Reconstruction tridimensionnelle du pelvis à partir de radiographies biplanes

Abstract

Biplanar radiography gives enough information to reconstruct a bone in three dimensions, while limiting the radiation dose received by the patient. This technique is particularly useful for the medical monitoring of pathologies like adolescent idiopathic scoliosis. The patients with such pathology need frequent examination and the 3D visualization of the pelvis may help with the diagnosis or the monitoring of the disease. However, 3D reconstruction of a structure as big as the pelvis using only two biplanar radiographs cans be complex. The objective of this work is to present a method for pelvic reconstruction as automatic as possible, with a precision similar or better than the methods currently used. We use a principal component analysis to create an active shape model which is deformed in order to maximize the similarity between its projected contours and the reference contours detected on the radiographs. The initial position of the model is calculated from the position of 15 manually detected landmarks and the reference contours are automatically detected using a convolutional neural network (R2U-Net). To assess the relevancy of using a gendered model or not, three active shape models (one female, one male and one non-gendered) were created from 115 reference pelvises (64 females and 51 males) obtained from EOS acquisitions (EOS Imaging). We reconstructed 20 pelvises with an RMS distance of 2,3 ± 3,4 mm; 2,2 ± 3,0 mm et 2,2 ± 3,0 mm for the female, male and neutral models. Our method allows to reconstruct a pelvis from two biplanar radiographs with a precision similar to the method implemented by EOS Imaging, but with a method fully automated in the exception of its initialization.

Date	15 Dec 2023
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Carlos Vázquez (Supervisor) & Jacques A. de Guise (Co-supervisor)