| Computer vision and pattern recognition techniques have been fostered to solve many practical problems of diverse areas. Medical image analysis using machine vision and learning intelligence is one of the most sought-after fields. Computer vision addresses problems of the use of computers to detect, partition, represent, group, track, and interpret crucial primitives from given visual inputs. By contrast, pattern recognition is the study of distinguishing and recognizing different patterns represented with quantitative measurements. As a result, both of these two components usually present themselves in medical image analysis research work.In this dissertation, two parts of new theoretical work and one new segmentation method are proposed. Additionally, three medical image analysis problems are solved with intensive use of computer vision and pattern recognition techniques. In the first theoretical work, we demonstrate a new paradigm of modelling the optimal geometry fitting problem with deformable statistical models. Two novel ellipse-specific fitting algorithms are developed from this schema, along with an algorithm specifically devised to allow for an appropriate choice of methods optimally matched with the input data distribution. The second theoretical analysis work is related to connections between the generalized and the ordinary eigenvalue decomposition problems often presented in Linear Discriminant Analysis (LDA). The third contribution is illustrated by the work aimed at the development of a whole-slide microscopy image analysis system. A complete work flow of analysis, including establishment of image hierarchy, segmentation, feature representation, feature selection, feature extraction, classification, classifier combination, and decision evaluation, is developed. With this work, a new segmentation approach using the Fisher-Rao criterion, embedded in the generic Expectation-Maximization algorithm, is proposed for segmenting images. As an extension, a combined microscopy image analysis system is next developed to integrate into the computerized decision making process two prognosis components (stromal detection and differentiation grading) clinically crucial for pathologists. Both linear and non-linear methods are investigated for reducing the dimensionality of the feature space where data are classified by the classification and regression tree (CART) boosted with a Multi-Class Gentle Boosting method. Finally, the work of detecting multiple layer boundaries within human retina OCT images is presented. The developed algorithm is a hybrid approach consisting of a segmented model fitting process and the probabilistic relaxation labeling procedure. In the initial model fitting step, a geometrical model concatenating three parabolas with downward openings is developed. In the second stage, the fine tuning problem is formulated as the probabilistic relaxation labeling process that transforms the image to a relational graph consisting of attributed nodes. Experimental results demonstrate a good correlation between the sub-layer locations identified by the automated process and those manual markers. |
