Understanding the bone kinematics of the human knee during dynamic motions is necessary to evaluate the pathological conditions, design knee prosthesis, orthosis and surgical treatments such as knee arthroplasty. Also, knee bone kinematics is essential to assess the biofidelity of the computational models.
A new methodology to reconstruct the 3D shape of the knee bones from multiple uncalibrated fluroscopic images using deep learning is developped. During the image acquisition using the SF, the subjects in this approach can freely rotate their leg (in a fully extended, knee-locked position), resulting in several images captured in arbitrary poses.
Relevant features are extracted from these images using a novel feature extraction technique before feeding it to a custom-built Convolutional Neural Network (CNN). The network, without further optimization, directly outputs a meshed 3D surface model of the subject's knee joint. The whole procedure could be completed in a few minutes.
The robust feature extraction technique can effectively extract relevant information from a range of image qualities. When tested on eight unseen sets of SF images with known true geometry, the network reconstructed knee shape models with a shape error (RMSE) of 1.91± 0.30 mm for the femur, 2.3± 0.36 mm for the tibia and 3.3± 0.53 mm for the patella.
The error was calculated after rigidly aligning (scale, rotation, and translation) each of the reconstructed shape models with the corresponding known true geometry (obtained through MRI segmentation). Based on a previous study that examined the influence of reconstructed shape accuracy on the precision of the evaluation of tibiofemoral kinematics, the shape accuracy of the proposed methodology might be adequate to precisely track the bone kinematics, although further investigation is required.
