Adaptive Segmentation Of Serial Chest Radiographs Using Dynamic Information On A Patient

Medical imaging is among the most important diagnostic tools in modern medicine. Today, a wide variety of (three-dimensional) imaging techniques is available, and many types of normal examinations have been, or are about to be replaced by computed tomography (CT) and/or magnetic resonance imaging (MRI) because of their high facility. While each has its strong points, the traditional radiological procedure is not to be replaced, which makes up at least a third of all exams in a typical radiology department. Segmentation of two-dimenstional radiographs is the topic of this thesis.When treating end-stage renal disease, an evaluation of the variation of the size of cardiac shadow from month to month in a simple chest radiograph remains crudial for treatment of the patients. This is because the evaluation of a patient's dry weight is usually based on the cardiac size measured during each dialysis session, and the misevaluation of dry weight often results in a fatal illness such as patient's death. Since segmentation of the lung fields in chest radiographs provides a means to measure the cardiac size, it is important to accurately segment the lung fields from regularly captured serial chest radiographs of each patient.Among the various methods about segmentation of lung field in chest radiographs, Model based method is the most prosperous one because it combines the prior of lung fields into the model. Concretely, model can capture human prior knowledge about anatomy, tissue appearance, modality characteristics, and efficiently employ this to infer functional or structural information from new data. As a model based method, active shape model is an appropriate start point applying to segmentation of lung field in chest radiographs.However, there exist several problems for original active shape model in segmenting chest radiographs. For example, general feature can not distinctively describe the feature of each point along boundary of lung fields. And also, shape model is not able to be adaptive to the boundary of lung fields. Therefore, this thesis provides corresponding means to improve the accuracy and robustness when using active shape model to segment the lung field in chest radiographs. These form the contribution in this thesis.A new feature descriptor, i.e. scale invariant feature transform (SIFT), is used to distinctively represent the rich information around each point along the boundary of lung fields. These image features can facilitate correspondence detection during the segmentation of lung fields in serial chest radiographs.Notice that the generic image features along the boundaries of lung fields in different chest radiographs coming from different scanning devices are not always consistent. Also there are some repeated patterns like ribs in the lung fields. All of these render generic features like edges not sufficiently distinctive to discriminate rib boundaries from lung field boundaries. Ideally, a distinctive local descriptor should be built for each point on lung field boundary in order to reliably differentiate it from other boundary points, thus helping detect correspondences during the deformable segmentation procedure.Therefore, a modified scale invariant feature transform local descriptor, more distinctive than the general intensity and gradient features, is used to characterize the image features in the vicinity of each pixel. This feature can improve the performance of active shape model by accurately locating the boundary of lung fields because of its facilitating the correspondence detection during segmentation of lung fields in chest radiographs.A new deformable model using both population-based and patient-specific shape statistics to segment lung fields from serial chest radiographs is presented. Formerly, when using deformable model to segment lung fields in chest radiographs from a patient, the image only scanned at current time-point was segmented and all images scanned during the whole treatement period were not considered to predict and deal with the image at current time-point. Therefore, the segmentation results were not enough to reflect the progress of health in treatment a patient. Accordingly, a new deformable model is presented to adaptively segment lung fields in serial chest radiographs by using all dynamic information on a patient during whole treatement period.Concretely, a new deformable model using both population-based and patient-specific shape statistics to segment lung fields from serial chest radiographs is presented. The deformable contour is constrained by both population-based and patient-specific shape statistics, and it yields more robust and accurate segmentation of lung fields for serial chest radiographs. In particular, for segmenting the initial time-point images, the population-based shape statistics is used to constrain the deformable contour; as more subsequent images of the same patient are acquired, the patient-specific shape statistics online collected from the previous segmentation results gradually takes more roles. Thus, this patient-specific shape statistics is updated each time when a new segmentation result is obtained, and it is further used to refine the segmentation results of all the available time-point images. Experimental results show that the proposed method is more robust and accurate than other deformable models in segmenting the lung fields from serial chest radiographs.An automatic and robust two-dimensional diagnostic measure is provided.The development of modern segmentation technique makes it possible to automaticly compute traditional diagnositic measure, and furthermore, it facilitates the computation of the unexpected diagnostic measure in the previous time. Therefore, a more interesting and informative diagnostice measure, i.e., two-dimensional cardiothoracic rate (2D-CTR) is provided. And the medical statistical methods validate that the 2D-CTR has strong agreement with original CTR, i.e., 1D-CTR. In addition, it is worth noting that 2D-CTR seems more useful since it can be still calculated in the cases that the left or the right cardiac margin is blurred, or sloping gently to diaphragm, or having a fatty pad at the apex. However, it is difficult to measure 1D-CTR in these cases. Therefore, 2D-CTR is robust and can be potentially more useful in computer-aided diagnosis.Finally, it is worth noting that experimental data in this study is provided by Hidaka hospital in Takasaki of Japan which will continue to provide chest radiographs for polishing our algorithm. At the same time, our automatic two-dimensional diagnostic measure is evaluated as a robust measure in clinical environment by Dr. Hidenori Matsuo of Hidaka hospital in Takasaki of Japan.