Diagnostic Accuracy of Artificial Intelligence Versus Musculoskeletal Radiologists for Foot and Ankle Fracture Detection on Radiographs

Abstract

Background/Objectives: To evaluate the diagnostic accuracy of a standalone artificial intelligence fracture detection software on foot and ankle radiographs, compared with board-certified musculoskeletal radiologists, with emphasis on midfoot injuries. Materials and Methods: This retrospective single-center diagnostic accuracy study included adult patients presenting to the emergency department with acute foot or ankle trauma. Radiographs were interpreted in real time, analyzed retrospectively by musculoskeletal (MSK) radiologists during the routine workflow and subsequently analyzed independently by an artificial intelligence system. The reference standard was structured clinical follow-up, complemented by cross-sectional imaging when the radiograph was equivocal. Diagnostic performance metrics and inter-reader agreement were calculated. Results: A total of 701 examinations (mean age, 42 ± 17 years; 376 men) were included; 319 fractures were identified, including 29 Chopart and 26 Lisfranc injuries. Overall sensitivity, specificity, and accuracy were 74.3%, 83.0%, and 79.0% for artificial intelligence, 84.0%, 95.5%, and 90.3% for radiologists, respectively (p < 0.0001; κ = 0.65). For Chopart and Lisfranc fractures, radiologists demonstrated higher sensitivity, while specificity remained excellent for both approaches, but the paired comparisons did not reach significance. Conclusions: Standalone artificial intelligence achieved high diagnostic performance for foot and ankle fracture detection on radiographs but was less sensitive than MSK radiologists for complex midfoot injuries.