Wavelets and sparse methods for image reconstruction and classification in neuroimaging

摘要

This dissertation contributes to neuroimaging literature in the fields of compressed sensing magnetic resonance imaging (CS-MRI) and image-based detection of Alzheimer’s disease (AD). It consists of three main contributions, based on wavelets and sparse methods.udThe first contribution is a method for wavelet packet basis optimisation for sparse approximation and compressed sensing reconstruction of magnetic resonance (MR) images of the brain. The proposed method is based on the basis search algorithm developed by Coifman and Wickerhauser, with a cost function designed specifically for compressed sensing. It is tested on MR images available from the Alzheimer’s Disease Neuroimaging Initiative (ADNI).udThe second contribution consists of evaluating and comparing several sparse classification methods in an application to detection of AD based on positron emission tomography (PET) images of the brain. This comparison includes univariate feature selection, feature clustering and classifiers that automatically select a small subset of features due to their mathematical or algorithmic construction. The evaluation is based on PET images available from ADNI.udThe third contribution is proposing an extension of wavelet-based scattering networks (originally proposed by Mallat and Bruna) to three-dimensional tomographic images. The proposed extension is evaluated as a feature representation in an application to detection of AD based on MR images available from ADNI.udThere are several possible extensions of the work presented in this dissertation. The wavelet packet basis search method proposed in the first contribution can be improved to take into account the coherence between the sparse approximation basis and the sensing basis. The evaluation presented in the second contribution can be extended with additional algorithms to make it more comprehensive. The three-dimensional scattering networks that are the core part of the third contribution can be combined with other machine learning methods, such as manifold learning or deep convolutional neural networks.udAs a whole, the methods proposed in this dissertation contribute to the work towards efficient screening for Alzheimer’s disease, by making MRI scans of the brain faster and helping to automate image analysis for AD detection.udThe first contribution is a method for wavelet packet basis optimisation for sparse approximation and compressed sensing reconstruction of magnetic resonance (MR) images of the brain. The proposed method is based on the basis search algorithm developed by Coifman and Wickerhauser, with a cost function designed specifically for compressed sensing. It is tested on MR images available from the Alzheimer’s Disease Neuroimaging Initiative (ADNI).udThe second contribution consists of evaluating and comparing several sparse classification methods in an application to detection of AD based on positron emission tomography (PET) images of the brain. This comparison includes univariate feature selection, feature clustering and classifiers that automatically select a small subset of features due to their mathematical or algorithmic construction. The evaluation is based on PET images available from ADNI.udThe third contribution is proposing an extension of wavelet-based scattering networks (originally proposed by Mallat and Bruna) to three-dimensional tomographic images. The proposed extension is evaluated as a feature representation in an application to detection of AD based on MR images available from ADNI.udThere are several possible extensions of the work presented in this dissertation. The wavelet packet basis search method proposed in the first contribution can be improved to take into account the coherence between the sparse approximation basis and the sensing basis. The evaluation presented in the second contribution can be extended with additional algorithms to make it more comprehensive. The three-dimensional scattering networks that are the core part of the third contribution can be combined with other machine learning methods, such as manifold learning or deep convolutional neural networks.udThis dissertation contributes to neuroimaging literature in the fields of compressed sensing magnetic resonance imaging (CS-MRI) and image-based detection of Alzheimer’s disease (AD). It consists of three main contributions, based on wavelets and sparse methods.udThe first contribution is a method for wavelet packet basis optimisation for sparse approximation and compressed sensing reconstruction of magnetic resonance (MR) images of the brain. The proposed method is based on the basis search algorithm developed by Coifman and Wickerhauser, with a cost function designed specifically for compressed sensing. It is tested on MR images available from the Alzheimer’s Disease Neuroimaging Initiative (ADNI).udThe second contribution consists of evaluating and comparing several sparse classification methods in an application to detection of AD based on positron emission tomography (PET) images of the brain. This comparison includes univariate feature selection, feature clustering and classifiers that automatically select a small subset of features due to their mathematical or algorithmic construction. The evaluation is based on PET images available from ADNI.udThe third contribution is proposing an extension of wavelet-based scattering networks (originally proposed by Mallat and Bruna) to three-dimensional tomographic images. The proposed extension is evaluated as a feature representation in an application to detection of AD based on MR images available from ADNI.udThere are several possible extensions of the work presented in this dissertation. The wavelet packet basis search method proposed in the first contribution can be improved to take into account the coherence between the sparse approximation basis and the sensing basis. The evaluation presented in the second contribution can be extended with additional algorithms to make it more comprehensive. The three-dimensional scattering networks that are the core part of the third contribution can be combined with other machine learning methods, such as manifold learning or deep convolutional neural networks.udAs a whole, the methods proposed in this dissertation contribute to the work towards efficient screening for Alzheimer’s disease, by making MRI scans of the brain faster and helping to automate image analysis for AD detection.