Purpose or Case Report: This review systematically examines diagnostic failure modes of artificial intelligence (AI) systems in pediatric imaging, analyzes root causes and clinical impact, and proposes strategies for safer deployment in pediatric radiology. Pediatric imaging poses unique challenges due to developmental anatomy, age-specific disease patterns, and technical variability that differ markedly from those in adult populations.
Methods & Materials: A comprehensive literature review was conducted examining AI performance in pediatric radiology across multiple subspecialties and imaging modalities. Clinical cases demonstrating AI errors were analyzed and categorized by failure type: developmental anatomy misinterpretation, diagnostic ambiguity/mimics, motion artifacts, technologist variability, modality-specific errors, artifacts/confounders, and dataset limitations. The review synthesized evidence from published studies, clinical cases, and clinical implementations of AI systems in pediatric imaging workflows.
Results: Six major categories of AI failure modes were identified: (1) Misclassification of normal developmental variants as pathology, particularly around growth plates and age-specific anatomical features; (2) Inability to distinguish between pathological mimics with similar radiographic appearances; (3) Vulnerability to motion artifacts and suboptimal positioning common in unsedated pediatric examinations; (4) Modality-specific technical limitations affecting AI performance; (5) Misinterpretation due to external artifacts, medical devices, and confounding visual elements; (6) Systematic biases from limited pediatric training data and annotation errors. These failures resulted in both false-positive and false-negative diagnoses, with potential consequences including unnecessary interventions, delayed care, and increased radiation exposure from repeat imaging.
Conclusions: Current AI systems demonstrate significant vulnerabilities primarily due to training on adult-dominant datasets and lack of pediatric-specific validation. Safe integration requires: building a pediatric-specific datasets, implementing context-aware model adaptation, prioritizing algorithm transparency, establishing feedback loops for continuous improvement, optimizing imaging environments and technologist training, ensuring access to pediatric sedation services, and creating centralized error registries. These interventions are essential for developing AI systems that can effectively address the challenges of pediatric radiology