Download files
Abstract
High Resolution Computer Tomography (HRCT) is a valuable imaging modality for detection and characterization of lung diseases. Manual assessment of disease findings and severity on lung HRCT images poses difficult problems, because the radiographic patterns observed are often varied and subtle. Labelling HRCT lung images is tedious, time consuming and error prone. Within this context, the thesis introduces computer-aided techniques based on multi-directional multi-resolution analysis of lung HRCT images and machine learning in order to classify diffuse lung disease (DLD) patterns: emphysema, honeycombing and ground glass opacities automatically. While a large amount of work has been conducted in the field, it is still an extremely active topic of research, with many problems still unsolved.
The main contributions of this thesis are to exploit multi-directional and multi-resolution analysis of lung HRCT images for classification of DLDs. First, the feature extraction method using wavelet and contourlet transform are investigated. A key contribution is to apply generalized Gaussian density model for a more accurate representation of coefficients of the outputs of wavelet and contourlet sub-bands of lung HRCT images. The proposed approach is shown to experimentally outperform other state-of-the-art methods in the literature.
Another approach is to use scale-space theory for multi-resolution modelling of lung HRCT images. Although the overall results of the method obtained is at the level of “acceptable”, it also proves the important role of scale-space features in classification of DLDs. Combination of the features from wavelet, contourlet and scale-space method is then investigated. A novel multiple kernel learning approach is proposed to optimally combine different feature spaces for classification of DLDs. The proposed method eliminates irrelevant features and selects relevant features by the score of individual feature through a primal formulation involving L2-norm regularization. The experimental results show the improvement in discriminating between DLDs, especially with two challenging patterns: ground glass opacity and honeycombing. Furthermore, the dimension of feature vector obtained using representations with non-zero weight values is significantly reduced or the complexity of the method is decreased.
In the final part of this thesis, the application of sparse representation based method to classification of DLDs in lung HRCT images is investigated. The sparse-representation model composed of three steps: dictionary learning, sparse coding and spatial pooling. After extracting the local features, the dictionary was learned by performing the K-SVD and Orthogonal Matching Pursuit (OMP) iteratively until the reconstruction error smaller than the stopping ruler. Then the original local features were represented by OMP using the learned dictionary. Finally, spatial pooling was adopted to combined the representation coefficients of the same ROI into a descriptor that is input to a SVM classifier. The performance of the method is comparable to the method using multiple kernel learning and it opens a new direction for future research of classification of DLD patterns on lung HRCT images.
The thesis focuses on multi-directional, multi-resolution analysis of lung HRCT images for classification of DLDs and makes significant contributions to lung CT image analysis.