Comprehensive Summary
This critical review evaluated how artificial intelligence and machine learning are being used to detect acute and occult scaphoid fractures on radiographs. The authors systematically reviewed nine retrospective studies, three of which were multicenter, conducted between 2020 and 2023. Sample sizes ranged from a few hundred to over eight thousand images, but all studies carried a high risk of bias. Model performance varied widely, with sensitivities between 0.72 and 0.97, specificities between 0.60 and 0.94, and overall accuracies up to 0.91. Models using transfer learning, such as DenseNet-121 and MobileNetV3, achieved the best results (AUCs around 0.81–0.88), though none of them consistently surpassed expert clinicians. The review found major weaknesses, including small test sets, inconsistent inclusion criteria, lack of MRI confirmation, and limited transparency in reporting. Overall, the evidence suggests that AI models hold promise for assisting fracture detection, particularly in challenging or occult cases, but they are not yet reliable enough for clinical use without larger, prospective, and MRI-validated studies
Outcomes and Implications
This review suggests that artificial intelligence could eventually aid clinicians in detecting scaphoid fractures—especially subtle or occult ones often missed on X-rays—helping reduce delayed diagnoses and unnecessary treatment. However, current evidence is too limited and inconsistent for clinical use. Small datasets, varied imaging standards, and unclear ground truths undermine reliability. To make AI tools clinically useful, future research must involve larger, multicenter, MRI-validated studies with transparent reporting and standardized evaluation. For now, AI should be seen as a supportive tool that complements, rather than replaces, clinician expertise