Show simple item record

dc.contributor.advisorGhosh, Debashis
dc.contributor.authorDalwani, Manish Shivkumar
dc.date.accessioned2017-09-12T19:53:08Z
dc.date.available2018-09-12T19:46:51Z
dc.date.submitted2017
dc.identifierDalwani_ucdenveramc_1639D_10462.pdf
dc.identifier.urihttps://hdl.handle.net/10968/2025
dc.descriptionIncludes bibliographical references.
dc.descriptionSummer
dc.description.abstractMachine learning in neuroimaging modalities is important for building a successful prediction model, especially in the field of Psychiatry. For example, successful classification of groups, tasks and behaviors leads to the possibility of automated diagnostic detection. Similarly, prediction of a behavioral outcome using regression approaches could provide an insight into a certain behavioral pattern in a patient population. Support vector machine (SVM) has been successfully implemented in neuroimaging classification and regression frameworks. Two popular kernels used in SVM are linear and radial basis function (RBF). We utilized non-linear wavelet kernels in conjunction with support vector machine. The reasoning behind this was that wavelets have useful properties that may fit well with the intrinsic nature of imaging data. We constructed wavelet kernels such as the Morlet, Mexican Hat and some of their variants. We ensured that these kernels functions provide valid Gram matrices that are positive definite and also proved that these functions are in Reproducible Kernel Hilbert Space (RKHS). We noticed found that in certain situations non-linear wavelet kernels outperformed linear and RBF. Based on our simulations for functional MRI data, we found that the proposed kernels worked well in situations of non-linear separation between groups and showed that non-linear wavelet kernels performed the best. Next, we proposed to extendextended this work to situations with more than one modality. We applied non-linear wavelet kernels in conjunction with SVM and Hadamard products to integrate across modalities. We do so with different combinations of modalities such as functional MRI, and behavioral Genetics and separately functional MRI and structural MRI. We found that Hadamard product in some cases outperformed use of a single kernel with a simple concatenation of features from two modalities. Finally, we built a Graphical User Interface that we called “Wavelet Machine” to insure usersso that others can apply this work to their own neuroimaging data.
dc.languageEnglish
dc.language.isoeng
dc.publisherUniversity of Colorado Anschutz Medical Campus. Strauss Health Sciences Library
dc.rightsCopyright of the original work is retained by the author.
dc.subjectWavelet Kernels
dc.subject.lcshHadamard transform spectroscopy
dc.subject.meshMachine Learning
dc.subject.meshMagnetic Resonance Imaging
dc.subject.meshSupport Vector Machine
dc.subject.meshClassification
dc.titleMachine learning in neuroimaging based modalities using support vector machines with wavelet kernels
dc.typeThesis
dc.rights.accessEmbargo Expires: 09/12/2018
dcterms.embargo.terms2018-09-12
dcterms.embargo.expires2018-09-12
dc.contributor.committeememberMikulich-Gilbertson, Susan K.
dc.contributor.committeememberSakai, Joseph T.
dc.contributor.committeememberHughes, John
dc.contributor.committeememberWagner, Brandie D.
thesis.degree.nameDoctor of Philosophy (Ph.D.)
thesis.degree.levelDoctoral
thesis.degree.disciplineBiostatistics
thesis.degree.grantorUniversity of Colorado at Denver, Anschutz Medical Campus


Files in this item

Thumbnail

This item appears in the following Collection(s)

Show simple item record