Accuracy of radiologists and radiology residents in detection of paediatric appendicular fractures with and without artificial intelligence.

Abstract

Objectives: We aim to evaluate the accuracy of radiologists and radiology residents in the detection of paediatric appendicular fractures with and without the help of a commercially available fracture detection artificial intelligence (AI) solution in the hopes of showing potential clinical benefits in a general hospital setting.

Methods: This was a retrospective study involving three associate consultants (AC) and three senior residents (SR) in radiology, who acted as readers. One reader from each human group interpreted the radiographs with the aid of AI. Cases were categorised into concordant and discordant cases between each interpreting group. Discordant cases were further evaluated by three independent subspecialty radiology consultants to determine the final diagnosis. A total of 500 anonymised paediatric patient cases (aged 2-15 years) who presented to a tertiary general hospital with a Children's emergency were retrospectively collected. Main outcome measures include the presence of fracture, accuracy of readers with and without AI, and total time taken to interpret the radiographs.

Results: The AI solution alone showed the highest accuracy (area under the receiver operating characteristic curve 0.97; AC: 95% CI -0.055 to 0.320, p=0; SR: 95% CI 0.244 to 0.598, p=0). The two readers aided with AI had higher area under curves compared with readers without AI support (AC: 95% CI -0.303 to 0.465, p=0; SR: 95% CI -0.154 to 0.331, p=0). These differences were statistically significant.

Conclusion: Our study demonstrates excellent results in the detection of paediatric appendicular fractures using a commercially available AI solution. There is potential for the AI solution to function autonomously.