Hoe goed detecteert AI fracturen van de cervicale wervelkolom op CT?*

Objective

To compare diagnostic accuracy of artificial intelligence (AI) for cervical spine (C-spine) fracture detection on CT with attending radiologists.

Design

Retrospective, diagnostic accuracy study.

Methods

AI analyzed 2368 scans from patients screened for C-spine fracture with CT (2007-2014, fracture prevalence 9.3%). With the use of a validated reference standard, which includes information on injuries in need of stabilizing therapy (IST), diagnostic accuracy of AI and radiologists was calculated and subsequently compared.

Results

Median age was 48 years. AI detected 158/221 fractures and radiologists 195/221, with a sensitivity of respectively 71.5% and 88.2% (p<0.001). Specificity of the AI and the radiologists was comparable: 98.6% and 99.2% (p=0.07). Of the fractures undetected by AI, 30/63 were an IST versus 4/26 for radiologists. AI detected 22/26 scans with fractures undetected by radiologists.

Conclusion

Compared to attending radiologists, AI has a lower sensitivity and misses more ISTs; however, it detected most fractures undetected by the radiologists, including ISTs.

Conflict of interest and financial support: potential conflicts of interest have been reported for this article. ICMJE forms provided by the authors are available online along with the full text of this article.