Biomedical image analysis using multi-fractal and multiscale techniques.
Thesis DisciplineComputer Science
Degree GrantorUniversity of Canterbury
Degree NameDoctor of Philosophy
This thesis proposes, develops and evaluates different multifractal methods for detection, segmentation and classification of medical images. This is achieved by studying the structures of the image and extracting the statistical self-similarity measures characterized by the Holder exponent, and using them to develop texture features for segmentation and classification. The theoretical framework for fulfilling these goals is based on the efficient computation of fractal dimension, which has been explored and extended in this work.
This thesis investigates different ways of computing the fractal dimension of digital images and validates the accuracy of each method with fractal images with predefined fractal dimension. The box counting and the Higuchi methods are used for the estimation of fractal dimensions. A prototype system of the Higuchi fractal dimension of the computed tomography (CT) image is used to identify and detect some of the regions of the image with the presence of emphysema. The box counting method is also used for the development of the multifractal spectrum and applied to detect and identify the emphysema patterns.
We propose a multifractal based approach for the classification of emphysema patterns by calculating the local singularity coefficients of an image using four multifractal intensity measures. One of the primary statistical measures of self-similarity used in the processing of tissue images is the Holder exponent (α-value) that represents the power law, which the intensity distribution satisfies in the local pixel neighbourhoods. The fractal dimension corresponding to each α-value gives a multifractal spectrum f(α) that was used as a feature descriptor for classification. A feature selection technique is introduced and implemented to extract some of the important features that could increase the discriminating capability of the descriptors and generate the maximum classification accuracy of the emphysema patterns.
We propose to further improve the classification accuracy of emphysema CT patterns by combining the features extracted from the alpha-histograms and the multifractal descriptors to generate a new descriptor. The performances of the classifiers are measured by using the error matrix and the area under the receiver operating characteristic curve (AUC). The results at this stage demonstrated the proposed cascaded approach significantly improves the classification accuracy. Another multifractal based approach using a direct determination approach is investigated to demonstrate how multifractal characteristic parameters could be used for the identification of emphysema patterns in HRCT images. This further analysis reveals the multi-scale structures and characteristic properties of the emphysema images through the generalized dimensions. The results obtained confirm that this approach can also be effectively used for detecting and identifying emphysema patterns in CT images.
Two new descriptors are proposed for accurate classification of emphysema patterns by hybrid concatenation of the local features extracted from the local binary patterns (LBP) and the global features obtained from the multifractal images. The proposed combined feature descriptors of the LBP and f(α) produced a very good performance with an overall classification accuracy of 98%. These performances outperform other state-of-the-art methods for emphysema pattern classification and demonstrate the discriminating power and robustness of the combined features for accurate classification of emphysema CT images. Overall, experimental results have shown that the multifractal could be effectively used for the classifications and detections of emphysema patterns in HRCT images.