Page 12 - Read Online

Page 219 Dababneh et al. Art Int Surg 2024;4:214-32 https://dx.doi.org/10.20517/ais.2024.50

compared to lateral views.

As for Mert et al., they evaluated ChatGPT 4’s capability in detecting distal radius fractures through
radiological images, comparing it with human specialists (a hand surgery resident, a medical student) and
TM [40]
an AI system (Gleamer BoneView ) . The results indicate that ChatGPT 4 demonstrates good diagnostic
accuracy, significantly outperforming the medical student, but was outperformed by both the hand surgery
TM
resident and Gleamer BoneView .

Scaphoid fracture
Our review identified 11 articles focusing on the role of AI in detecting scaphoid fractures, with two
specially examining its potential in identifying occult fractures.

In 2020, Ozkaya et al. conducted a study to evaluate a CNN model’s ability to detect scaphoid fractures
[41]
using AP wrist radiographs . The CNN model achieved an AUC of 0.840, performing comparably to the
less experienced orthopedic surgeon, and it surpassed the ED physician (0.760 AUC), but was outperformed
by the expert hand surgeon (0.920 AUC).

Similarly, Hendrix et al. compared a self-developed CNN model’s performance in detecting scaphoid
[42]
fractures on AP and PA hand radiographs to that of 11 radiologists . The segmentation CNN achieved a
Dice coefficient of 0.974 while the fracture detection CNN achieved an AUC of 0.87, performing
comparably to the radiologists. It is important to note that in this study, radiologists were limited to a single
view for fracture detection, whereas multiple views are generally used in clinical settings.

In a follow-up study, Hendrix et al. expanded on their previous work by using a larger dataset to train the
[43]
same CNN model and compared its performance to five MSK radiologists . In this study, the CNN model
achieved an AUC of 0.88, slightly outperforming the radiologists. The inclusion of ulnar-deviated and
oblique views enhanced the model’s accuracy. The results also demonstrated that the CNN model reduced
the reading time for four out of five radiologists by 49.4%. Nevertheless, it was noted that AI integration did
not significantly improve most radiologists’ diagnostic accuracy.

In 2021, Tung et al. also published on this topic by comparing multiple CNN models in detecting scaphoid
[44]
fractures . Among the models without additional transfer learning training, DN121 had the highest AUC
with 0.810, while VGG16 demonstrated the highest precision with 100% accuracy and 1.00 specificity. After
the application of transfer learning, RN101 achieved the highest AUC with 0.950. Yang et al. also proposed a
[45]
combination of two CNN models for scaphoid area segmentation and fracture detection . The Faster R-
CNN was used to identify the fracture region, followed by ResNet to detect the presence of fracture. The
study utilized a dataset of scaphoid radiographs, which included 31 images of occult fractures. The proposed
algorithm achieved an AUC of 0.917 for scaphoid fracture detection.

Scaphoid fracture prediction
Bulstra et al. trained five ML models to calculate scaphoid fracture probability using clinical and
demographic features such as mechanism of injury, sex, age, affected side, and examination maneuvers .
[46]
All models achieved an AUC above 0.72, with Boosted decision tree outperforming the others with an AUC
of 0.77. Pain over the scaphoid on ulnar deviation and male sex were the predictors with the highest
correlation to scaphoid fractures. Using these features, an algorithm was developed suggesting that patients
with radial-sided wrist pain, negative radiographs, and a fracture probability of 10% or more should
undergo further imaging. When applied to the study’s patients, the algorithm achieved 100% sensitivity and