Foundation models in radiology: a primer for pediatric radiologists

Foundation models (FMs) are large deep learning models pre-trained on vast heterogeneous datasets through self-supervised learning that are adaptable to diverse downstream tasks with minimal fine-tuning. In pediatric radiology, where data scarcity, rare pathologies, and anatomical variability present significant hurdles, FMs offer a robust mechanism for broad feature learning. Complementary to traditional machine learning pipelines, FMs serve as flexible backbones that can be adapted to specific clinical needs through established techniques such as transfer learning and parameter-efficient fine-tuning. These models can facilitate multiple tasks, such as pathology detection and classification, lesion segmentation, report generation, and visual question answering, with the potential to improve diagnostic accuracy, workflow efficiency, and decision support in pediatric care. However, FMs' implementation in pediatric imaging faces key challenges, including pediatric unique disease spectra and anatomical variability, limited datasets, ethical and privacy issues, and the absence of pediatric-specific validation. Broader limitations include hallucinations, lack of model explainability, resource disparities, and risks of radiologists' deskilling. Future perspectives to overcome these barriers are represented by techniques such as federated and continual learning, and synthetic data generation. Our review introduces the principles and architectures of FMs, presents the current and emerging applications in pediatric radiology, and offers an overview of the challenges and future directions for FMs' safe, equitable, and effective integration into clinical practice. FMs are a promising frontier for transforming pediatric imaging and advancing child-centered healthcare.