Negative Screening AI in Ankle Stress Radiography to Reduce Workload.

ABSTRACT

RATIONALE AND OBJECTIVES: Interpreting ankle stress radiographs is subjective and time-consuming. We aimed to train an AI model that efficiently screens negative cases, assess the agreement with expert with and without AI-assistance, and compare the workload reduction. MATERIAL AND METHODS: We collected anterior draw test (ADT) and talar tilt test (TTT) ankle stress radiographs from Seoul St. Mary's Hospital and St. Vincent's Hospital. Patients with prior surgery, severe joint fusion, or incomplete images were excluded. Expert measurements of tibio-talar distance (TTD) and tibio-talar angle (TTA) served as reference, defining positive labels as TTD ≥ 8.3 mm or TTA ≥ 6.2°. We trained a VGG16 model on data from hospital A and tested it on three separate test sets (testset1, 2 from St. Mary's Hospital, and testset3 from St. Vincent's Hospital). Three readers (expert, reader2, and the collective reading reports) evaluated the test sets, with and without AI-assistance (focusing only on AI-predicted positive cases). We measured agreement with the expert using Cohen's weighted Kappa and assessed the hypothetical workload reduction. RESULTS: AI-assistance did not significantly affect agreement with the expert for any reader in all test sets. Reader2 showed moderate-substantial agreement for all test sets, while collective reports reached fair agreement. The AI alone demonstrated fair to moderate agreement with the expert. AI-assistance reduced the hypothetical workload by 68.8-89.2% for ADT and 58.3-70.4% for TTT. CONCLUSION: We successfully trained an AI model for ankle stress radiography, achieving an average of 70% workload reduction while maintaining agreement with expert radiologists.