Méthode automatique pour la localisation et l’identification de structures osseuses dans les radiographies EOS

Abstract

The calculation of clinical parameters and the three-dimensional reconstruction of the bone structures from biplane radiographs in clinical practice are essential tasks in the field of medical imaging, particularly in orthopedics, to establish the diagnosis and plan the medical procedure. The processes of clinical measurements and 3D reconstruction are based on the task of extracting information (for example : identifying on frontal and lateral radiographs the center of the femoral head, segmenting the tibia). This extraction task, mainly performed manually by healthcare professionals, is repetitive, time consuming and can lead to results with greater variability. As a result, it could lead to 3D reconstruction error and miscalculation of clinical parameters (for example : neck-shaft angle, femoro-tibial angle), which are important when making a decision before, during and after surgery; which could compromise the patient’s health. In EOS stereo-radiographs (images from the low-dose EOS radiology system) presenting strong bone superpositions, different fields of view (full-body images, lower limb images or images with partial structures) and orientations (0°, 20°, 45°, 70° or 90°), this extraction task is more difficult. This explains the non-existence of an automatic method of extracting lower limb information covering this diversity of EOS images. The non-resolubility of this problem explains why currently in clinical practice, the initialization step of the processes of clinical measurement and three-dimensional reconstruction of the lower-limb bone structures step consisting in identifying the primitives of interest (for example : sphere, points) on each bone structure is still performed manually. The main objective of this work is to overcome this problem by automating the task of extracting information related to the lower limbs on EOS images presenting different points of view and orientations in order to automatically initiate the processes of clinical measurements and 3D reconstruction for EOS radiographs. We have developed a fully automatic method of segmenting bone structures and identifying landmarks of the lower limbs for EOS radiographs. The proposed segmentation approach is based on a data augmentation approach, allowing the generation of images containing both partial and complete structures and on a new neural architecture called RobustNet. The landmark identification approach is mainly based on a Siamese network which takes EOS image pairs as input and predicted three-dimensional points as output. The proposed methods for bone structures segmentation and lower limb landmark identification were respectively tested on 70 and 30 EOS images presenting different points of view (frontal and sagittal images of the lower limb and of the whole body) and the orientations of the patient (0 °, 90 ° and 45 °). The results obtained are promising to be integrated into the EOS image analysis software platform for future evaluation. This work will make it possible to automatically extract bone structures, regions or points of interest of the lower limbs in databases of EOS images. It will make it possible to automatically initiate the process of clinical measurements and three-dimensional reconstruction of the lower limb bone structures and will allow a reduction in the processing time performed on EOS radiographs.

Date	15 Nov 2021
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Carlos Vázquez (Supervisor), Jacques A. de Guise (Co-supervisor) & Matthew Toews (Co-supervisor)