AbstractRespiratory motion a ects a wide range of techniques in the eld of medical image acquisition and analysis. In image-guided interventions it may cause misalignment of static road maps with the patient's anatomy. In imaging such as magnetic resonance (MR) imaging or positron emission tomography (PET) it may cause the images to appear blurred which may impede disease diagnosis and staging.
Manifold learning is a powerful tool for the non-linear dimensionality reduction of imaging data, which can be used to uncover the data's dominating sources of motion. By aligning low-dimensional embeddings of multiple datasets, which vary due to the same motion, in a joint low-dimensional space accurate correspondences between the datasets can be established.
In the first part of this thesis, manifold alignment is investigated for the robust reconstruction of high-resolution 4D (3D+time) MR imaging sequences of respiratory motion from sequentially acquired coronal 2D MR slices. In particular, a novel groupwise manifold alignment scheme is presented which outperforms two current state-of-the-art reconstruction techniques. From such 4D MR images very accurate motion estimates are derived, which, in turn, are used to correct for motion in simulated PET-MRI data.
In the second part of this thesis, a patient-specific respiratory motion model presented based on groupwise manifold alignment. Such a motion model can be used to correct for 3D organ motion during an image-guided intervention where only 2D images are available. It is shown that the aligned low-dimensional representations obtained using manifold alignment may be viewed directly as a surrogate-driven motion model. By updating this low-dimensional manifold with points obtained from new 2D imaging data, the model can automatically adapt to previously unseen breathing patterns.
Lastly, in the third part of this thesis, a novel manifold alignment method is outlined which does not require any of the datasets to be comparable in the image space. To this end a novel similarity kernel is proposed which allows comparison of visually distinct datasets by analysing their internal graph structure using a random walk approach. This allows to embed all data simultaneously in one step rather than in groups. It is shown that improved 4D MR reconstruction from multiple sagittal 2D MR slices can be achieved using this approach. Furthermore, the method is demonstrated on the time-resolved compounding of multiple 3D ultrasound views of the same organ.
|Date of Award||2016|
|Supervisor||Andrew King (Supervisor), Daniel Rueckert (Supervisor) & Jamie McClelland (Supervisor)|