Automatic segmentation of the lumbar spine from medical images

Segmentation of the lumbar spine in 3D is a necessary step in numerous medical applications, but remains a challenging problem for computational methods due to the complex and varied shape of the anatomy and the noise and other artefacts often present in the images. While manual annotation of anatomical objects such as vertebrae is often carried out with the aid of specialised software, obtaining even a single example can be extremely time-consuming. Automating the segmentation process is the only feasible way to obtain accurate and reliable segmentations on any large scale. This thesis describes an approach for automatic segmentation of the lumbar spine from medical images; specifically those acquired using magnetic resonance imaging (MRI) and computed tomography (CT). The segmentation problem is formulated as one of assigning class labels to local clustered regions of an image (called superpixels in 2D or supervoxels in 3D). Features are introduced in 2D and 3D which can be used to train a classifier for estimating the class labels of the superpixels or supervoxels. Spatial context is introduced by incorporating the class estimates into a conditional random field along with a learned pairwise metric. Inference over the resulting model can be carried out very efficiently, enabling an accurate pixelor voxel-level segmentation to be recovered from the labelled regions. In contrast to most previous work in the literature, the approach does not rely on explicit prior shape information. It therefore avoids many of the problems associated with these methods, such as the need to construct a representative prior model of anatomical shape from training data and the approximate nature of the optimisation. The general-purpose nature of the proposed method means that it can be used to accurately segment both vertebrae and intervertebral discs from medical images without fundamental change to the model. Evaluation of the approach shows it to obtain accurate and robust performance in the presence of significant anatomical variation. The median average symmetric surface distances for 2D vertebra segmentation were 0.27 mm on MRI data and 0.02 mm on CT data. For 3D vertebra segmentation the median surface distances were 0.90 mm on MRI data and 0.20 mm on CT data. For 3D intervertebral disc segmentation a median surface distance of 0.54 mm was obtained on MRI data.