Page 13 - Read Online

Dababneh et al. Art Int Surg 2024;4:214-32 https://dx.doi.org/10.20517/ais.2024.50 Page 220

reduced the need for additional imaging by 36% without overlooking any fractures.

Occult fracture detection
Langerhuizen et al. advanced research in this field by exploring AI’s potential in detecting occult scaphoid
[47]
fractures, aiming to enhance radiologists’ detection capabilities . The single-step pretrained CNN used in
this study achieved an AUC of 0.77, performing similarly to orthopedic surgeons but with a lower
specificity. The model was also able to detect five out of six occult scaphoid fractures that were missed by
human experts, but it struggled with detecting fractures that were obvious to human observers. Yoon et al.
[48]
also explored the detection of occult scaphoid fractures by developing a three-step model . The first CNN
was tasked with segmentation, while the second model was used to detect scaphoid fracture. As for the
third, its role was to analyze the cases that were considered negative by the previous AI, aiming to detect
overlooked fractures. This CNN model successfully detected 90.9% of occult fractures, correctly identifying
20 out of 22 cases.

Raisuddin et al. published a study focusing on the detection of occult distal radius fractures, requiring CT
imaging for detection. In this study, the authors developed a DL model, Deep Wrist, and evaluated its
[49]
performance in challenging cases . To validate the model’s efficacy, it was initially tested on a general
population test set, where it achieved a diagnostic accuracy of 99% and an AUC of 0.99. However, when
tested on the occult fracture dataset, the model’s accuracy dropped to 64%. The model performed slightly
better when both lateral and AP views were used together, compared to using the lateral view alone.

Reducing X-ray projections
Building upon the established efficacy of AI in detecting fractures on radiographs, Janisch et al. explored
CNN’s potential to reduce the standard X-ray requirements for diagnosing torus fractures of the distal
[50]
radius . Currently, common practice often requires at least two complementary projections, AP and lateral
views of the wrist. This traditional approach, while thorough, results in increased radiation exposure and
patient discomfort. Three CNNs were trained on a pediatric dataset and achieved AUCs ranging from 0.945
to 0.980. EfficientNet-B4 emerged as the most accurate, outperforming radiologists and pediatric surgeons.

Applications in pediatric
Zech et al. have made significant contributions to the application of DL models in pediatric fracture
detection, publishing three articles. The first study analyzed the performance of an open-source AI
algorithm in the detection of pediatric upper extremity fractures based on 53,896 radiographs . In this
[51]
study, attendings’ accuracy in detecting fractures improved slightly with AI. In contrast, radiology and
pediatric residents showed significant improvement with AI. AI showed its superiority especially in
identifying non-obvious fractures (non-displaced or non-angulated). The second study focused on wrist
injuries. The Faster R-CNN model accurately identified distal radius fractures with an AUC of 0.92 .
[52]
Additionally, the use of AI significantly improved residents’ fracture detection rates from 69% to 92%. To
further enhance the model’s performance, Ilie et al. conducted a subsequent, more comprehensive study
utilizing a database of 58,846 upper extremity fracture images . This successfully improved the model’s
[53]
diagnostic accuracy across various fracture types and anatomical regions, notably by an increase of the AUC
to 0.96.

A compilation of published articles focusing on AI-driven fracture detection in the hand or wrist (including
the distal radius and ulna, carpal bones, and fingers) is presented in Table 2.