Deep learning applied to medical image registration

Sloan, James M. (2018) Deep learning applied to medical image registration. EngD thesis, University of Glasgow.

Due to Embargo and/or Third Party Copyright restrictions, this thesis is not available in this service.


The investigations of this thesis are in applying machine learning, and in particular deep learning, to the image registration problem. The thesis investigates how machine learning may be used to:
o Synthesise a modality X from a modality Y, such that a multi-modality registrationtask involving registering images of different modalities X and Y can be converted to amono-modality registration between a synthetic image of modality X from Y, and X.
o Learn rigid transformations to align two given images in a single pass, using deep learning.
o Learn non-rigid, non-parametric transformations to align two given images in a singlepass, using deep learning.
In addition to these proposed methods for registration, a novel training paradigm for neural networks for image regression tasks is proposed and investigated. Image denoising, modality conversion and displacement field regression are the applications trained within the proposed training paradigm, and superior results observed. Each of the proposed registration methods is validated with a series of synthetic registration experiments, such that the data being aligned was real data, primarily from the OASIS [101] but the underlying deformations are synthetic. Modality conversion did not improve the accuracy of the registration schemes, and the modality conversion results were as good, in mean and standard deviation of the registration error.
The learned rigid and non-rigid registration schemes, where a neural network predicts the transformation parameters to align two given images, gave substantially better results compared to the baseline, multi-scale, iterative registration schemes with significant lower errors between the ground truth and predicted transforms. In addition to this, substantial speed-up was observed for the deep learning methods when compared to the runtimes from the baseline methods, as the deep learning methods estimate the transformation parameters in a single pass while the baseline method iteratively computes the optimum transformation.
The findings of this thesis suggest that it is possible to rigidly align two images, either of the same or different modality, in a single pass. In addition, the experiments demonstrate it is possible to non-rigidly align mono-modal images in a single pass, but our experiments show it is not possible, with the data publicly available, to non-rigidly align two images of different modalities with the proposed methodology.

Item Type: Thesis (EngD)
Qualification Level: Doctoral
Keywords: Deep learning, adversarial methods, medical image registration, MRI, head, neck, rigid, non-rigid, modality conversion.
Subjects: Q Science > QC Physics
T Technology > T Technology (General)
Colleges/Schools: College of Science and Engineering > School of Computing Science
Supervisor's Name: Siebert, Dr. J.P. and Goatman, Dr. K.A.
Date of Award: 2018
Embargo Date: 31 August 2024
Depositing User: Mr J. M. Sloan
Unique ID: glathesis:2018-30894
Copyright: Copyright of this thesis is held by the author.
Date Deposited: 11 Nov 2019 16:33
Last Modified: 10 Mar 2020 09:47

Actions (login required)

View Item View Item