TitleA hierarchical spectral clustering and non-linear dimensionality reduction scheme for detection of prostate cancer from magnetic resonance spectroscopy
NameTiwari, Pallavi (author), Cai, Li (chair), Madabhushi, Anant (internal member), Kim, Sobin (internal member), Rosen, Mark (outside member), Rutgers University, Graduate School - New Brunswick,
Magnetic resonance imaging
DescriptionMagnetic Resonance Spectroscopy (MRS) is a unique non-invasive method which has recently been shown to have great potential in screening of prostate cancer (CaP). MRS provides functional information regarding the concentrations of different biochemicals present in the prostate at single or multiple locations within a rectangular grid of spectra superposed on the structural T2-weighted Magnetic Resonance Imaging (MRI). Changes in relative concentration of specific metabolites including choline, creatine and citrate compared to "normal" levels is highly indicative of the presence of CaP. Most previous attempts at developing computerized schemes for automated prostate cancer detection using MRS have been centered on developing peak area quantification algorithms. These methods seek to obtain area under peaks corresponding to choline, creatine and citrate which is then used to compute relative concentrations of these metabolites. However, manual identification of metabolite peaks on the MR spectra, let alone via automated algorithms, is a challenging problem on account of low SNR, baseline irregularity, peak-overlap, and peak distortion. In this thesis work a novel computer aided detection (CAD) scheme for prostate MRS is presented that integrates non-linear dimensionality reduction (NLDR) with an unsupervised hierarchical clustering algorithm to automatically identify cancerous spectra. The methodology comprises of two specific aims. Aim 1 is to first automatically localize the prostate region followed in Aim 2 by automated cancer detection on the prostate obtained in Aim 1. In Aim 1, a hierarchical spectral clustering algorithm is used to distinguish between informative and non-informative spectra in order to localize the region of interest (ROI) corresponding to the prostate. Once the prostate ROI is localized, in Aim 2, a non-linear dimensionality reduction (NLDR) scheme in conjunction with a replicated k-means clustering algorithm is used to automatically discriminate between 3 classes of spectra (normal, CaP, and intermediate tissue classes). Results of qualitative and quantitative evaluation of the methodology over 18 1.5 Tesla (T) in-vivo prostate T2-w and MRS studies obtained from the multi-site, multi-institutional ACRIN trial, for which corresponding histological ground truth of spatial extent of CaP is available, reveal that the CAD scheme has a high detection sensitivity (89.60) and specificity (78.98). Results further suggest that the CAD scheme has a higher detection accuracy compared to such commonly used MRS analysis schemes as z-score and PCA.
NoteIncludes bibliographical references (p. 47-49).
Noteby Pallavi Tiwari
CollectionGraduate School - New Brunswick Electronic Theses and Dissertations
RightsThe author owns the copyright to this work.