Artificial intelligence-assisted detection of fractures on radiographs with BoneView: a systematic review

To systematically review the added value of the artificial intelligence tool BoneView in detecting fractures on radiographs.

Medline and Embase were searched for original studies that reported the diagnostic performance of human reading in detecting fractures on radiographs with and without BoneView. Study quality was assessed. Diagnostic accuracy data and reading speed were extracted.

Eight studies were included. There was high risk of bias with respect to patient selection (5 studies), reference standard (1 study), and flow and timing (3 studies). There was high concern regarding the applicability of the execution of the index test in one study. Sensitivities and specificities were heterogeneous (p ≤ 0.0001). Sensitivity was significantly higher (p < 0.05) among the far majority of the readers in the included studies when radiographs were evaluated with BoneView. Specificities and diagnostic odds ratio results were mixed, with either no significant change or significant increase or decrease. Four studies assessed reporting time. In 3 studies, reading speed was faster with BoneView (mean of 5.3-15.7 s, p ≤ 0.046), whereas in one study there was no change (p = 0.12).

BoneView appears to improve sensitivity, whereas the results regarding specificity and overall diagnostic accuracy are mixed. There are methodological quality concerns in the existing literature and further research is needed to explore causes of heterogeneity. The use of BoneView appears not to compromise reading speed and may even improve it.