Research On Thoracic CT Scans Segmentation For Pulmonary Diseases Detection

As lung cancer is now the world’s deadliest cancers, its early detection, diagnosis andtreatment are important means to improve the survival of patients with lung cancer. Theearly symptoms of the lung cancer are in the form of small pulmonary nodules, so theearly detection of lung nodules and treatment for it in time are significant to save thepatients. With the progress of medical CT (computed tomography) technology and theincrease of image data gotten in the medical CT examination, it provides moreinformation about organs and tissues, and at the same time gives a great burden to thedoctors. In order to improve the doctor’s working efficiency, reduce the workingstrength, overcome the influence of artificial factors in the reading, and improve thedetection rate of lung disease, it is an urgent task to study the automatic detection oflung diseases. And for it, its primary task is to segment the organs or tissues in thethoracic CT scans correctly. Therefore, the dissertation takes the goal to automaticallydetect pulmonary diseases based on the thoracic CT scans, and studies the segmentationmethod of every lung tissues through the combination of the anatomical structure of thelungs and the knowledge of the CT imaging mechanism. It generally involves thefollowing contents:As segmenting lung correctly from the thoracic CT scans is an important step and thefirst priority for the analysis, diagnosis and treatment of pulmonary diseases, it isproposed a segmentation algorithm which integrated with the homomorphismprocessing and the CT threshold segmentation. Under the guidance of lung anatomicknowledge model, it is able to segment the lung fields with lobar fissure, thin junctionbetween the lung fields, and indentation of the blood vessel and bronchi-wall in thehilus pulmonisvery well.For the regular lung fields with the juxta-plural lung nodules of high density, it isproposed a method which an active contour model constrainted of the prior shape cansegment out the correct lung fields. It is to classify the lung field shape, decomposethese shapes into the PCA (Principal Component Analysis) shape vectors, and followedby the prior model vectors the improved active contour can fit the boundary and segments out the correct ones. Research results show it is completely feasible that themethod segments the lung fields whose edge affected by high density pathology.In order to solve the segmentation problem better, the dissertation has studied anothersegmentation method for the lung fields by using the similar characteristics. To investthe shapes similarity manifold of the lung fields in a lung and the method to construct amanifold with the PCA, based on the manifold relationship among the lung fields, theaffected lung fields can be reconstructed by the manifold interpolating, and it registersto decrease the error and segments out the lung field. Results show that the method is aneffective method, and the sensitivity, the specificity and the accuracy illustrate that itcan segment out the correct ones except the apex and the bottom of lung.Due to a lot of noise in the plain CT scans, low-dose CT scans, after a deepinvestigation into the composition filtering, medical image enhancement, imagesegmentation method and the enhancing ability to subtle details of the fractional orderdifferential, it is proposed a pulmonary blood vessels segmentation method. It segmentsevery local region of the thoracic CT scans enhanced by fractional differential operatorwith each optimal threshold. Studies have shown the method can effectively extract theblood vessels. Compared with the traditional pulmonary vascular segmentation ones, ithas more accurate segmentation ability of pulmonary vascular.For the influences of the noises of the imaging, the reconstructing of the image andthe partial volume effects, and the pathological parts which lead to the border ofdifference tissures or organs blurry, it is proposed a segmentation method which iscombination of the four neighborhood connection power pulse coupled neural networkand an active contour model with the prior shape to extract the candidate lung nodules.Studies results show that the method is a feasible, and an effective one to segment thecandidate pulmonary nodule.In the last part of the dissertation, it is elaborated the trends of detection for lungdisease in thoracic CT scans, as well as the next step to carry out research works. Thenits summarized the main content of the research work, the innovation points, anddiscuses further research problems in the future.