Effect of artificial intelligence assistance on skull radiograph diagnosis of craniosynostosis

Diagnosing craniosynostosis from skull radiographs is challenging, often leading to diagnostic delays or overreliance on computed tomography (CT). While ultrasound and MRI reduce radiation exposure, their need for specialized expertise and equipment limits routine triage adoption. In a multireader, multicase paired study, 23 clinicians interpreted 59 independent skull radiograph cases unassisted and subsequently with AI assistance. Using a DBM jackknife multireader multicase analysis that accounted for both reader and case variability, AI assistance significantly improved diagnostic accuracy from 75.0% to 84.7% (difference, 0.097; 95% CI, 0.053–0.141; p = 0.00015), bridging the performance gap between human readers and a standalone AI model (cross-validation mean AUROC 0.975). Improvements were most pronounced among neurosurgery residents, suggesting that AI may reduce diagnostic variability across experience levels. Furthermore, AI markedly improved the accuracy of high-confidence decisions. These findings indicate that AI assistance improves radiograph-based classification accuracy across experience levels. Whether this benefit translates into fewer misdiagnoses, more appropriate referral, or reduced CT utilization requires prospective evaluation.