Significantly better precision with new machine learning versus manual image registration software in processing images from single-plane fluoroscopy to determine tibiofemoral kinematics following total knee replacement

One common method to determine tibiofemoral kinematics following total knee replacement (TKR) is to capture single-plane fluoroscopic images of a patient activity and determine anterior-posterior (AP) positions of the femoral condyles and internal-external (IE) tibial rotation. Although JointTrack is widely used to analyze such images, precision (i.e. repeatability) in determining AP positions and IE tibial rotations using the two publicly available programs has never been quantified. The objectives were to determine the precision and reproducibility of results using both programs. Fluoroscopic images of 16 patients who performed a weight-bearing deep knee bend following TKR were analyzed. JointTrack Manual (JTM) and JointTrack Machine Learning (JTML) were used to perform 3D model-to-2D image registration after which AP positions of the femoral condyles and IE tibial rotations were determined. Precision in AP positions and IE rotations was quantified. Intraclass correlation coefficients (ICCs) for both repeatability (i.e. intraobserver) and reproducibility (i.e. interobserver) also were determined. Precision using JTM was worse than JTML for AP positions of the medial and lateral femoral condyles (1.0 mm and 0.9 mm vs 0.3 mm and 0.4 mm, respectively; p < 0.001 for both). For IE tibial rotation, precision also was worse using JTM versus JTML (1.1º vs 0.9°, p = 0.010). ICC values for JTML indicated good to excellent agreement (range: 0.82–0.98) whereas ICC values for JTM indicated only moderate to good agreement (range: 0.58–0.88). JTML has better precision and reproducibility than JTM and also is more efficient to use. Therefore, use of JTML over JTM is strongly recommended.