Abstract
Purpose:
To materialize 3D kinematic analysis of total knee arthroplasty (TKA), 2D/3D registration techniques, which use X-ray fluoroscopic images and the knee implants CAD, have been applied to clinical cases. However, most conventional methods have needed time-consuming and labor-intensive manual operations in some process. In previous study, we addressed a manual operations problem when setting initial pose of implants model for 2D/3D registration, and reported a semi-automated initial pose estimation method based on an interpolation technique. However, this method still required appropriate initial pose estimation of the model with manual operations for some X-ray images (key frames). Additionally, in the situation like fast knee motion and use of low frame rate, good registration results were not obtained because of the large displacement between each frame silhouette. To overcome these problems, this study proposes an improved semi-automated 3D kinematic estimation method.
Methods:
Our 2D/3D registration technique is based on a robust feature-based algorithm. In improved initial pose estimation method, for the only first frame, the initial pose is manually adjusted as close as possible. That is, we automatically estimate appropriate initial pose of the model for X-ray images except for the first frame.
To automatically estimate the initial pose of the model, we utilize a transformation with feature points extracted from the previous and next frames. A transform matrix which has three DOF (translations parallel to the image, and a rotation perpendicular to the image) is calculated by registration of corresponding feature points between the previous and next frame extracted with SURF algorithm. While, the corresponding point sets extracted by SURF sometimes include some error sets. Therefore, in this study, LmedS method was employed to detect the error corresponding sets and calculate a transform matrix accurately. In Fig. 1(a) and (b), the orange square shows the region defined with the boundary box of the model, and some lines show the combined corresponding point sets. The blue lines are correct corresponding point sets, and the pink lines are error corresponding point sets detected with LmedS method.
Finally, 3D pose of the model estimated in previous frame is transformed with accurately calculated transform matrix, and the transformed pose is used as an initial 3D pose of the model in next frame.
Experimental results:
To validate the feasibility of the improved semi-automated 3D kinematic estimation method, experiments using X-ray fluoroscopic images of 4 TKA patients during knee motions were performed. In order to assess the performance of the improved method, automation rate was calculated, and the rate was defined as the X-ray frame number of satisfying clinical required accuracy (error within 1 mm, 1 degree) relative to all X-ray frame number.
As results of the experiments, 3D pose of the model for all X-ray images except for the first frame is automatically stably-estimated, the automation rate of the femoral and tibial component were 83.7% and 73.5%, respectively.
Conclusions:
The present method doesn't need labor-intensive manual operations for 3D kinematic estimation of TKA, and is thought to be very helpful for practical clinical applications.
