UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Automatic localization and labelling of spine vertebrae in MR images using deep learning Danaei, Pardiss


Magnetic Resonance (MR) and Computed Topography (CT) are the most common modalities for spine imaging. Localization and identification of vertebrae is an essential first step in examining these volumes for diagnosis, surgical planning and management of patients with disc or vertebra pathologies and conditions. With large volumes of spinal scans acquired at imaging centres, development of a computerized solution for spine labelling has received attention from several research groups, as it can help save radiologists time and clicks. It can also expedite the imaging-dependent pre- and post-operation procedures. Nonetheless, automatic spine labelling in CT and MR is non-trivial and has proven challenging. This is due to: 1) limited and variable field-of-view (FOV); 2) variability in imaging parameters and resolution; 3) variability in shape, size and appearance of the spinal anatomies, especially in the presence of various pathologies or implants; 4) the repetitive nature of the spine and similar appearance of the vertebrae; and particularly for learning-based solutions, 5) dependence on expert annotations. In this thesis, learning-based approaches that perform simultaneous identification and localization of vertebrae are introduced. The principal goal is to design a supervised spine labelling approach that requires minimal manual annotations, and can perform both identification and localization tasks within a unified framework. We achieved an identification rate of 89.76%.

Item Citations and Data


Attribution-NonCommercial-NoDerivatives 4.0 International