幼児の2歳未満の脳年齢予測を改善したミエリン形成に焦点を当てた深層学習モデル

Myelination is a key biomarker of healthy brain maturation, and its disruption can signal neurodevelopmental disorders.

The study aimed to enhance the accuracy and interpretability of brain age prediction in early infancy by incorporating the biological process of myelination as an attention mechanism into deep learning models.

A fully automated deep learning framework, called myelination-attention-empowered model (MAENet), was developed through retrospective analysis of structural magnetic resonance imaging (sMRI) data from 603 participants who met the inclusion criteria, aged 0-2 years, collected in a local hospital between July 2017 and June 2024. The MAENet consisted of four modules: a multiscale information fusion channel (MSIF-channel) on the T2WI brain image, a myelination-empowered feature extraction channel (MEFE-channel) on an automated and standardized segmentation of the white matter image, a communication mechanism that enabled inter-channel information flow and enhanced the MSIF-channel's sensitivity to myelination-related features, and a myelination-attention mechanism that dynamically emphasized myelination-sensitive regions.

The proposed MAENet model exhibited superior performance over multiple deep learning models, including ResNet-50, VGG, Inception, SFCN, Skewed, FiA-Net, and TSAN. The mean absolute error (MAE) between the predicted brain age and chronological age was significantly reduced by 18%-41% in the subgroup of 0-1-year-old infants, 25%-37% in the subgroup of 1-2-year-old infants, and 18%-40% in the whole group of 0-2-year-old infants in the experimental comparison (P < 0.05). The brain regions attended to by the MAENet model were visualized and consistent with the well-known developmental trajectories of white matter myelination in early infancy.

The MAENet model demonstrated a significant improvement in brain age prediction accuracy in 0-2-year-olds by effectively leveraging the developmental process of myelination.