Élaboration d'outils de traitement d'images pour aider à la reconstruction 3D du rachis à partir d'images radiographiques postopératoires

Abstract

Adolescent idiopathic scoliosis is an illness that touches mostly young adolescents. It’s a tridimensional deformation of the spine that leads to modification of the posture. Unfortunately, the origins of this illness are still unknown. An important step in diagnosing scoliosis consists in manually identifying the anatomical landmarks on x-rays. These landmarks will help determine the clinical indexes to help diagnose and treatment of the illness. There are also used in 3-D reconstruction. The visualization of the scoliotic spine in 3-D allows a better evaluation of the deformation and a better diagnosis. One the diagnosis established, severe cases of scoliosis require surgery. This operation consists in straightening the spine et maintaining it with the help of instrumentation (rod, screws and hook). Automatic detection algorithms have been developed to facilitate this task however the presence of metal instrumentation prevents the algorithms from working properly. The problematic is due to the interference of the instrumentation after a surgery. This inconvenience is both visual for the reconstruction expert. It affects also the automatic detection algorithms. The objective of this thesis is to propose a solution that diminishes the disturbance for the operators during the reconstruction process. It also necessary to ensure that the automatic approaches be applicable to postoperative cases. This allows postoperative reconstructions to be accessed thanks to the developed methods for the pre operation without modifying the process, whether it be manual or automatic. The literature reviews and the feedback of different reconstruction experts on the subject allowed to determine the most suitable solution to answer our problematic. This solution comprises two main phases : 1. Segment the instrumentation and creation of a related mask; 2. Inpaint the previous mask with information which will allowed to continue the process. The mask replacement is realized with the help of filling methods called “Inpainting”. There exist two types of filling algorithms, algorithms based on diffusion and those based on filling by textures. A diffusion algorithm and three texture based algorithms were applied in the case of filling surgical instrumentation present in postoperative x-rays. The goal was to define which algorithm would be the most effective in this context. It was necessary to create a reference to evaluate the fill. The addition of an instrumentation mask on a non-scoliotic x-ray allowed to simulate the rod, screws and hook installations all while knowing the masked structures. The method proposed allowed a reduction in discomfort for the operators, the use of automatic approaches for the spine reconstruction on real x-rays with simulated instrumentation and reel postoperative images. The simulated instrumentation was created by adding an instrumentation mask on a healthy preoperative x-ray. The use of the Peak-Signal-Noise Ratio (PSNR) showed a 49 % improvement of the image in relation to the instrumented image. The metric Multi-Scale Structural Similarty (MSSIM) demonstrated that the filling of the instrumentation mask by using the algorithm type “diffusion-based” increased 30 % the similarity between an instrumented x-ray and a non instrumented x-ray. We have tested the application of this method in the framework of a 3D semi-automatic and automatic reconstruction. The semi-automatic application showed that the proposed method would not bring improvement in terms of time of reconstruction et precision. The user comfort is however increased thanks to the instrumentation filling. Furthermore, the average of error detection calculated by using the Deep Neural Network (DNN) was improved by 4,7 mm. The filled image as an error average of detection of 3,0 mm against 7,7 for the instrumentation simulation and 1,3 mm error average for the healthy x-ray. This method allows for the developed algorithms for preoperative images to function on postoperative images.

Date	15 Feb 2017
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Jacques A. de Guise (Supervisor) & Carlos Vázquez (Co-supervisor)