Automatic segmentation and quantitative analysis of brain CT volume in 2-year-olds using deep learning model

Objective Our research aims to develop an automated method for segmenting brain CT images in healthy 2-year-old children using the ResU-Net deep learning model. Building on this model, we aim to quantify the volumes of specific brain regions and establish a normative reference database for clinical and research applications. Methods In this retrospective study, we included 1,487 head CT scans of 2-year-old children showing normal radiological findings, which were divided into training ( n = 1,041) and testing ( n = 446) sets. We preprocessed the Brain CT images by resampling, intensity normalization, and skull stripping. Then, we trained the ResU-Net model on the training set and validated it on the testing set. In addition, we compared the performance of the ResU-Net model with different kernel sizes (3 × 3 × 3 and 1 × 3 × 3 convolution kernels) against the baseline model, which was the standard 3D U-Net. The performance of the model was evaluated using the Dice similarity score. Once the segmentation model was established, we derived the regional volume parameters. We then conducted statistical analyses to evaluate differences in brain volumes by sex and hemisphere, and performed a Spearman correlation analysis to assess the relationship between brain volume and age. Results The ResU-Net model we proposed achieved a Dice coefficient of 0.94 for the training set and 0.96 for the testing set, demonstrating robust segmentation performance. When comparing different models, ResU-Net (3,3,3) model achieved the highest Dice coefficient of 0.96 in the testing set, followed by ResU-Net (1,3,3) model with 0.92, and the baseline 3D U-Net with 0.88. Statistical analysis showed that the brain volume of males was significantly larger than that of females in all brain regions ( p < 0.05), and age was positively correlated with the volume of each brain region. In addition, specific structural asymmetries were observed between the right and left hemispheres. Conclusion This study highlights the effectiveness of deep learning for automatic brain segmentation in pediatric CT imaging, providing a reliable reference for normative brain volumes in 2-year-old children. The findings may serve as a benchmark for clinical assessment and research, complementing existing MRI-based reference data and addressing the need for accessible, population-based standards in pediatric neuroimaging.