Research On Key Technologies Of3D Patient-specific Pose Estimation Of Femurs Based On Digital Radiographs

Femoral shaft fracture is a common clinical orthopedic trauma. Walk dysfunction, limb deformity or other consequences may happen to the patients if it is not treated properly, which will significantly affect patients’ life quality. In modern medical diagnosis, the common clinical approaches for three-dimensional (3D) imaging mainly include computed tomography (CT) and magnetic resonance imaging (MRI). Although the3D data of limbs can be obtained directly by these two imaging methods, the cost, convenience in operation, imaging speed, and harmful radiation amount are inferior to that of digital radiograph (DR) imaging. However, DR images are2D projected images, and lacking of3D spatial information is the biggest shortcoming in clinical diagnosis with DR imaging equipment. Therefore, developing one approach for recovering3D pose of femur before surgery only by using one single or few limited two-dimensional (2D) DR images is of extremely clinical values and broad application prospects.This dissertation aims to design a method for quickly and effectively reconstructing femur patient-specific pose by using femoral DR images, and then realize description of femur in3D space. In this dissertation, corresponding solutions have been studied to handle the existed key problems such as DR imaging system calibration, feature extraction issue of inhomogeneous intensity distribution image,2D-3D image registration, and multiple extremum optimization of2D point set registration, which realizes patient-specific pose visualization of3D preoperative femoral fracture. It is studied deeply about three key technologies which are segmentation of inhomogeneous intensity distribution image, multiple extremum optimization problem and2D-3D image registration. The results promote the image-based guided visualization technology. The main contents of this dissertation are as follows:DR imaging system cannot be calibrated since images got from common plane calibration board are not clear under condition of X-rays. To handle this problem, PCB board containing copper grid was used to be new calibration board, the same calibration method based on plane calibration board was applied, and finally we obtained calibration parameters for the DR imaging system.Femoral fracture DR image is of complicated topology and inhomogeneous intensity distribution. To respond such kind characteristics, a new active contour model combined with grayscale fluctuation information was proposed. In details, we get grayscale fluctuation curve of an image basing on the concept of grayscale fluctuation; taking grayscale fluctuation transform to the image according to the judgment on the fluctuation range of a monotonic curve; introducing grayscale fluctuation transform function into a Chan-Vese (C-V) model and redefining the energy function, we finally obtain the new active contour model. Synthetic and medical image segmentation experiments showed that this new model not only inherits characteristics of perfect anti-noise performance, low computational complexity, initial contour insensitivity, etc., but also overcomes the shortcoming that C-V model cannot segments inhomogeneous intensity images. In addition, the new model can extract intact and continuous characteristic contour of femur fracture and marrow cavity in DR image with high accuracy.Point set registration is essentially a multiple extremum optimization process. Traditional intelligent optimization algorithms easily fall into local minimum and are inefficient in searching when they are used for complicated or special optimization problems. To address this issue, a characteristics statistical algorithm (CSA) was adopted to the point set registration and we proposed a point registration algorithm based on Gaussian mixture model (GMM). After determining the characteristics statistical items which are closely related to target function through recombining the registration parameters, it will be better to address the problem that parameter estimation of point set registration easily fall into local minimum and improve the efficiency of the algorithm in searching. Registration results of testing data proved that the proposed algorithm has good anti-noise, anti-point-missing and outliers resistant abilities. This algorithm is of higher registration accuracy and success rate for registration.To address the limitation of lacking of3D spatial information in DR images for clinical diagnosis, a3D pose estimation method based on affine2D-3D image registration using bi-planar DR images was proposed. After extracting characteristic contour of femur, we got2D registration parameters between segmented contours of DR image and projected contours of femur generic model by using point set registration. We further obtained3D pose matrix which is got through determining the transform relationship between2D plane and3D space based on pinhole calibration method. And then applying the pose matrix to generic model we finally obtain the patient-specific femoral pose. Experimental results showed that, under condition of lacking of3D patient-specific spatial information, this method can reconstruct patient-specific femur pose which is only depending on DR images from anteroposterior and lateral views with femoral generic model.