Perturbation-Based Deep Learning Reveals Dynamic Spatiotemporal Shifts in Pediatric Brain Development from CT Imaging

Abstract

Understanding pediatric brain maturation is essential for detecting neurodevelopmental disorders, yet quantitative analysis using CT has been limited. This study proposes a deep learning method that estimates brain age from pediatric CT scans while simultaneously elucidating the contribution of each brain region during developmental stages. The method involves training a 3D ResNet model on CT data from 201 patients aged 0 to 47 months and evaluating the importance of each region through perturbation analysis. A two-level perturbation analysis is conducted, using lobe-wise masks for evaluation based on anatomical information and patch-wise masks for more detailed localization. As a result, the model achieved high accuracy, estimating brain age with a Mean Absolute Error of 4.007 months and a correlation coefficient of 0.925. Furthermore, the results of the perturbation analysis also revealed a dynamic posterior-to-anterior shift in the process of brain development.