A deep learning algorithm for contour detection in synthetic 2D biplanar X-ray images of the scapula: towards improved 3D reconstruction of the scapula

Master Thesis

2020

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

University of Cape Town

Publisher
License
Series
Abstract
Three-dimensional (3D) reconstruction from X-ray images using statistical shape models (SSM) provides a cost-effective way of increasing the diagnostic utility of two-dimensional (2D) X-ray images, especially in low-resource settings. The landmark-constrained model fitting approach is one way to obtain patient-specific models from a statistical model. This approach requires an accurate selection of corresponding features, usually landmarks, from the bi-planar X-ray images. However, X-ray images are 2D representations of 3D anatomy with super-positioned structures, which confounds this approach. The literature shows that detection and use of contours to locate corresponding landmarks within biplanar X-ray images can address this limitation. The aim of this research project was to train and validate a deep learning algorithm for detection the contour of a scapula in synthetic 2D bi-planar Xray images. Synthetic bi-planar X-ray images were obtained from scapula mesh samples with annotated landmarks generated from a validated SSM obtained from the Division of Biomedical Engineering, University of Cape Town. This was followed by the training of two convolutional neural network models as the first objective of the project; the first model was trained to predict the lateral (LAT) scapula image given the anterior-posterior (AP) image. The second model was trained to predict the AP image given the LAT image. The trained models had an average Dice coefficient value of 0.926 and 0.964 for the predicted LAT and AP images, respectively. However, the trained models did not generalise to the segmented real X-ray images of the scapula. The second objective was to perform landmark-constrained model fitting using the corresponding landmarks embedded in the predicted images. To achieve this objective, the 2D landmark locations were transformed into 3D coordinates using the direct linear transformation. The 3D point localization yielded average errors of (0.35, 0.64, 0.72) mm in the X, Y and Z directions, respectively, and a combined coordinate error of 1.16 mm. The reconstructed landmarks were used to reconstruct meshes that had average surface-to-surface distances of 3.22 mm and 1.72 mm for 3 and 6 landmarks, respectively. The third objective was to reconstruct the scapula mesh using matching points on the scapula contour in the bi-planar images. The average surface-to-surface distances of the reconstructed meshes with 8 matching contour points and 6 corresponding landmarks of the same meshes were 1.40 and 1.91 mm, respectively. In summary, the deep learning models were able to learn the mapping between the bi-planar images of the scapula. Increasing the number of corresponding landmarks from the bi-planar images resulted into better 3D reconstructions. However, obtaining these corresponding landmarks was non-trivial, necessitating the use of matching points selected from the scapulae contours. The results from the latter approach signal a need to explore contour matching methods to obtain more corresponding points in order to improve the scapula 3D reconstruction using landmark-constrained model fitting.
Description

Reference:

Collections