Abstract
This research thesis
addresses the problems of microscopic pathology analysis. It presents a general
framework for computer-aided diagnostics. The aim of the framework is
unification and improvement of the pathological examination routine. The
framework combines telepathology with computer-aided diagnostics algorithms.
The framework targets image acquisition and interpretation stages. The image
acquisition subsystem solves problems related to microscopical slide
digitization such as biomedical image registration, data representation, and
processing. The interpretation subsystem uses a support vector machine
classifier together with a feature selection for computer-aided diagnostics.
The histopathological and cytopathological systems for computer-aided
diagnostics are implemented using the presented framework.
In microscopy, regions
of interest are considerably smaller than the whole slide area. Various
microscopy related medical applications, such as telepathology and computer
aided diagnosis, are liable to benefit greatly from automatic determination of
field of view coordinates. In this thesis we present a method for image-based
positioning on a microscope slide. The method is based on localization of a
microscopic query image using a previously acquired slide map. It uses
geometric hashing, a highly efficient technique of object recognition. The
algorithm exhibits high tolerance to possible variations of microscopic field
of view due to slide rotations and illumination changes.
Problems of
computer-aided diagnostics systems based on analysis of color or gray-scale
images are addressed. The first problem is large color variance of the same
dyes attached to the same specimen types. The second problem concerns the
physical overlapping of different kinds of objects resulting in color mixtures.
The algorithm for automated decomposition of microscopic hyperspectral
biological images is proposed. The decomposed compounds are described by their
spectral characteristics and optical density. The multiplicative physical model
of image formation in transmission light microscopy is presented, a reduction
of a hyperspectral image decomposition problem to a Blind Source Separation
(BSS) problem is justified, and a method for hyperspectral restoration of
separated compounds is provided. The obtained results may be used for improving
automatic microscope hardware calibration and computer-aided diagnostics.
The method for
decomposition of hyperspectral images is extended to a case of multichannel
noisy signals separation. The cases when the source signals have a sparse
representation in some dictionary are considered. The dependency of
decomposition coefficients is modeled by the joint sparsity prior. The joint
sparsity prior is explored in the context of BSS problem and denoising problem,
where the mixing matrix and the sources are unknown.
