Radiologists in India are increasingly exploring pediatric AI imaging to improve patient outcomes. However, developers often train these tools using adult datasets. A recent scoping review reveals that such models perform poorly when used for children without adaptation. Consequently, the lack of age-specific training creates significant diagnostic risks.
The Challenges of Pediatric AI Imaging
Adult-trained AI models consistently demonstrate reduced accuracy in younger populations. Researchers found that performance drops significantly across various modalities like CT, MRI, and ultrasound. Therefore, applying these tools directly to children can lead to incorrect diagnoses. Specifically, detection tasks for pulmonary nodules showed a massive decline in sensitivity. Sensitivity fell from nearly 100% in adults to as low as 26% in children. This discrepancy occurs because a child’s anatomy differs fundamentally from an adult’s.
Vulnerable Age Groups and Task Specificity
The study highlights that children under two years old face the highest risks. Within this group, the performance gap is most pronounced across all AI tasks. Additionally, segmentation tasks for organs showed significant errors when using adult parameters. Furthermore, India-based studies suggest that ethnic differences in bone geometry also require localized AI validation. Because of these factors, healthcare professionals cannot assume that adult-approved software is safe for pediatric use.
Ensuring Clinical Safety and Accuracy
Clinicians must demand rigorous validation before implementing AI in pediatric workflows. Instead of relying on adult data, vendors should provide evidence of performance in children. Moreover, fine-tuning models with pediatric datasets can help bridge the current accuracy gap. This proactive approach protects the youngest patients from potential medical errors. By prioritizing age-appropriate tools, radiologists can harness the full potential of AI safely.
Frequently Asked Questions
Q1: Why do adult AI models fail in young children?
Adult models fail because children have unique anatomical structures and growth patterns that differ from the adult data used for training.
Q2: Which imaging tasks are most affected by performance drops?
Detection tasks, such as finding lung nodules, show the most severe deterioration when adult models are applied to children.
References
- Laborie LB et al. Performance of adult-trained artificial intelligence models in paediatric imaging-a scoping review. Eur Radiol. 2026 Feb 12. doi: 10.1007/s00330-026-12354-5. PMID: 41673142.
- Gassenmaier S, et al. Artificial intelligence in pediatric radiology: overview and current status. Pediatr Radiol. 2023.
- Khadilkar AV, et al. Adaptation and validation of an artificial intelligence based digital radiogrammetry tool for assessing bone health of indian children and youth. Endocrine. 2024.