Biologically Informed Brain Age Estimation in Infants Using T1w/T2w Ratio Imaging and Anatomical Priors

Introduction: Accurately estimating brain age in infants is essential for monitoring neurodevelopment and identifying early signs of atypical maturation. However, conventional MRIs (T1w and T2w) have limited sensitivity to early myelination. To address this, we propose a biologically informed deep learning model that leverages T1w/T2w ratio images (RI) and structural priors—white matter volume (WMV) and intracranial volume (ICV)—to enhance brain age estimation (BAE) by focusing on developmentally relevant features.

Methods: We retrospectively collected 630 clinical T1w and T2w brain MRI scans from infants aged 0–24 months. Ground-truth labels represented biological ages of normal subjects, based on radiologist review with clinical information. The dataset was divided into training, validation, and test sets (3:1:1).

To generate RI, we performed the following preprocessing steps: 1) Linear registration of T2w to the corresponding T1w images. 2) Brain extraction and bias field correction using FSL. 3) Nonlinear registration to an 18-month-old brain template to normalize brain size and improve model generalizability. RI was calculated via voxel-wise division of the preprocessed T1w by the T2w (RI = T1w/T2w) and used as input to 3D convolutional

neural network.

WMV and ICV were extracted using Infant FreeSurfer, with four registration samples used for subject-specific atlas alignment. WMV was measured by combining left and right WM regions, and ICV was estimated from the transformation matrix of atlas registration. These anatomical priors were injected after intermediate residual blocks to modulate feature maps via conditional scaling, enabling the model to better capture biologically relevant features.

Results: We conducted 5-fold cross-validation to evaluate model robustness. Performance was evaluated using mean absolute error (MAE), root mean square error (RMSE), Pearson correlation coefficient (PCC), and coefficient of determination (R2).

The model trained on RI achieved significantly lower MAE (1.149 months) compared to models trained on T1w (1.437) or T2w (1.330) alone. This demonstrates the RI's superior sensitivity to early myelination, especially important during early infancy when conventional contrasts are less informative. Additionally, incorporating anatomical priors consistently improved prediction accuracy across all input modalities by providing complementary structural information that enhanced biologically relevant feature learning.

Discussion & Conclusion: By integrating biologically sensitive imaging features (RI) with anatomical priors (WMV and ICV), our framework significantly improves the accuracy of infant BAE. RI provides enhanced contrast for early myelination without the need for additional imaging sequences, making this a scan-efficient and clinically practical tool. This approach shows promise as a reliable imaging biomarker for tracking early brain development and enabling timely intervention in pediatric neurodevelopmental care.