Deeply Supervised Multitask Autoencoder for biological brain age estimation using structural magnetic resonance imaging

Brain age estimation from magnetic resonance imaging is challenging due to subtle and heterogeneous age-related structural changes, along with substantial anatomical and demographic variability across individuals. To address this, this study proposes a deep learning-based artificial intelligence framework termed the Deeply Supervised Multitask Autoencoder (DSMT-AE), which combines deep supervision for stable optimization with multitask learning to enhance representation quality. DSMT-AE jointly optimizes brain-age regression with two auxiliary objectives: biological sex classification and self-supervised image reconstruction, enabling the model to capture both demographic and anatomical variation. The proposed method is evaluated on the large publicly available cohort, the Open Big Healthy Brains (OpenBHB), and demonstrates superior performance, achieving a Mean Absolute Error (MAE) of 2.64 years, a Root Mean Square Error (RMSE) of 3.08 years, and a Coefficient of Determination ( R 2 ) of 0.94, substantially improving upon prior methods. These results demonstrate robust performance across age and sex subgroups and highlight the potential of DSMT-AE for sensitive detection of brain-aging patterns relevant to neurodegenerative processes.