Automatic generation of knee kinematic models from medical imaging

Abstract

Background and ObjectiveThree-dimensional spatial mechanisms have been used to accurately predict passive knee kinematics, and have shown potential to be used in optimized multibody kinematic models. Such multi-body models are anatomically consistent and can accurately predict passive knee kinematics, but require extensive medical image processing and thus are not widely adopted. This study aimed to automate the generation of kinematic models of tibiofemoral (TFJ) and patellofemoral (PFJ) joints from segmented magnetic resonance imaging (MRI) and compare them against a corresponding manual pipeline. MethodsFrom segmented MRI of eight healthy participants (four females; aged 14.0 ± 2.6 years), geometric parameters (i.e., articular surfaces, ligament attachments) were determined both automatically and manually, and then assembled into TFJ and PFJ kinematic models to predict passive kinematics. The TFJ model was a six-link mechanism with deformable ligamentous constraints, whereas PFJ was a modified hinge. The ligament length changes through TFJ flexion were prescribed to literature strain profile. The geometric parameters were optimized to ensure physiological kinematic predictions through a Multiple Objective Particle Swarm Optimization. ResultsGeometric parameters showed strong agreement between automatic and manual pipelines (median error of 2.8 mm for anatomical landmarks and 1.5 mm for ligament lengths). Predicted TFJ and PFJ kinematics from the two pipelines were not statistically different, except for tibial superior/inferior translation near terminal TFJ extension. The TFJ kinematics predicted from the automatic pipeline had mean errors of 3.6° and 12.4° for adduction/abduction and internal/external rotation, respectively, and <7 mm mean translational error compared to the manual pipeline. Predicted PFJ had <9° mean rotational errors and <6 mm mean translational errors. ConclusionsThe automatic pipeline developed and presented here can predict passive knee kinematics comparable to a manual pipeline, but removes laborious manual processing and provides a systematic approach to model creation.