Reconstruction 3D automatique de la colonne vertébrale à partir de radiographies bi-planaires EOS

Abstract

Three-dimensional (3D) modeling of the spine from postero-anterior (PA) and lateral (LAT) biplanar calibrated radiographs represents a useful tool for surgeons and clinicians. It provides accurate spine and pelvis 3D-2D clinical parameters used for patient diagnosis and follow-up. However, an accurate 3D modeling process involves numerous time-consuming user interventions (between 10 to 40 min.) that hinders its use in clinical routine. Therefore, the process still requires improvements in order to reduce both user supervision and reconstruction time. In this thesis, machine learning and image processing algorithms were used to develop a fully automated 3D spine reconstruction method. As a first step, a global spine and pelvis detection method in biplanar radiographs is proposed using a combination of a statistical spine model (SSM) (based on principal components analysis and composed of 23 vertebrae, spanning from C2 level to sacral endplate), and point detections based on convolutional neural networks (CNN). The SSM can predict realistic 3D vertebra models from a few information identificated in radiographs. The training database includes 799 adolescent idiopathic scoliosis (AIS) patients as well as 553 adults (65% had a scoliosis diagnosis, 3% had a degenerative spine and 32% were without pathologies). The vertebral body center locations detected using CNN are used to gradually deform the SSM in images. The SSM also serves to regularize the location of the points using statistically admissible deformation modes. A step of 3D/2D registration then locally refines the first estimate of the 3D models to increase their shape accuracy. The registration includes a prior step of image-to-image translation based on generative adversarial network in order to convert the original X-ray image into the same image domain of virtual X-ray generated from the 3D model. This process improves the similarity level between both image and allowed a robust image-image matching. For the method validation, the test set includes 60 patients, 40 AIS patients (mean age 14 year old, Cobb angle range [32° - 102°]) and 20 adult patients (mean age 40 year old, Cobb angle range [13° - 56°]). The automated 3D reconstructions are compared to reference reconstructions, build as “bronze standards”, which are generated from the 3D reconstructions made by three different experts. The mean 3D localization error of anatomical landmarks were 1.7 mm and 1.3 mm, for the pedicles and the middle of vertebral endplates, respectively. For the clinical parameters extracted from the 3D modeling (six parameters of kyphosis/lordosis, three Cobb angles and linked axial vertebral rotations, five pelvis parameters and seven alignment parameters), the automated algorithm success (rate of clinical measurement directly useable in clinics, i.e. with errors falling in the expert confidence interval) were 82% and 84%, for adults and AIS patients, respectively. For instance, the main Cobb angle mean error is 0.8 ± 2.9° (for AIS). The computation time (1’30” on average for the whole spine) is negligible for the clinician since the 3D modeling may be started each time an X-ray is taken thanks to the proposed 3D modeling, which is designed to be launched automatically. The proposed automated 3D spine reconstruction method, reducing both the user supervision and the modeling time, should help the dissemination and adoption of 3D measurements in clinical routine.

Date	15 Dec 2020
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Jacques A. de Guise (Supervisor) & Carlos Vázquez (Co-supervisor)