.png){kind=icon}
.png){kind=icon}
Return
The AI implementation gap in trauma radiography: standalone versus discretionary AI-integrated fracture detection
.png){kind=icon}
OA
AI
DOI：10.1186/s41747-026-00763-6.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In emergency trauma care, artificial intelligence (AI) may aid fracture detection on radiographs, potentially reducing radiologists’ workload. We evaluated the role of deep learning-based decision-support software in the reporting of trauma cases. We retrospectively analyzed 2317 trauma radiographs acquired at a single center: 1,174 images obtained from November 1 to 16, 2023, without access to the AI tool during reporting, and 1,143 images from February 1 to 13, 2024, with discretionary use of the AI output during reporting. The AI software output was compared with final radiology reports, with ground truth established by a musculoskeletal radiologist with 9 years’ experience. Accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated at both the fracture and patient levels. The dataset included 1,914 patients with 1,188 acute fractures (621 in November, 567 in February). At the fracture level, standalone AI achieved 90.7% accuracy, 87.8% sensitivity, 94.0% specificity, 94.3% PPV, and 87.2% NPV in November, 94.1%, 93.5%, 94.6%, 94.5%, and 93.6% in February, respectively. Non-AI-assisted radiologists reached 92.4%, 89.0%, 96.2%, 96.3%, and 88.7%, AI-assisted radiologists 93.4%, 90.0%, 96.7%, 96.4%, and 90.7%, respectively. At the patient level, AI’s overall performance reached up to 96.5% accuracy and 95.6% sensitivity. Discrepancies between AI and radiologists occurred in 326 cases, often related to anatomical variants such as accessory ossicles. Standalone AI demonstrated near-expert accuracy and sensitivity in fracture detection at both fracture and patient levels. PPV increased with AI support, indicating more accurate detection of actual fractures. By examining discretionary real-world use of AI in trauma radiography, this study shows that clinical benefit is not guaranteed by algorithmic performance alone, as optional AI integration does not consistently improve radiologist sensitivity, underscoring a critical implementation gap in practice.
Keywords:
Artificial intelligence
Deep learning
Emergency medicine
Fractures (bone)
Radiography
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
E
IF:
3.6
Papers:
651
Citations:
1.7K