Pattern Recognition-Based Analysis of COPD in CT
PhD-defense by Lauge Sørensen
Computed tomography (CT), a medical imaging technique, offers a detailed view of the human body that can be used for direct inspection of the lung tissue. This allows for in vivo measurement of subtle disease patterns such as the patterns associated with chronic ob-structive pulmonary disease (COPD). COPD, also commonly referred to as ``smokers' lungs'', is a lung disease characterized by limita-tion of the airflow to and from the lungs causing shortness of breath, and the disease is expected to rank as the fifth most burdening dis-ease worldwide by 2020 according the World Health Organization. Though the components of COPD can be seen as a family of texture patterns in CT, commonly employed CT-based quantitative measures in the clinical literature are rather simplistic and fail to take the texture appearance of the lung tissue into account.
This thesis presents several methods for texture-based quantification of emphysema and/or COPD in CT images of the lungs. The methods rely on image processing and pattern recognition. The image processing part deals with characterizing the lung tissue texture using a suitable texture descriptor. Two types of descriptors are considered, the local binary pattern histogram and histograms of filter responses from a multi-scale Gaussian derivative filter bank. The pattern recognition part is used to turn the texture measures into a quantitative measure of disease. Different classification systems are considered, and we will in particular use the pattern recognition concepts of supervised learning, multiple instance learning, and dissimilarity representation-based classification.
The proposed texture-based measures are applied to CT data from two different sources, one comprising low dose CT slices from subjects with manually annotated regions of emphysema and healthy tissue, and one comprising volumetric low dose CT images from subjects that are either healthy or suffer from COPD. Several experiments demonstrate that it is clearly beneficial to take the lung tissue texture into account when classifying or quantifying emphysema and/or COPD in CT. Compared to common clinical CT-based measures, the texture-based measures are better at discriminating between CT images from healthy and COPD subjects, they correlate better with the lung function of the subjects, they are more reproducible, and they are less influence by the inspiration level of the subject during CT scanning - a major source of variability in CT.
Chairman: Associate Professor Francois Lauze, DIKU
Member 1: Professor Dr. Horst Bischof, Institute for Computer Graphics and Vision, TU Graz
Member 2: Professor Dr. Joseph Reinhard, Department of Biomedical Engineering, University of Iowa
For an electronic copy of the thesis, please contact Dina Riis Johannessen, email@example.com