Robust texture descriptors and algorithms for micro-calcification detection and breast density estimation in mammograms.
| dc.contributor.author | Li, Haipeng | |
| dc.date.accessioned | 2021-08-24T02:34:32Z | |
| dc.date.available | 2021-08-24T02:34:32Z | |
| dc.date.issued | 2021 | en |
| dc.description.abstract | This thesis proposes, implements and evaluates novel algorithms used for processing and analysing mammograms, with two distinct research aims: microcalcifications (MCs) detection and mammographic density (MD) classification. Mammography is the most widely used imaging technology for breast cancer screening, which outputs mammogram images for reading and interpretation by radiologists. The work presented in this thesis deals with robust and effective methods which can be used in computer aided detection systems to automatically detect diagnostic features and to assist radiologists in the interpretation of mammograms. Accurate MCs detection is difficult due to their heterogeneous features: varied size, shape, and local contrast. This thesis presents a MCs detection framework with two stages: candidate MCs detection and false positive number reduction. In the first stage, a linear structure detector, multifractal features, and a support vector machine are applied to develop a MCs detector, which produces initial MCs detection results. In the second stage, a Weber’s law-based multifractal measure is developed to enhance the texture features related to MC spots, and a patch-wise convolutional neural network (CNN) is proposed to classify the detected MC spots in the first stage into true positive or false positive groups. The false positive number is reduced significantly using the proposed scheme. A digital mammogram dataset, INbreast, is used to test the proposed detection framework, and experimental results show that the overall detection accuracy is higher than that obtained using other methods in the literature. Mammographic density is a key biomarker of breast cancer, which indicates the risk of women developing breast cancer in the future, and MD evaluation plays an important role in preventing breast cancer. This thesis builds four robust and effective classification models for grouping mammogram images into different density categories, with an in-depth investigation of extracting and analysing relevant image features of mammograms. The image features used for MD evaluation in this thesis can be divided into three groups: cascaded image features, robust texture descriptors, and texture feature based spatial information. For using the cascaded image features, this thesis presents a patch-wise classification method using multifractal spectrum, histogram information and Chi-square test statistic (model 1). In addition, histogram information based on both multifractal features and local binary patterns are incorporated to form a new feature set to be used in a density classification model (model 2). An autoencoder network and principal components analysis are adopted to compress the feature space for improving the efficiency of the classifier. This thesis conducts a complete study based on binary encodings of local texture patterns, and proposes a MD classification method (model 3) by developing and testing novel texture descriptors. A robust rotation invariant texture descriptor is developed based on local quinary patterns using high transition numbers. This thesis also explores spatial distribution characteristics of feature points based on the above novel texture descriptor to capture rich image representations and to improve the classification accuracy. Baddeley’s K-inhom method is employed to describe the spatial information of feature points in mammograms and a new texture feature vector, K-spectrum, is constructed and used in the classification model (model 4). The proposed classification method is tested on two mammogram datasets, INbreast and MIAS. Quantitative assessment and comparative analysis are given, which demonstrate the superiority and robustness of the proposed model. Finally, this thesis summarises the main contributions and research accomplishments, and also outlines extensions and future directions. | en |
| dc.identifier.uri | https://hdl.handle.net/10092/102326 | |
| dc.identifier.uri | http://dx.doi.org/10.26021/11375 | |
| dc.language | English | |
| dc.language.iso | en | |
| dc.publisher | University of Canterbury | en |
| dc.rights | All Right Reserved | en |
| dc.rights.uri | https://canterbury.libguides.com/rights/theses | en |
| dc.subject | Mammogram, Breast cancer, Microcalcification detection, Breast density classification, Multifractal analysis, Texture feature extraction, Image enhancing technique, Convolutional neural network, Autoencoder network, Local binary patterns, Local quinary patterns, Rotation invariant uniform encoding, Spatial distribution analysis, Feature selection | en |
| dc.title | Robust texture descriptors and algorithms for micro-calcification detection and breast density estimation in mammograms. | en |
| dc.type | Theses / Dissertations | en |
| thesis.degree.discipline | Computer Science | en |
| thesis.degree.grantor | University of Canterbury | en |
| thesis.degree.level | Doctoral | en |
| thesis.degree.name | Doctor of Philosophy | en |
| uc.bibnumber | 3087087 | |
| uc.college | Faculty of Engineering | en |