Statistical Shape Model-Based Reconstruction Of A Scaled, Patient-Specific Surface Model Of The Pelvis From A Single Standard Ap X-Ray Radiograph

Guoyan Zheng

DOI: https://doi.org/10.1118/1.3327453

IF: 4.506

2010-01-01

Medical Physics

Abstract:Purpose: The aim of this article is to investigate the feasibility of using a statistical shape model (SSM)-based reconstruction technique to derive a scaled, patient-specific surface model of the pelvis from a single standard anteroposterior (AP) x-ray radiograph and the feasibility of estimating the scale of the reconstructed surface model by performing a surface-based 3D/3D matching.Methods: Data sets of 14 pelvises (one plastic bone, 12 cadavers, and one patient) were used to validate the single-image based reconstruction technique. This reconstruction technique is based on a hybrid 2D/3D deformable registration process combining a landmark-to-ray registration with a SSM-based 2D/3D reconstruction. The landmark-to-ray registration was used to find an initial scale and an initial rigid transformation between the x-ray image and the SSM. The estimated scale and rigid transformation were used to initialize the SSM-based 2D/3D reconstruction. The optimal reconstruction was then achieved in three stages by iteratively matching the projections of the apparent contours extracted from a 3D model derived from the SSM to the image contours extracted from the x-ray radiograph: Iterative affine registration, statistical instantiation, and iterative regularized shape deformation. The image contours are first detected by using a semiautomatic segmentation tool based on the Livewire algorithm and then approximated by a set of sparse dominant points that are adaptively sampled from the detected contours. The unknown scales of the reconstructed models were estimated by performing a surface-based 3D/3D matching between the reconstructed models and the associated ground truth models that were derived from a CT-based reconstruction method. Such a matching also allowed for computing the errors between the reconstructed models and the associated ground truth models.Results: The technique could reconstruct the surface models of all 14 pelvises directly from the landmark-based initialization. Depending on the surface-based matching techniques, the reconstruction errors were slightly different. When a surface-based iterative affine registration was used, an average reconstruction error of 1.6 mm was observed. This error was increased to 1.9 mm, when a surface-based iterative scaled rigid registration was used.Conclusions: It is feasible to reconstruct a scaled, patient-specific surface model of the pelvis from single standard AP x-ray radiograph using the present approach. The unknown scale of the reconstructed model can be estimated by performing a surface-based 3D/3D matching.