AI Fracture Detection on X-Rays: What the Evidence Shows, from Wrist to Mandible

By the BossCE Editorial Team.

Key Points

• Across 42 studies, AI fracture detection performed on a par with clinicians: pooled sensitivity 92% for AI and 91% for clinicians, specificity 91% and 92% [1]. Only one study tested AI prospectively in a real clinical setting [1].
• When 24 doctors from six specialties read 480 trauma radiographs with AI assistance, per-patient sensitivity rose by 10.4% with no loss of specificity, and reading was 6.3 seconds faster per patient [2].
• NICE allows four AI fracture-detection tools in UK urgent care, but only as an addition to standard care and only while further evidence is collected [3].
• For mandibular fractures on panoramic radiographs, the evidence is still preliminary: five small studies, none fully at low risk of bias [4].

Why a Second Look Matters

Missed fractures are a common reason for disagreement between the first reading of a radiograph and the final report by a radiologist [2]. In emergency departments and urgent care, the first reader is often an emergency doctor, a nurse practitioner or a junior colleague, working at speed. Fracture-detection AI is designed for that moment: it marks suspected fractures on the image so the reader looks again.

What the Evidence Shows

AI alone versus clinicians. A systematic review and meta-analysis in Radiology pooled 42 studies, 37 on radiographs and five on CT [1]. On internal test sets, pooled sensitivity was 92% for AI and 91% for clinicians, and pooled specificity was 91% and 92% [1]. The authors described the field as early and largely pre-clinical: few studies assessed clinicians working with AI, and only one evaluated AI prospectively in practice [1].

Clinicians with AI versus without. In a multi-reader study of 480 examinations across eight body regions, 24 readers (radiologists, orthopaedic surgeons, emergency physicians, physician assistants, rheumatologists and family physicians) read every case with and without AI, at least a month apart [2]. With AI, sensitivity per patient was 10.4% higher, specificity did not fall, and reading time dropped by 6.3 seconds per patient [2]. The study design set fracture prevalence at 50%, higher than in a typical emergency department, so real-world gains may differ.

How NICE Frames It

NICE’s early value assessment allows four named AI technologies to be used in urgent care while evidence is generated, two of them for any age and two for patients aged 2 years and over; a fifth is recommended for research only [3]. The conditions matter as much as the decision:

• An addition, not a replacement: AI is used alongside standard care, and healthcare professionals make the treatment decisions [3].
• Read first, then check the AI: NICE advises clinicians to interpret the X-ray before viewing the AI result, so the tool does not anchor their judgement [3].
• Plan for disagreement: local protocols should say what happens when the clinician and the AI disagree [3].

The Dental and Maxillofacial Angle

Facial and mandibular fractures often present first to emergency teams, and dentists may see them on a panoramic radiograph. A 2024 systematic review and meta-analysis found five studies of deep learning for mandibular fracture detection on panoramic radiographs, with a pooled specificity of about 81% [4]. No included study fully met the QUADAS-2 risk-of-bias criteria, and the authors concluded that the evidence is preliminary, limited by small and narrow datasets [4]. For now, these tools are research findings, not a substitute for clinical examination and appropriate imaging.

Practical Safeguards If Your Service Uses AI Triage

• Read the image yourself first, then compare with the AI marks [3].
• Treat a flag as a prompt to look again, not as a report. A clear AI result does not rule out a fracture when the history and examination suggest one.
• Check what the tool is approved for, including age range and body region [3].
• Record your own interpretation in the notes, and follow the local protocol when you and the AI disagree [3].
• Audit missed and over-called fractures locally. Published accuracy may not match your case mix [1].

Using BossCE for Imaging Questions

BossCE AI searches guidelines and published evidence for each question and numbers its sources under the answer, so you can check every statement. Remove all patient identifiers before uploading any image or report.

Try it: ask BossCE AI about AI fracture detection in emergency care, or browse the reviewed guidelines in the BossCE Library.

References

1. Kuo RYL, Harrison C, Curran TA, Jones B, Freethy A, Cussons D, Stewart M, Collins GS, Furniss D. Artificial intelligence in fracture detection: a systematic review and meta-analysis. Radiology. 2022;304(1):50-62. doi:10.1148/radiol.211785
2. Guermazi A, Tannoury C, Kompel AJ, et al. Improving radiographic fracture recognition performance and efficiency using artificial intelligence. Radiology. 2022;302(3):627-636. doi:10.1148/radiol.210937
3. National Institute for Health and Care Excellence. Artificial intelligence (AI) technologies to help detect fractures on X-rays in urgent care: early value assessment. HealthTech guidance HTG739 (formerly HTE20). nice.org.uk/guidance/htg739
4. Dashti M, Ghaedsharaf S, Ghasemi S, Zare N, Constantin EF, Fahimipour A, Tajbakhsh N, Ghadimi N. Evaluation of deep learning and convolutional neural network algorithms for mandibular fracture detection using radiographic images: a systematic review and meta-analysis. Imaging Science in Dentistry. 2024;54(3):232-239. doi:10.5624/isd.20240038

This article is for qualified healthcare professionals and is not a substitute for clinical judgement or the full guideline. Check the current version of each source before applying it.