VIENNA – A commercially available AI tool for fracture detection performed comparably to radiologists in detecting most fractures, but fell short on the most clinically challenging injuries, according to a study presented on 6 March at ECR 2026.
Dr. Paul Botti of Geneva University Hospitals in Switzerland and colleagues tested how BoneView (Gleamer, Paris) stacked up against board-certified musculoskeletal radiologists when interpreting emergency x-rays for adult patients with foot/ankle trauma over six months.
“Our findings support the AI use as complementary help for radiologists in a busy workflow, with targeted attention to midfoot injuries where detection remains challenging," Botti told attendees in a packed session on AI in emergency imaging.
Dr. Paul Botti of Geneva University Hospitals in Switzerland, presented a study on 6 March at ECR 2026.The ESR and Dr. Paul Botti.
Trauma to the foot and ankle is a common cause of visits to the emergency department. Midfoot fractures, particularly those involving the Chopart and Lisfranc joints, are notoriously difficult to detect on plain x-rays and carry a high risk of being missed even by experienced clinicians, Botti explained.
Thus, the researchers sought to evaluate whether BoneView could approach radiologist-level performance in this setting, with particular attention to high-stakes injury patterns.
The retrospective single-center study included all emergency foot and ankle radiographs obtained over six months for adult trauma patients (n = 701). Radiographs were interpreted independently by radiologists in routine clinical workflow and by the AI tool in stand-alone mode, with each blinded to the other. The researchers calculated diagnostic performance metrics and interreader agreement for overall fractures, as well as for Chopart and Lisfranc injuries specifically. Conebeam CT and clinical follow-up served as the reference standard.
Of 701 studies analyzed, 319 fractures were identified (a prevalence of 45.5%), including 24 Chopart fractures (7.6% of all fractures), 22 Lisfranc fractures (6.8%), and 273 involving other bony structures (85.6%).
For overall fracture detection, AI achieved a sensitivity of 74.3%, a specificity of 83%, and an accuracy of 79%, compared with the radiologists' sensitivity of 84%, specificity of 95.5%, and accuracy of 90.3%.
For Chopart fractures, AI’s sensitivity was 62.1%, versus 82.8% for radiologists, with similarly high specificity in both groups and no statistically significant difference (p = 0.18). For Lisfranc fractures, AI sensitivity was 65.4%, compared with 80.8% for radiologists, again without reaching statistical significance (p = 0.453), though the gap remained clinically notable, Botti noted.
"Radiologists and AI achieve high diagnostic performance for fracture detection on foot and ankle radiographs, with high agreement. Radiologists showed a better performance in detecting Lisfranc/Chopart fractures than AI," Botti said.
The study was limited by its retrospective, single-center design and the relatively small number of Chopart and Lisfranc cases, but warrants further investigation, he concluded.
Our full coverage of ECR 2026 can be found here.
