Statistical Models for Segmentation from MR Localizer Images (English shop)

Preview

Table of Contents, Datei (33 KB)
Extract, Datei (140 KB)

This thesis presents methods for the segmentation of anatomical structures from magnetic res-
onance (MR) localizer images. Localizer images are obtained from fast pre-scan protocols and
are usually used for scan planning. Segmentations of anatomical structures from these images
can leverage applications in automated scan planning like organ localization, slice positioning
tasks, sequence adaptations, etc. Since localizer images are not optimized for image quality
but for scan time and scan range, statistical and model based approaches are preferred for the
segmentations.
Two methods are proposed: The first is a method for reconstructing liver shape, position
and orientation from a set of stacked sparse 2D localizer images. For this purpose an active
shape model, which represents the average liver shape and its major modes of statistical varia-
tion, is created from a set of liver segmentations from 3D training data images. Corresponding
landmark vertices on the surfaces are found by remeshing the surfaces using conformal map-
pings in the spherical domain and tuning of the correspondences by means of a measure based
on minimum description length (MDL) which describes the compactness of the model. Seg-
mentation of the liver from the stacked localizer image data is done by computing the model
instance of the active shape model which best fits the given image data. This is achieved by
iteratively computing optimal displacements of the landmark vertices based on image profiles
according to the statistics of local grey value appearance in the training image data. The in-
stance of the active shape model which best represents the displaced landmark vertices is found
by a projection to the linear space of the model. This yields the instance of the model which is
closest to the landmark displacements and ensures that the shape is valid.
The results of the segmentations from generated localizer images are compared to the
ground truths of the manual segmentations by comparing them according to four error met-
rics. The results show that the method can compete with standard 3D active shape model based
segmentation approaches, though with less precision due to lower input image quality.
The second method presented in this thesis is an automatic anatomical labeling or multi-
organ segmentation of anatomical structures from FastView images, which is a modern MR
imaging protocol producing 3D localizer images by acquiring 2D slices during continuous
movement of the patient table. The segmentation is based on a statistical atlas of the human
body created from an image data set of representative FastView images from a group of volun-
teers. The statistical atlas on the one hand consists of a statistical model of deformation, which
can be used to create valid instances of deformation fields to warp unseen image data sets to
the mean shape of the atlas. On the other hand, the atlas contains a statistical model of grey
value appearance, which can be used to create instances of atlas images with the mean shape
and valid grey value appearance. Both statistical models can be used to register an unseen data
set by finding the instances of the deformation and local model of appearance, which maximize
the similarity between the atlas with the grey values from the local model of appearance and the
unseen image, which is warped according to the deformation model. Propagating anatomical
labels from the atlas to the registered image then yields the anatomical labeling of the image
data. Since the registration process requires optimizing a multi-parameter cost function with
various degrees of freedom and computing its derivatives and since each evaluation of the cost
function requires a 3D image warp, performance speed-up strategies have to be considered. It
was decided to shift the statistical models of deformation and appearance completely to the
GPU (Graphics Processing Unit), which permits computing the complete non-rigid registra-
tions with GPU hardware acceleration, including the calculations of the instances of both the
deformation field and grey value appearance which require evaluating the linear models by
matrix multiplications and vector additions. The running times of the registration process are
about 10 to 30 seconds which is the same magnitude as the image acquisition itself. Running
times of this magnitude allow for practical usage of the method in clinical MR routine.
Validation of the atlas based segmentations is done by manually placing corresponding
landmarks at descernible points in the input FastView image data sets and the atlas. The dis-
placements between the landmarks in the registered data sets and the atlas are taken as an error
criterion for the exactness of the method and can be compared to the reproducibility of the
manual landmark placements.