Age and sex estimation from Japanese femoral computed tomographic images using foundation models

This study evaluated the feasibility and performance of a deep learning-based multitask framework for estimating sex and age using femoral CT images in a Japanese population.

Postmortem CT (PMCT) data from 1,489 Japanese individuals (1,003 males, 486 females) aged 20-90 years were analyzed. Femora were automatically segmented using the TotalSegmentator library, and three-dimensional volumetric data were processed using a three-dimensional Vision Transformer foundation model (3DINO) adapted with low-rank adaptation. The model was trained in a multitask learning framework to perform sex classification and age estimation simultaneously. Performance was assessed, and attention map analysis was conducted to identify anatomically relevant regions contributing to predictions.

On the independent test dataset, sex estimation showed high performance, with balanced accuracies of 0.897 (95% confidence interval [CI], 0.887-0.948) for the left femur and 0.929 (95% CI, 0.907-0.961) for the right femur. Area under the curve values were also high (0.979 and 0.988 for the left and right femora, respectively). Age estimation achieved a mean absolute error of approximately 8 years and a root mean square error of approximately 10 years; female models showed lower mean absolute error and root mean square error values than male models. Attention map analysis indicated that the distal femur contributed most to sex estimation, whereas the femoral shaft was the primary region associated with age estimation.

A foundation model-based approach enables accurate simultaneous estimation of sex and age using femoral CT images and has potential applications in automated forensic biological profiling.