Abstract
Image analysis techniques have been broadly used in computer aided diagnosis tasks in recent years. Computer-aided image analysis is a popular tool in medical imaging research and practice, especially due to the development of different imaging modalities and due to the increased volume of image data. Image segmentation, a process that aims at identifying and separating regions of an image, is crucial in many medical applications, such as in identication (delineation) of anatomical structures and pathological regions, providing objective quantitative assessment and monitoring of the onset and progression of the disease. Multidetector CT (MDCT) allows acquisition of volumetric datasets with almost isotropic voxels, enabling visualization, characterization and quantication of the entire extent of lung anatomy, thus lending itself to characterization of Interstitial Lung Diseases (ILDs), often characterized by non-uniform (diffuse) distribution in the lung volume. Interpretation of ILDs i ...
Image analysis techniques have been broadly used in computer aided diagnosis tasks in recent years. Computer-aided image analysis is a popular tool in medical imaging research and practice, especially due to the development of different imaging modalities and due to the increased volume of image data. Image segmentation, a process that aims at identifying and separating regions of an image, is crucial in many medical applications, such as in identication (delineation) of anatomical structures and pathological regions, providing objective quantitative assessment and monitoring of the onset and progression of the disease. Multidetector CT (MDCT) allows acquisition of volumetric datasets with almost isotropic voxels, enabling visualization, characterization and quantication of the entire extent of lung anatomy, thus lending itself to characterization of Interstitial Lung Diseases (ILDs), often characterized by non-uniform (diffuse) distribution in the lung volume. Interpretation of ILDs is characterized by high inter and intra-observer variability, due to lack of standardized criteria in assessing its complex and variable morphological appearance, further complicated by the increased volume of image data being reviewed. Computer-Aided Diagnosis (CAD) schemes that automatically identify and characterize radiologic patterns of ILDs in CT images have been proposed to improve diagnosis and follow-up management decisions. These systems typically consist of two stages. The first stage is the segmentation of left and right Lung Parenchyma (LP) region, resulting from lung eld segmentation and vessel tree removal, while the second stage performs classication of LP into normal and abnormal tissue types. The segmentation of Lung Field (LF) and vessel tree structures are crucial preprocessing steps for the subsequent characterization and quantication of ILD patterns. Systems proposed for identication and quantification of ILD patterns have mainly exploited 2D texture extraction techniques, while only a few have investigated 3D texture features. Specifically, texture feature extraction methods that have been exploited towards lung parenchyma analysis are: first order statistics, grey level co-occurrence matrices, gray level run length matrices, histogram signatures and fractals. The identication and quantification of lung parenchyma into normal and abnormal tissue type has been achieved by means of supervised classication techniques (e.g. Articial Neural Networks, ANN, Bayesian classier, linear discriminant analysis (LDA) and k-Nearest Neighboοr (k-NN). However, the previously proposed identification and quantification schemes in-corporate preprocessing segmentation algorithms, effective on normal patient data. In addition the effect of the preprocessing stages (i.e. segmentation of LF and vessel tree structures) on the performance of ILD characterization and quantification schemes has not been investigated. Finally, the complex interaction of such automated schemes with the radiologists remains an open issue. The current thesis deals with identification and quantification of ILD in lung CT. The thesis aims at optimizing all major steps encountered in a computer aided ILD quantification scheme, by exploiting 3D texture feature extraction techniques and supervised and unsupervised pattern classication schemes to derive 3D disease segments. The specic objectives of the current thesis are focused on: • Development of LF segmentation algorithms adapted to pathology. • Development of vessel tree segmentation adapted to presence of pathology. • Development of ILD identification and quantification algorithms. • Investigation of the interaction of an ILD identification and quantification scheme with the radiologist, by an interactive image editing tool.
show more
