3D approximation of scapula bone shape from 2D X-ray images using landmark-constrained statistical shape model fitting

Master Thesis


Permanent link to this Item
Journal Title
Link to Journal
Journal ISSN
Volume Title

University of Cape Town

Two-dimensional X-ray imaging is the dominant imaging modality in low-resource countries despite the existence of three-dimensional (3D) imaging modalities. This is because fewer hospitals in low-resource countries can afford the 3D imaging systems as their acquisition and operation costs are higher. However, 3D images are desirable in a range of clinical applications, for example surgical planning. The aim of this research was to develop a tool for 3D approximation of scapula bone from 2D X-ray images using landmark-constrained statistical shape model fitting. First, X-ray stereophotogrammetry was used to reconstruct the 3D coordinates of points located on 2D X-ray images of the scapula, acquired from two perspectives. A suitable calibration frame was used to map the image coordinates to their corresponding 3D realworld coordinates. The 3D point localization yielded average errors of (0.14, 0.07, 0.04) mm in the X, Y and Z coordinates respectively, and an absolute reconstruction error of 0.19 mm. The second phase assessed the reproducibility of the scapula landmarks reported by Ohl et al. (2010) and Borotikar et al. (2015). Only three (the inferior angle, acromion and the coracoid process) of the eight reproducible landmarks considered were selected as these were identifiable from the two different perspectives required for X-ray stereophotogrammetry in this project. For the last phase, an approximation of a scapula was produced with the aid of a statistical shape model (SSM) built from a training dataset of 84 CT scapulae. This involved constraining an SSM to the 3D reconstructed coordinates of the selected reproducible landmarks from 2D X-ray images. Comparison of the approximate model with a CT-derived ground truth 3D segmented volume resulted in surface-to-surface average distances of 4.28 mm and 3.20 mm, using three and sixteen landmarks respectively. Hence, increasing the number of landmarks produces a posterior model that makes better predictions of patientspecific reconstructions. An average Euclidean distance of 1.35 mm was obtained between the three selected landmarks on the approximation and the corresponding landmarks on the CT image. Conversely, a Euclidean distance of 5.99 mm was obtained between the three selected landmarks on the original SSM and corresponding landmarks on the CT image. The Euclidean distances confirm that a posterior model moves closer to the CT image, hence it reduces the search space for a more exact patient-specific 3D reconstruction by other fitting algorithms.