Close
Supervisors names
Dr Jon Clayden
Dr Leevi Kerkelä, Mr Kristian Aquilina, Prof Chris Clark
Background: Registration is the process of transforming images to bring them into alignment with each other. Registration algorithms have been developed over several decades, with a range of approaches having been explored, and in the field of medical imaging they have been extended into three dimensions (Oliveira et al., 2012; DOI 10.1080/10255842.2012.670855). Registration is an important step in many medical image processing pipelines, as it allows for nuisance effects such as patient motion to be corrected, and enables images to be combined to create new composite images bringing together the information available in each of the components. It is also one of the areas of medical image analysis in which modern machine learning methods such as deep learning have shown promise (Chen et al., 2022; DOI 10.1016/j.media.2022.102444).
Intraoperative magnetic resonance imaging (iMRI) offers paediatric neurosurgeons up-to-date images to guide sensitive surgery that may involve resecting (removing) tissue close to functionally important brain tissue. Registering these images with preoperative scans and/or atlas images is often important, but this context can present unique challenges (Alam et al., 2018; DOI https://doi.org/10.1016/j.bbe.2017.10.001). For example, patients are under anaesthesia and therefore are unable to hold their heads in the conventional position, and image artefacts can be appreciably worse due to the skull being open for the surgery. Few registration approaches have been designed with this particular scenario in mind.
Aims: Registration methods typically use an optimisation algorithm to find the most suitable parameters needed to align the images of interest, over a certain family of transformations, with respect to some measure of alignment quality. Since the computational cost of evaluating many possible transformations can be large, a heuristic search strategy is generally used to explore the parameter space. The unusual characteristics of iMRI scans can, however, violate the assumptions of some of these strategies, as can the removal of some tissue relative to the reference scan. The purpose of this project will be to appraise current approaches to image registration in the iMRI context, evaluate how well they work for paediatric image data acquired at GOSH, and explore new or improved approaches specialised for this purpose. This would also benefit other image processing projects running at GOS ICH targeting the same type of images, such as a recently developed method for rapid white matter segmentation.
Methods/Timeline: Following an initial literature review and familiarisation with medical image analysis tools (months 1-3), retrospective image data from iMRI cases scanned at GOSH will be collated and the performance of several existing registration methods compared for preoperative-to-intraoperative and between-modality intraoperative registration (months 3-6). Novel approaches to address shortcomings in the results will be investigated, and tested on a subset of the cases as well as scans from healthy adults and children (months 7-15). The left-out cases will then be used for a detailed evaluation of the approach (months 16-18), followed by integration of the novel registration strategy into image segmentation pipelines and investigation of additional uses (months 19-30). The student will then write up their thesis (months 31-36).
Contact
Jon Clayden (j.clayden@ucl.ac.uk)