Highly Accurate Automated Patient-Specific 3D Bone Pose and Scale Estimation Using Bi-Planar Pose-Invariant Patches in a CNN-Based 3D/2D Registration Framework

Nahid Babazadeh Khameneh, Carlos Vázquez, Thierry Cresson, Frédéric Lavoie, Jacques A. de Guise

Abstract: This paper proposes an automatic CNN-based 3D/2D registration method to achieve highly accurate and robust seven degrees of freedom (7DOF) pose and isotropic scale of a generic 3D model. This step is a key enabler for reconstructing a patient-specific 3D bone surface model from a wide range of EOS® 2D bi-planar X-rays acquired with various fields of view and patients' orientations. Based on a coarse-to-fine strategy, first a CNN-based semantic segmentation followed by a PCA-based registration are used to roughly locate the bone. Similarity in pose-invariant local patches using CNN regression models is used to refine the 3D pose. The accuracy of the method is validated on 60 bi-planar X-rays. The mean of Mean Absolute pose Errors (MAE) of 3D translations, 3D rotations, and isotropic scaling are 0.19 mm, 0.33°, and 0.05 (%), respectively. The success rate is of 100 % at MAE lower than 1 mm, 1°, and 0.1 (%).