Supervoxel-based image registration and analysis with application to lung cancer

Lung diseases, including lung cancer, are amongst the largest burdens to healthcare systems worldwide. Improving the extraction of information from imaging data has the potential to provide more accurate diagnosis and more effective treatment. For example, assessing regional lung function could guide more effective radiotherapy treatment to spare well-functioning parts of the lungs. One meaningful regional representation is achieved by locally clustering similar image voxels together using the concept of supervoxels, which allows for bulk regional processing and analysis, whilst removing redundant information. This thesis aims to lay further foundations for supervoxel-based image analysis by presenting novel deformable lung image registration frameworks, as well as a method to estimate lung ventilation using supervoxels. We demonstrate how a supervoxel-based image representation can be combined with graph cuts as a discrete optimisation-based approach to provide effcient and accurate 3D deformable image registration. Our work overcomes some of the previous limitations by posing the problem on a graph created by adjacent supervoxels. We further show that the application of a relaxed graph representation of the image, followed by guided image filtering over the estimated deformation �field, allows us to model `sliding motion'. Applying this method to lung image registration results in highly accurate image registration and anatomically plausible estimations of the deformations. The resulting deformation �fields might be further applied to estimate lung ventilation maps. We present a novel approach for estimating regional lung ventilation from dynamic lung CT imaging. Our method combines a supervoxel-based image representation with deformable image registration, performed between peak breathing phases, during which we track changes in intensity for a number of layers of previously extracted supervoxels. Such a region-based approach is expected to be physiologically more consistent with lung anatomy than methods relying on voxel-wise relationships. Our work also presents novel approaches for performing more accurate hyperpolarised 129Xenon MRI (XeMRI) analysis. We propose a multimodal lung image registration enhanced with personalised motion model approach derived from lung 4DCT. The approach with a prior motion model is particularly important for regions where there is not enough information to reliably drive the registration process, as in the case of XeMRI and proton density MRI (pMRI) to CT registration. Subsequently we introduce a framework for breathing motion correction of the dynamic sequence of XeMRI using a lung atlas-based approach. Such a method for the breathing motion correction is expected to facilitate the analysis of temporal lung ventilation.