Evaluation of a convolutional neural network to identify scaphoid fractures on radiographs

Abstract

This study aimed to develop and evaluate a convolutional neural network for identifying scaphoid fractures on radiographs. A dataset of 1918 wrist radiographs (600 patients) was taken from an orthopaedic referral centre between 2010 to 2020. A YOLOv3 and a MobileNetV3 convolutional neural network were trained for scaphoid detection and fracture classification, respectively. The diagnostic performance of the convolutional neural network was compared with the majority decision of four hand surgeons. The convolutional neural network achieved a sensitivity of 82% and specificity of 94%, with an area under the receiver operating characteristic of 92%, whereas the surgeons achieved a sensitivity of 76% and specificity of 96%. The comparison indicated that the convolutional neural network’s performance was similar to the majority vote of surgeons. It further revealed that convolutional neural network could be used in identifying scaphoid fractures on radiographs reliably, and has potential to achieve the expert-level performance. Level of evidence: III

Exploring the role of artificial intelligence chatbot on the management of scaphoid fractures

“…Recent studies have demonstrated the potential of a subset of AI convolutional neural networks in the field of radiology. For example, one study showed that convolutional neural networks had a sensitivity of 82% and specificity of 94% for identifying scaphoid fracture on radiographs, whereas surgeons achieved sensitivity of 76% and specificity of 96%, superior of emergency physicians (Li et al., 2022; Ozkaya et al., 2022). Conversely, a similar study found that convolutional neural networks had a high number of false positives, suggesting that further research is needed in developing such algorithm (Langerhuizen et al., 2020).…”

What’s New in Hand Surgery

“…There has been focus during the preceding year on technology to assist with this. Li et al applied a convolutional neural network trained to identify scaphoid fractures to 1,918 wrist radiographs and showed sensitivity and specificity comparable with those of surgeons 34 . Another trial randomized 36 patients with acute, nondisplaced scaphoid fractures to robotic-assisted or freehand placement of a guidewire 35 .…”

Insights and trends review: artificial intelligence in hand surgery

“…Such interventions could reduce the burden of emergency reporting in cases of binary decisions (fracture present or not present), reducing missed fractures and streamlining referral to orthopaedic clinics or automating further radiological investigations. AI in hand/wrist radiology has focused on using neural networks for scaphoid fracture detection (Hendrix et al., 2021; Li et al., 2022; Yoon and Chung, 2021; Yoon et al., 2021) and to a lesser extent, hand fractures (Üreten et al., 2022) with promising results. While they are not out-performing humans (Kuo et al., 2022; Olczak et al., 2017; Yoon et al., 2021), these studies are setting a benchmark for future research.…”