Neonatal brain age estimation from T2-weighted magnetic resonance images using shallow regression analysis.

This study develops a method for estimating brain age (postmenstrual age, PMA) in preterm and term neonates using T2-weighted Magnetic Resonance Imaging (MRI) data and shallow regression analysis. The goal is to provide an accurate biomarker for brain maturation in early life. We used T2-weighted imaging data from 885 neonates from the Developing Human Connectome Project (dHCP). To investigate the effects of linear and nonlinear registration, as well as the impact of template selection on brain age estimation error, we registered the input images using both linear and linear-nonlinear transformations to align the images with three neonatal brain templates from the dHCP: (i) the age-specific template, (ii) the full-term 40-week neonatal brain atlas template, and (iii) an arbitrary template randomly selected from a pool of 17 templates designed for neonates. We then applied principal component analysis (PCA) to reduce the dimensionality of the input feature sets in voxel space. The most relevant components, showing higher correlation with PMA, were identified and used to train and test three commonly used shallow regression models: Gaussian Process Regression (GPR), Support Vector Regression (SVR), and Relevance Vector Regression (RVR) using ten-fold cross-validation. Using an age-specific template with linear-nonlinear registration, GPR and SVR achieved mean absolute errors of 0.438 ± 0.06 weeks and 0.437 ± 0.06 weeks, respectively, outperforming RVR (0.441 ± 0.058 weeks) with 350 principal components and an R² of 0.95. SVR showed instability with more components. Template selection and registration type significantly affect brain age estimation performance in neonates.