英文摘要 | Molecular imaging is a medical imaging technology which involves
molecular biology, chemistry, physics, radiology, and computer science. It
can image the physiological activities of the organism at tissue level, cell
level, and even subcellular level. Compared with traditional medical
imaging, molecular imaging has the advantages of high sensitivity and high
specificity. Nowadays, molecular imaging has been widely used in
biomedical research. Optical molecular imaging is an important part of
molecular imaging. It takes optical signal as imaging medium, and can
conduct noninvasive, qualitative, and quantitative research on specific
physiological activities at the cellular and molecular level. In recent years,
optical molecular imaging has been greatly developed, and has been
successfully applied to preclinical research and clinical experiments.
Fluorescence molecular tomography (FMT) is an important optical
molecular imaging technology. Compared with two-dimensional
fluorescence molecular imaging, FMT can obtain the three-dimensional
distribution information of fluorescence probe in the organism, and realize
the three-dimensional visualization of certain biomarkers. In the process of
FMT imaging, the fluorescence information on the surface of organism is
first collected by high sensitivity detector, and then the distribution
information of fluorescence sources in the organism is obtained by
reconstruction method. The conventional reconstruction methods are based
on photon propagation model. The deviation of model description affects
the accuracy of FMT reconstruction. In addition, the inverse problem of
FMT reconstruction is ill-posed and under-determined, which limits the
quality and speed of morphological reconstruction in FMT.
This research focuses on FMT reconstruction method based on the
neighborhood information of fluorescence source to solve the above
problems, and can be divided into two parts. The first part depends on the
conventional photon propagation model-based reconstruction strategy. It
aims to design the regularization terms in inverse problem using the
neighborhood information of fluorescence distribution. On the basis of the
previous research, the second part proposes a machine learning-based
reconstruction strategy. This strategy utilizes data-driven method to train
the reconstruction model, and avoids building the photon propagation
model and solving the inverse problem. The major contributions of this
thesis are listed as follows:
A FMT reconstruction method based on nonlocal total variation
regularization has been proposed to improve the morphological
imaging quality. This algorithm refers to the intensity distribution
rule of fluorescence source in the organism, i.e., the fluorescence
difference between spatial points is positively related to their
distance. Based on this assumption, this algorithm designs the
nonlocal total variation regularization. This regularization
transforms the physical distance between spatial points into
Gaussian distance, and uses Gaussian weight as radial function to
construct nonlocal total variation regularization matrix. Thus, this
matrix heavily penalizes the intensity difference between the close
pairs of points, and lightly penalizes the intensity difference between
the distant pairs of points. The results of numerical simulation
experiments and orthotopic glioma mouse model experiments
demonstrate that the proposed method achieved more accurate
morphological reconstruction than the conventional methods.
A FMT reconstruction method based on adaptive Gaussian
weighted Laplace prior regularization has been proposed. The
prior regularization of the algorithm is improved from the Gaussian
weighted Laplace prior (GWLP) regularization. Besides the prior
assumption that the fluorescence difference of spatial points is
positively related to the corresponding distance, adaptive Gaussian
weighted Laplace prior (AGWLP) regularization assumes that the
fluorescence distribution is related to the energy region. Based on
the above two assumption, this algorithm builds an AGWLP
regularization matrix with both Gaussian distance and fluorescence
intensity. The regularization matrix not only considers the physical
distance between spatial points, but also takes into account the
energy range of neighborhood to which the spatial points belong.
The design of the regularization matrix reduces the fluorescence
difference in the fluorescence source, and improves the fluorescence
difference at the edge of the fluorescence source. The results of
numerical simulation experiments and orthotopic glioma mouse
model experiments demonstrate that the proposed method improved
the contrast-to-noise of reconstruction results, and achieved accurate
morphological reconstruction.
A FMT reconstruction method based on K-nearest neighbor
based locally connected network has been proposed. Different
from the conventional photon propagation model-based
reconstruction methods, this algorithm is a machine learning-based
reconstruction strategy. This reconstruction strategy uses neural
network to directly learn the inverse process of photon propagation
in biological tissue, which avoids building the photon propagation
model and solving the inverse problem. K-nearest neighbor based
locally connected (KNN-LC) network is improved from fully
connected (FC) network. It utilizes the prior assumption that the
fluorescence difference between spatial points is positively related
to the physical distance, and designs locally connected (LC) subnetwork. KNN-LC network cascades a FC sub-network with a LC
sub-network, which improves the imaging quality of morphological
reconstruction. In addition, in order to obtain a large number of FMT
samples, this algorithm simulates a number of single-source samples
with Monte Carlo method, and assembles a large number of dualsource samples and big-source samples using the method of sample
assembling. The results of numerical simulation experiments and
orthotopic glioma mouse model experiments demonstrate that KNNLC network can achieve both accurate source localization and
morphological reconstruction. Compared with the conventional
methods, machine learning-based reconstruction strategy greatly
improves the reconstruction speed and reduces the reconstruction
time by three orders of magnitude.
|
修改评论