Automatic personalized 3D model reconstruction of the distal tibiofibular joint from biplanar ankle X-Ray images for syndesmosis assessment

Abstract

The tibiofibular joint plays a critical role in ankle stability, and its injuries, particularly syndesmosis injuries, can lead to significant functional impairments. Accurate assessment of such injuries is crucial for diagnosis and treatment planning. Traditional imaging techniques, including conventional radiographs, Computed Tomography (CT) scans, and Magnetic Resonance Imaging (MRI), each have limitations in terms of loss of 3D information, radiation exposure, cost, and the ability to capture weight-bearing conditions. The EOS Imaging system, which provides biplanar X-ray images with low radiation in a standing position, presents a promising alternative. However, its application in 3D reconstruction of the tibiofibular joint remains challenging due to factors such as two-dimensional projection, which leads to a loss of three-dimensional information and superimposition of structures, and the small inter-bone distances of the tibiofibular joint. This thesis proposes a novel method for personalized 3D reconstruction of the tibiofibular joint from biplanar radiographs. The approach integrates statistical shape and intensity modeling and deep learning-based 2D/3D registration techniques to achieve accurate joint reconstruction. A statistical shape and intensity model (SSIM) of the tibiofibular joint is developed using a dataset of CT scans, ensuring an anatomically coherent representation. The reconstruction pipeline includes automatic extraction of the joint from radiographs using deep autoencoder-based neural networks, followed by a one-shot deep-learning-based 2D/3D registration framework. Experimental results demonstrate the effectiveness of the proposed method for reconstructing anatomically accurate 3D models of the tibiofibular joint from ankle biplanar radiographs. The reconstructed models enable automatic calculation of clinically relevant syndesmosis parameters, offering a non-invasive alternative to CT scans for injury assessment. This research contributes to the advancement of personalized orthopedic diagnostics by leveraging AI-driven reconstruction techniques, ultimately improving injury detection and treatment outcomes.

Date	18 Nov 2025
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Carlos Vázquez (Supervisor), Jacques A. de Guise (Co-supervisor) & Marie Lyne Nault (Co-supervisor)