Segmentation de l'articulation de la hanche à partir de radiographies biplanes en utilisant une approche multi-structures

Abstract

EOS system which is less radiating than CT scan allows acquiring biplanar radiographs in a weight-bearing position in order to enable 3D reconstruction of bony structures. To achieve 3D reconstruction of a bony structure from biplanar radiographs, extracted contours from images are combined with 3D a priori information. This study focuses on simultaneous contour extraction of adjacent bony structures of the hip joint, i.e. femoral head and acetabulum. This task is considered hard due to high radiographic noise, low contrast and structure superposition. We have adopted a multi-structure approach which enables us to obtain simultaneously two contours corresponding each to one structure. Our proposed method requires an initialisation of the femoral head by a circle in order to extract and localize the joint region. The segmentation process is further applied to this region which is transformed into a straightened space, where the sought contours correspond to two aligned paths. Vertical edges of the straightened image are then detected by applying a horizontal gradient operator. Two edge images are generated by separating opposed signs of the gradient in which each structure contour is represented by a path. A straightened volume is constructed by combining the two edge images, which allows the combination of both paths in one. We propose a 3D minimal path searching algorithm in a volume where inter-structure constraints are imposed. The projection of the obtained 3D minimal path on edge images returns two distinct and labeled paths. The labeled contours of the two structures are recuperated in the original image space by converting the two paths. We applied this method on 100 radiographic images of the lower limb, containing each two hip joints, acquired in frontal, oblique and lateral directions. The method accuracy was evaluated and global errors point-contour (RMS±STD), equal to 1.27±0.79 for the femoral head and to 1.26±0.74 for the acetabulum, were observed. 80% of local repeatability errors were less than 1 mm. The results show that the method is accurate, but it also depends on the complexity of the segmented image and the presence or not of pathology such as hip osteoarthritis. This work will definitely improve the 3D reconstruction workflow in terms of accuracy.

Date	6 May 2015
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Jacques A. de Guise (Supervisor) & Carlos Vázquez (Co-supervisor)