate of 10⁻⁴, which is decayed by a factor of 10 every 20 epochs. Early stopping was applied when the validation loss did not decrease for 10 consecutive epochs. The experiments are performed on a Quadro A6000 GPU (48GB). We set the batch size of 32 when training the ResNet-18-based method and the batch size of 6 for other backbone-based models due to the limitation of the GPU memory. The selection of the best models for each method was based on the best MAE performance index on the validation set.

The source code is available at https://github.com/Netherlands-Cancer-Institute/Mammo-AGE↗. Mammo-AGE was developed based on standard DICOM mammograms. First, we converted the images into 16-bit PNG format and segmented the whole breast region to exclude the background⁶⁵. For unifying the size of all images, we simultaneously perform resizing and padding to standardize the image size to 1536 × 768 pixels, which can preserve the original aspect ratio of the mammogram image and avoid structural distortion. We also employ data augmentation techniques (i.e., random flip, brightness, contrast, Gaussian blur, Gaussian noise, and coarse dropout) during training for model robustness and overfitting prevention. All augmentations were executed using the Albumentations Python library (https://github.com/albumentations-team/albumentations↗). Besides, during model inference, it is hardware-friendly and runs efficiently (~4 s per exam on a 2080 Ti (12GB) GPU). It can be easily deployed on clinical servers or cloud-based systems, supporting integration into existing radiology workflows.

To optimize our model design and assess the effectiveness of each proposed component in enhancing breast age estimation performance, we conducted a series of ablation experiments. In the first set of analyses, we explored the model's performance with various backbones, such as ResNet-18, ResNet-50, ConvNeXt-Tiny, DenseNet-121, EfficientNet-B0, and an ensemble model that combined all the predictions of these five models. The ensemble model's predictions were selected for further analysis and consideration as our final results in this study. Additionally, we evaluated the model's performance by removing specific components, including the proposed transformer block (w/o CI-Transformer), POE loss (w/o POE), MV loss (w/o MVL), and the breast density learning (w/o Multi-task). These experiments were implemented using ResNet-18 as the backbone. Furthermore, we investigated the model's performance with different sizes of mammograms, specifically 256 × 128, 512 × 256, 1024 × 512, 1536 × 768, and 2048 × 1024 pixels. All results are based on the five-fold cross-validation.

The presence of systematic bias, induced by regression dilution, introduces a correlation between the estimated age gap and chronological age^19,68. This leads to a tendency to overestimate age in younger individuals and underestimate it in older individuals^19,68,69. To mitigate this phenomenon, we applied the linear bias correction method as outlined by Smith et al. ⁴².

The systematic bias is defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Delta }_{{bias}}=\hat{y}-x$$\end{document}Δbias=y^−x, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{{{\rm{y}}}}$$\end{document}y^ denotes the predicted breast age and x is the chronological age. To correct for bias, we fit a linear regression model: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat{\Delta }}_{{bias}}={ax}+b$$\end{document}Δ^bias=ax+b, using the healthy reference population to estimate systematic bias \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Delta }_{{bias}}$$\end{document}Δbias. Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{a}}}$$\end{document}a and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\rm{b}}}$$\end{document}b are the represent the slope and intercept derived from the healthy group. The bias-corrected estimated breast age is then computed as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat{y}}_{{\mathrm{corrected}}}=\hat{y}-{\hat{\Delta }}_{{bias}}=\hat{y}-({\mathrm{ax}}+b)$$\end{document}y^corrected=y^−Δ^bias=y^−(ax+b), and the corrected breast age gap is:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Delta }_{{\mathrm{breast}}{\mathrm{age}},{\mathrm{corrected}}}={\hat{y}}_{{\mathrm{corrected}}}-x$$\end{document}Δbreastage,corrected=y^corrected−x.

Following the standard aging study method¹⁹, bias correction coefficients were obtained from the validation set (healthy population) to ensure that the breast age gap reflects true biological aging rather than systematic prediction bias, remaining independent of health status. These coefficients, derived from the healthy group, are then applied consistently across breast cancer diagnostic groups for the purpose of bias correction. This linear correction ensures that the estimated breast age gap accounts for systematic bias, providing a more accurate assessment of biological aging across different diagnostic groups.

To reflect the importance of different regions of the breast in estimating biological age, we employed the occlusion sensitivity analysis¹⁹ to calculate the age-specific saliency map. This method calculates the difference in the model's performance when masking portions of input mammograms, offering insights into the significance of various breast structures, such as calcifications, fibroglandular structures, fatty tissue, and lymph nodes. Given the considerable size of mammograms and the varying sizes of their contents, reflecting distinct structures within the breast, we utilized masks of different sizes (32 × 32, 64 × 64, 128 × 128, and 256 × 256). The occlusion process involved iteratively masking the original mammograms along a grid, with each mask size corresponding to a grid cell. The occluded images, represented by the masked areas with a black image, were then evaluated using the MAE metric to assess the model's performance. The delta MAE, calculated as MAE_occluded- MAE_original, was employed to quantify the extent to which the model's performance was affected. Here, MAE_occluded signifies the performance with occluded images, while MAE_original represents the performance with the original mammogram. Then, four delta MAE metrics, each associated with a specific mask size, were integrated and reconstructed as a saliency map with the same size as the mammograms through cubic interpolation. Therefore, the saliency map can accurately and finely represent the importance of breast tissue related to breast aging, regardless of the size difference of breast tissue.

In this study, the primary results are derived from our proposed model, ensembled Mammo-AGE-Combined, which is trained on a combined dataset. To assess the robustness of the age-related biomarkers extracted by the Mammo-AGE model from mammograms, we conducted four separate dataset retraining experiments, resulting in four specific-dataset retrained Mammo-AGE models. Each model was retrained from scratch using the corresponding dataset. For the EMBED dataset, we also selected normal mammograms (BI-RADS 1 and 2) and split them into training, validation, and test sets at the patient level. Subsequently, we replicated the breast age gap analysis to ascertain the consistency of the results compared to those based on the Mammo-AGE-Combined model. This retraining experiment allows us to evaluate the generalizability and reliability of the age-related biomarkers extracted by our model across different datasets.

To measure the performance of our breast age prediction model, we employed four key evaluation metrics: Mean Absolute Error (MAE), the Pearson correlation coefficient (r), the Spearman's rank correlation coefficient (r_s), and Cumulative Score (CS). These metrics align with established methods in previous studies^17,19. The MAE is defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{MAE}}=\frac{{1}}{n}{\sum }_{i{-}{1}}^{n}\left|y{-}x\right|\times 100\%$$\end{document}MAE=1n∑i−1ny−x×100% where y is the predicted breast age and x is chronological age. It is widely used for evaluating the performance of age estimation; lower MAE means more accurate prediction. The Pearson correlation coefficient (r) is calculated based on the assumed linear relationship between predicted breast age and chronological age. The nonparametric based Spearman's rank correlation coefficient (r_s) is also reported. The CS represents the accuracy of age estimation given a threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}α (5 years in this study), calculated as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$cs\left(\alpha \right){=}\frac{{N}_{\varepsilon {\le }\alpha }}{N}\times 100\%$$\end{document}csα=Nε≤αN×100% where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${N}_{\varepsilon {{\boldsymbol{\le }}}\alpha }$$\end{document}Nε≤α is the number of samples with an absolute error of age estimation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varepsilon$$\end{document}ε no higher than the threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha$$\end{document}α. A higher CS score signifies better performance.

Descriptive statistics, including means and standard deviations (SDs), numbers, and percentages, were utilized to report baseline characteristics. Chronological age and breast density adjusted analysis of Covariance (ANCOVA)⁷⁰ was used to compare the difference in breast age gap between breast cancer patients and the healthy group. Cox proportional hazards regression models considering breast age gap as a continuous linear term were fitted to estimate the effect of a 1-year increase in breast age gap on breast cancer risk. Adjustment of covariates in Cox models was made for chronological age, breast density, and additional known risk factors or covariates. We tested the proportional hazards assumption using Schoenfeld residuals^15,44. For covariates that violated this assumption, we employed stratification in the Cox models to account for their nonproportionality. Several sensitivity analyses were conducted, including covariate adjustments and dataset-specific model validation. Differences between Kaplan–Meier survival curves of high- and low-risk groups were assessed using the log-rank test. All the statistical tests are two-sided. The analyses were performed using R (version 3.3.0, R Foundation for Statistical Computing, www.R-project.org↗, Vienna, Austria) and Stata (version 13, StataCorp, Texas, USA). The 95% confidence intervals (CI) of AUROC and AUPRC metrics are estimated by bootstrapping with 1000 bootstraps for each measure. No statistical method was used to predetermine sample size. All available eligible mammogram examinations meeting the inclusion criteria were included in the analyses. No data were excluded from the analyses.