gms | German Medical Science

Objectives: The measurement of the implant alignement on long leg radiographs (LLR) after total knee arthroplasy (TKA) is time consuming and dependent on the expertise of the observer. Artificial-Intelligence (AI) technology may be used to automate and standardize this process and would also allow the analysis of large datasets. In this study we evaluated the reliability of a newly developed AI-algorithm for TKA alignment assessment and testedthe software on a large dataset.

Methods: The AI software was trained on over 15,000 radiographs from the OAI (Osteoarthritis Initiative study; US six-site multi-center), MOST (Multicenter Osteoarthritis Study, US two-site multi-center), CHECK (Cohort Hip and Cohort Knee study; Netherland single center) studies as well as five sites in Austria. For the validation cohort 200 calibrated LLRs of eight different common unconstrained and constrained knee systems were analyzed. The AI-algorithm-measurements from LAMA Version 1.03 (ImageBiopsy Lab) were compared to manual reads. The hip knee ankle (HKA) as well as femoral (FCA) and tibial component (TCA) were measured by two experienced orthopedic surgeons. An evaluation cohort consisting of all institutional LLRs with TKAs of 2018 (n=1312) were evaluated to assess the algorithms' ability of handling large datasets.

Figure 1 [Fig. 1]

Descriptive statistic was employed, including mean (M), standard deviation (SD), statistical mean absolute deviation about the mean (sMAD), and percentage. For the validation we allocated the measured results into reader-1, reader-2, mean of measures from reader-1 and -2 (=ground truth), and AI software measurements. Differences between measurements were assessed by perfoming an analysis of variance and a post-hoc test. Intraclass correlation (ICC) coefficients were calculated to describe conformity between the AI software and our manual reads, as well as between the two readers.

Results: Validation cohort: Reproducibility was 96% and reliability was 92.1% for the use of the AI-software on LLRs with an output and was excellently reliable on all measured angles (ICC>0.97). Excellent results were found for primary unconstrained TKAs. In constrained TKAs landmark setting on the femoral and tibial component failed in 12.5% of LLRs (n=9). Evaluation cohort: The mean overall HKA, FCA and TCA values were 0.3° ± 2.9°, 89.4° ± 2.0° and 89.1° ± 1.8° respectively. The mean HKA angle of postoperative LLRs was 0.2° varus ± 2.5° (n=1240) and the HKA of pre-revision LLRs was 1.6 varus ± 6.4° (n=74).

Conclusion: Lower limb alignement in LLRs with TKAs can be reliably analyzed by AI-algorithms. AI-solutions allow the convenient handling of large datasets.