Abstract:
Age estimation from medical images can provide insight into biological ageing and disease risk. In domains such as brain MRI, retinal photographs, and chest radiographs, deep learning models have shown that tissue appearance contains information predictive of chronological age. In mammography, however, this question has received limited attention. This thesis investigates whether chronological age can be estimated directly from full-field digital mammograms and whether the resulting age gap (predicted age minus true age) may capture clinically relevant variation. A multi-view deep learning model was developed to process the four standard screening mammographic views and fuse information across projections. The model was trained for age regression on the OPTIMAM Mammography Image Database (52,556 patients; ages 47–73 years) and evaluated on an internal evaluation test set and two external cohorts: EMBED (6,111 patients, United States) and CMMD (455 patients, China), spanning different populations and scanner manufacturers. The best individual model, ConvNeXt- Small, achieved a mean absolute error of 3.33 years on the internal test set, 4.20 years on EMBED, and 3.69 years on CMMD. These findings show that mammographic age estimation is feasible and can generalize across heterogeneous datasets. Comparisons with recently published approaches, including MammoAge, should be interpreted cautiously because cohort composition, preprocessing, and evaluation protocols differ. As a secondary exploratory analysis, this thesis examined whether age gap differed across pathology-defined groups using models trained either on all mammograms or on normal mammograms only. On EMBED, significant differences were observed between normal and malignant groups under both training conditions, with a slightly larger effect for the normal-only model. However, these results are sensitive to regression bias and dataset-specific confounding. Therefore, this thesis does not establish age gap as a validated breast cancer risk marker, but instead identifies it as a hypothesis-generating signal for future longitudinal study.
