Neurodegenerative and cerebrovascular diseases, and mental disorders, are the leading
cause of physical and mental disability in the population of developed countries and thus
have a huge and increasing socio-economic impact. Recent researches show that these
diseases and disorders also affect the physical structure of human brain, e.g. through
neuronal loss that can be observed as atrophy of normal brain tissues or the occurrence
of scars or lesions in the brain tissues. Currently, magnetic resonance (MR) tomographic
imaging is by far the most sensitive technique for visualizing normal and pathological
brain structures, but is also being increasingly used for their quantification. This Thesis is
focused on the problem of automated delineation or segmentation of brain MR images
into normal and pathological structures. Accurate, reliable and computationally efficient
segmentation methods are nowadays required in clinical routine so as to extract various
neuroimaging biomarkers, an important paraclinical tool used to characterize, monitor
and optimize treatment of several neurological and cerebrovascular diseases and mental
disorders. Segmentation of brain MR images can be performed manually by delineating each of the structures of interest, however, this task is cumbersome, time-consuming, expensive,
but most of all subjective and thus unreliable. Especially in large clinical trials that involve
processing of a large number of MR images, there is a need for accurate, reliable and
computationally efficient automated segmentation methods so as to deliver timely and
consistent quantitative measurements or biomarkers. Nevertheless, automated segmentation
is difficult and may be hampered due to numerous sources of variability in MR images
that are inherent to brain MR imaging, e.g. intensity inhomogeneity that may depend on
the imaged object, differences in multi-center MR image acquisition protocols, anatomical
variability across population and heterogeneity of pathology manifestation across different
patients. This Thesis advances state-of-the-art in the field of automated brain MR image
segmentation in several major aspects. Models of multi-sequence MR intensity of normal
brain structures, which capture the aforementioned MR intensity variabilities, and accurate
and robust estimation of the models are a crucial part of most state-of-the-art segmentation
methods. In the Thesis a novel method is proposed for estimation of a mixture model of
normal brain structures that is robust to unbalanced samples with outliers. The method is
particularly suited for the estimation of MR intensity models and thus for brain structure
segmentation methods, while its main advantage is high tolerance to the presence and
large variations of pathological structures. Secondly, a novel model of otherwise complex
whole-brain multi-sequence MR intensities is proposed that employs spatial stratification of
the complex model into several simplified models and their independent, robust estimation.
Subsequent recombination of the obtained simplified models resulted in accurate wholebrain
MR intensity model. Thirdly, a method for segmenting normal and pathological
brain structures based on locally-adaptive MR intensity model of normal-appearing brain
structures is proposed. The main advantage of the method is its robustness against different
imaging artifacts and anatomical variations due to a successful compilation of adaptive local
modeling and robust model estimation. Finally, a quantitative and comparative evaluation
of the proposed and several state-of-the-art methods for segmenting normal-appearing
structures and white-matter lesions in MR images of multiple sclerosis patients, having
different disease severity characterized by total lesion load, revealed that herein developed
methodological contributions substantially improve the performance of segmentation of
both normal appearing and pathological brain structures. Based on the observed accuracy,
reliability and efficiency the proposed methods seem as good tools for the extraction of
neuroimaging biomarkers.
|
