Sparse Reconstruction of Compressive Sensing Magnetic Resonance Imagery using a Cross Domain Stochastic Fully Connected Conditional Random Field Framework
Loading...
Date
2016-07-08
Authors
Li, Yaguang
Advisor
Wong, Alexander
Journal Title
Journal ISSN
Volume Title
Publisher
University of Waterloo
Abstract
Prostate cancer is a major health care concern in our society. Early detection of prostate
cancer is crucial in the successful treatment of the disease. Many current methods used in
detecting prostate cancer can either be inconsistent or invasive and discomforting to the
patient. Magnetic resonance imaging (MRI) has demonstrated its ability as a non-invasive
and non-ionizing medical imaging modality with a lengthy acquisition time that can be
used for the early diagnosis of cancer. Speeding up the MRI acquisition process can greatly
increase the number of early detections for prostate cancer diagnosis.
Compressive sensing has exhibited the ability to reduce the imaging time for MRI by
sampling a sparse yet sufficient set of measurements. Compressive sensing strategies are
usually accompanied by strong reconstruction algorithms. This work presents a comprehensive
framework for a cross-domain stochastically fully connected conditional random
field (CD-SFCRF) reconstruction approach to facilitate compressive sensing MRI. This
approach takes into account original k-space measurements made by the MRI machine
with neighborhood and spatial consistencies of the image in the spatial domain. This
approach facilitates the difference in domain between MRI measurements made in the
k-space, and the reconstruction results in spatial domain. An adaptive extension of the
CD-SFCRF approach that takes into account regions of interest in the image and changes
the CD-SFCRF neighborhood connectivity based on importance is presented and tested as
well. Finally, a compensated CD-SFCRF approach that takes into account MRI machine
imaging apparatus properties to correct for degradations and aberrations from the image
acquisition process is presented and tested.
Clinical MRI data were collected from twenty patients with ground truth data examined
and con firmed by an expert radiologist with multiple years of prostate cancer diagnosis
experience. Compressive sensing simulations were performed and the reconstruction
results show the CD-SFCRF and extension frameworks having noticeable improvements
over state of the art methods. Tissue structure and image details are well preserved while
sparse sampling artifacts were reduced and eliminated. Future work on this framework
include extending the current work in multiple ways. Extensions including integration into
computer aided diagnosis applications as well as improving on the compressive sensing
strategy.
Description
Keywords
Sparse reconstrction, Biomedical Imaging, Image processing, Computer Vision