AbstractsComputer Science

Efficient multi-modal image registration based on gradient orientations of minimal uncertainty

by Dante De Nigris Moreno

Institution: McGill University
Department: Department of Electrical and Computer Engineering
Degree: PhD
Year: 2015
Keywords: Applied Sciences - Computer Science
Record ID: 2058321
Full text PDF: http://digitool.library.mcgill.ca/thesisfile130290.pdf


This thesis presents a general framework for the registration of medicalimages across multiple clinical contexts involving rigid and non-rigidapplications. The proposed framework relies on gradient orientations asprimitive geometric descriptors so as to locally assess image similarity basedon orientation alignment and evaluates the metric on sparse locationscorresponding to anatomical boundaries of interest. The two main advantagesbrought forward by the proposed approach are: (1) a substantial reduction incomputational complexity and processing time and (2) a significant improvementin robustness against multi-modal contexts with widely different imageformation models and significant non-homogeneities.The proposed approach is evaluated in multiple clinical contexts and comparedagainst state-of-the-art techniques. In the context of neurosurgery, imageregistration can be employed so as to update a pre-operative magneticresonance image (MRI) based on an intra-operative ultrasound volume. Theproposed approach is evaluated in this challenging time-sensitive scenario andis shown to provide robust performance with sub-second processing times. Inthe context of the rigid registration of computational tomography (CT) and MRIbrain volumes, the proposed approach is evaluated with a publicly availabledataset and compared against previously proposed techniques. The quantitativeresults demonstrate that the proposed approach can employ highly reducedsampling rates (e.g. only 0.05% of the voxels in the image) while stillyielding a median registration error inferior to 1mm. In the context of thenon-rigid registration of inter-patient MRI brain volumes, the proposedapproach is evaluated with a publicly available dataset which measuresregistration accuracy in terms of the agreement of spatially mapped labelswith expert annotated labels. The use of such dataset allows for a fair andunbiased comparison with over fourteen competing techniques. The quantitativeresults show that the proposed approach achieves slightly inferior accuracythan the top performing method but with only one sixth of the processing timerequired by the alternative technique. Finally, the proposed approach isevaluated in the context of automatic brain lesion detection which relies onhealthy tissue probability maps obtained via registration to a brain atlas.The quantitative comparison against two leading image registration techniquesshows that the proposed approach can lead to a slightly improved performanceof brain lesion detection algorithms while requiring only one sixth of theprocessing time used by competing registration approaches. Cette thèse porte sur de nouvelles techniques de recalage d'images médicalesdans différents contextes cliniques et avec applications rigides etnon-rigides. Le principe de base est l'utilisation d'orientation de gradientsen tant que descripteurs géométriques primitifs. Cette technique permetd'évaluer localement la similitude de l'image en se basant sur l'alignement del'orientation et de restreindre l'évaluation de la mesure de…