Validation of an AI model for pediatric myelin maturation age assessment and benefits for junior radiologists

BACKGROUND: Artificial intelligence (AI) demonstrates potential for estimating myelin maturation age from pediatric brain magnetic resonance imaging (MRI) scans. However, its generalizability across diverse clinical settings and performance in infants with delayed myelination remain unvalidated. Additionally, the utility of such AI models in assisting radiologist interpretation is unclear. This study aimed to validate an AI model for myelin maturation age prediction in our institution and compare the assessment performance of junior radiologists with and without AI assistance. MATERIALS AND METHODS: A pretrained convolutional-neural-network (CNN) model was tested on 817 brain MRI scans (0–12 months): 768 with normal myelination and 49 with documented delay. The ground truth of myelin-maturation age was set by consensus of two experienced pediatric neuroradiologists. Model performance was evaluated using mean absolute error (MAE) and Spearman’s correlation coefficient between ground-truth and AI-predicted age. For the utility study, three junior radiologists independently reviewed 147 consecutive cases in a crossover design (AI-assisted vs. unassisted), separated by a 4-week washout. MAE and reading time were compared with paired t-tests. RESULTS: The AI model achieved an MAE of 1.3 months (95%CI: 1.2, 1.4), demonstrating a strong positive correlation (Spearman’s r = 0.933, P < 0.001). There was no significant difference in MAE between the delayed myelination group (1.3 months, 95% CI: 1.0, 1.6) and the normal myelination group (1.3 months, 95% CI: 1.2, 1.4) (P = 0.978). Correlation strength differed significantly between groups (P < 0.001). When using AI assistance, junior radiologists significantly reduced their MAE from a baseline range of 0.5–0.8 months to 0.2 months (P < 0.001 for each radiologist) and decreased reading times from 53.8 to 65.2 s to 48.6–56.8 s (P < 0.001). CONCLUSIONS: The AI model was successfully validated at our institution, demonstrating accurate prediction of myelin maturation age, even in infants with delayed myelination. Furthermore, providing AI predictions significantly enhanced both the accuracy and efficiency of myelin maturation age assessments performed by junior radiologists. This supports the potential clinical utility of AI assistance in pediatric neuroimaging workflows.