Infant Brain Age Estimation Using Deep Learning on Myelin-Sensitive T1w/T2w Ratio Images

Purpose: To develop a neural network (NN) that quantitatively estimates infant brain age using the T1w/T2w ratio image (RI), reflecting the degree of myelination.
Materials and Methods: A total of 630 clinical T1w and T2w brain MRIs from infants aged 0-24 months were retrospectively collected. Ground-truth labels represented biological ages of normal subjects, based on radiologist review with clinical information. The dataset was split 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 correction using FSL, assuming white matter intensity uniformity. 3) Intensity normalization to reduce inter-modality contrast differences. 4) Nonlinear registration to an 18-month-old brain template to normalize brain size and improve model generalizability. The RI was obtained by voxel-wise division of the preprocessed T1w by the T2w (RI = T1w/T2w) and used as inputs to 3D convolutional NNs such as SFCN (a VGGNet variant) and ResNet18 for biological brain age estimation (BAE). To assess robustness, 5-fold cross-validation was performed.
Results: Model performance was evaluated using mean absolute error (MAE), root mean square error (RMSE), Pearson correlation coefficient (PCC), and coefficient of determination (R2R2). The model trained on RI achieved significantly lower MAE (1.136 months) compared to T1w (1.368) or T2w (1.235) alone.
The enhanced accuracy of RI is due to its biologically informative contrast, effectively capturing myelination patterns. This is particularly valuable in infants, where T1w and T2w often lack clear tissue contrast and myelination is rapidly until 24 months. Grad-CAM showed the model focuses on myelinated regions, supporting the relevance of the RI. Incorporating T1w, T2w, and RI as multi-channel inputs further improved performance, demonstrating the benefits of integrating imaging modalities.
Conclusions: Due to its biological relevance and sensitivity to myelination, RI significantly improves BAE over T1w, T2w, or their combination. Accurate BAE is crucial for the early detection of atypical neurodevelopmental trajectories. By capturing myelination without extra scans or complex techniques such as myelin water fraction imaging, RI improves model accuracy and interpretability. This makes it a practical tool for pediatric neuroimaging, offering neurologists a biomarker for monitoring development and enabling timely intervention in at-risk infants.