Published In(March 2023)
AuthorsLouis Lassalle, Nor-eddine Regnard, Jeanne Ventre, Vincent Marty, Lauryane Clovis, Zekun Zhang, Ali Guermazi, Jean-Denis Laredo
To assess the diagnostic performances of artificial intelligence (AI)-based software to perform automatic measurements on full-leg radiographs.
We retrospectively collected 393 consecutive full-leg radiographs including 50 EOS radiographs and 343 conventional radiographs from 4 imaging institutions. Two senior musculoskeletal radiologists independently annotated key points to measure the Hip-Knee-Ankle (HKA) angle, femur, tibial, and full-leg lengths, and pelvic obliquity. The gold standard was defined as the mean of their two measurements.
Statistical analysis consisted of mean absolute error (MAE), bias assessed with Bland-Altman analysis between the gold standard and the AI prediction and intraclass coefficient (ICC) between the two manual ratings.
153 full-leg conventional radiographs and 23 full-leg EOS radiographs were included, 217 radiographs were excluded. MAE for the HKA angle, femur, tibial, and full-leg lengths and pelvic obliquity were respectively 0.33° (95% CI: [0.29 ; 0.37], bias=0.28mm, ICC=0.99), 1.5mm (95% CI: [1.2 ; 1.9], bias=1.31mm, ICC=0.99), 1.5mm (95% CI: [1.2 ; 1.8], bias=0.33mm, ICC=0.99), 1.7mm (95% CI: [1.3 ; 2.1], bias=1.48mm, ICC=0.99),
0.7mm (95% CI: [0.6 ; 0.9], bias=-0.01mm, ICC=0.99). Bias and MAE between the gold standard and the AI prediction were low across all measurements. ICC between the two manual ratings was excellent across all measurements.
AI allows accurate and automatic anatomic measurements on full-leg conventional and EOS radiographs.
The study is retrospective with a small number of radiographs and no comparison to an independent manual rating.