3D liver segmentation from abdominal computed tomography scans based on a novel level set model

Research Doctorate - Doctor of Philosophy (PhD)The liver is one of the most important organs in the human body. It carries out a variety of functions including filtering the blood, making bile and proteins, processing sugar, breaking down medications, and storing iron, minerals, and vitamins. However, the liver is prone to many diseases such as hepatitis C, cirrhosis, and cancer. As computer science and technology advances, computer-aided surgical planning systems have played an important role in the diagnosis and treatment of liver diseases. These systems can present the structures of various liver vessels, generate resection proposals, offer 3D visualizations, provide surgical cutting simulations, and shorter planning times. However, among these systems, one of the most challenging issues is the accurate segmentation of the liver from its surrounding organs in computed tomography images. Factors contributing to the challenge in carrying out accurate liver segmentation include the similar intensity values between adjacent organs, geometrically complex liver structure, and the injection of contrast media that causes all tissues to have similar gray-level values. Several artefacts of pulsation and motion, and partial volume effects, also increase difficulties for automatic liver segmentation in computed tomography images. Significant individual variations in shape and volume of the liver also add to the difficulties. Therefore, liver segmentation from medical images remains an open problem.In this research, we aim to perform accurate and automatic 3D liver segmentation from the latest multi-slice spiral/helical computed tomography (CT) scans, an achievement which would be very useful for computer-aided surgical planning systems. The development and evaluation of a clinically applicable segmentation algorithm, and its integration into software that could be used by medical experts, represents the major focus of the thesis. Level set methods have been widely used in medical image segmentation and perform well in segmenting irregularly shaped objects such as the liver. However, level set methods fail to segment meaningful objects from images if the objects are occluded by other objects, if some parts have low contrast (or are even missing), or if the target object has similar intensity values to adjacent objects. Since all these factors exist in the case of the liver, classical level set methods are not well suited to accurately segment the liver from abdominal CT scans. In this thesis, the enhanced level set method has been modified to make it suitable for segmenting the liver from an abdominal CT scan. We have improved the level set method to enable segmentation of the liver parenchyma from CT images by introducing a priori knowledge about the liver into the level set framework. These improvements make it possible to distinguish unclear liver boundaries, prevent surrounding organs from confusing the boundaries, and enhance segmentation performance. An important aspect of our improvements is that implementation of the necessary prior knowledge is not long or difficult compared to other segmentation methods. In initial exploratory work, the novel liver segmentation algorithm we first developed used the level set method together with an intensity prior (IP). The IP model improved the level set method by adding a priori statistical knowledge about the intensity distribution inside and outside the liver to the level set framework. The main merits of this approach were found to be its strong ability to dynamically guide the direction of the evolving contour and prevent it from leaking into regions with unclear boundaries. Examples of applying the proposed IP algorithm on real computed tomography images are presented. We show that the proposed method can deliver superior segmentation compared to the distance-regularized level set (DRLS) method. The average accuracy values for the IP model and the DRLS model are 99% and 89%, respectively. However, the IP model does have some limitations. We need to train the algorithm on liver slices that have a very similar intensity distribution to the target. This indicates that the statistical learning applied a priori in the training stage cannot be generally transferred to a large range of liver slices. Consequently, the method is not capable of segmenting a sequence of liver slices and building a complete liver volume. This motivated us to develop a liver segmentation algorithm which used the level set method together with density matching and a shape prior (DMSP). The DMSP model we developed provides accurate and automatic 3D liver segmentation from abdominal CT images. The algorithm is novel in that it combines density matching with prior knowledge about the liver shape. Density matching is a tracking method which maximizes the Bhattacharyya similarity measure between the photometric distribution inside the evolving curve and a model photometric distribution learned a priori. Density matching provides adaptive shrinkage or expansion to the evolving contour, while the shape prior improves robustness of the density matching and discourages the evolving contour from exceeding liver boundaries at regions which are unclear. For the purpose of comparison, we improved the IP model by adding a shape prior to its framework, producing an intensity prior and shape prior model, or IPSP model. However, even with this modification, the learning of the a priori statistical model applied during the training stage could still not correctly allow a liver volume to be reconstructed from a sequence of liver slices. Comparison experiments have shown that the DMSP model outperformed the IPSP model and performed well for all the investigated liver cases in our test data. The average overlap values for the IPSP model and the DMSP model were 76% and 91%, respectively. We compared the DMSP model with several other reported methods: the density matching (DM) model, the overlap prior (OP) model, and the DRLS method. Comparisons showed that the proposed method achieved better performance than any of these aforementioned approaches. The proposed method was shown to be more effective in overcoming over- and under-segmentation problems. The average overlap values of segmentation (compared to the ground truth) were estimated to be 69%, 77%, 63%, and 93% for the DM model, OP model, DRLS model, and DMSP model, respectively. Since the DMSP model achieves better performance than previous analogous studies, it has the potential to be used in clinical practice or in a computer-aided surgical planning system