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

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Pediatric Radiology

Background:

• Accurate brain volume quantification in pediatric CT scans is crucial for clinical assessment and research.
• Existing methods for brain segmentation in children can be labor-intensive and may lack standardization.
• Establishing normative reference data for brain development in early childhood is essential.

Purpose of the Study:

• To develop and validate an automated deep learning method for segmenting brain CT images in healthy 2-year-old children.
• To quantify regional brain volumes and establish a normative reference database.
• To compare the performance of the ResU-Net model with different kernel sizes against a standard 3D U-Net.

Main Methods:

• A retrospective study included 1,487 pediatric head CT scans (2-year-olds).
• The ResU-Net deep learning model was trained and validated for brain image segmentation.
• Performance was evaluated using the Dice similarity score; regional volumes were derived and statistically analyzed for sex, hemisphere, and age correlations.

Main Results:

• The ResU-Net model achieved high segmentation accuracy (Dice coefficient of 0.96 on the testing set).
• ResU-Net (3,3,3) outperformed other models, including the baseline 3D U-Net (0.88).
• Significant differences in brain volume by sex and hemisphere were observed, with positive correlations between brain volume and age.

Conclusions:

• Deep learning, specifically the ResU-Net model, is effective for automated brain segmentation in pediatric CT imaging.
• The study provides a reliable reference for normative brain volumes in 2-year-old children.
• These findings establish a benchmark for clinical assessment and research, complementing MRI data and addressing the need for accessible pediatric neuroimaging standards.