| The rapid industrial development has caused more and more serious environmental issues, especially air pollution, which led to higher incidence of a variety of lung diseases with higher mortality rates. Common lung diseases include: emphysema, pulmonary atelectasis, lung real variable, fibrosis, calcification, lumps or nodules, lung cancer, tuberculosis and pneumonia. Lung cancer has become the world's top cancer killer, and in recent years the incidence of lung cancer still continue to show an upward trend on a global scale, especially in China. The incidence of lung cancer is relevant with the ageing of the population, urban industrialization, pollution and unhealthy lifestyles. Among them, smoking and other unhealthy lifestyle is key factors led to lung cancer. However, lung cancer can be prevented and controlled if found in early stage. Noting that if there are recurrent symptoms such as chest pain, hoarseness, coughing up blood, or lung and shoulder pain, numbness in the fingers, headache, vomiting, you will need to take an X-ray screen as soon as possible to exclude the possibility of lung cancer, and get early treatment. In addition, early imaging findings of lung cancer are usually shown as pulmonary nodule, and ordinary digital chest X-ray films can check out 70?80 percent of small pulmonary nodules. Digital X-ray chest, a fast and convenient medical imaging method, especially with low radiation dose, has long been a widely available medical screening for lung diseases. Chest X-ray inspection has counted as 40 percent of the area of imaging diagnosis of pulmonary diseases, which is sufficient to demonstrate the application value of Chest X-ray inspection. But the number of patients is getting more and more, and if radiologists have to find lesions with the naked eyes and make judgment fully by their experience, at last the whole process will be very time-consuming, and no doubt it will result in low productivity, backlog of patient issues. However, with the development of computer technology and medical image processing technology, computer-aided diagnosis system has emerged in clinical practice. Due to the particularity of medical images, such as small differences in image contrast, occlusion of the adjacent tissues or organs, different imaging conditions and other factors, you need to do some pre-process before analysis and diagnosis of the disease, like automatic segmentation of lesions or vital organs. Due to the complexity of the image content, imaging method and imaging conditions, such as in individual differences, local effects, artifacts and noise caused by heart beats or breathing, with one segmentation method for segmentation of different organizations or lesions may not work. Therefore, chest X-ray automatic segmentation of lung field has long been a hot research area of Medical Image Segmentation. There are many segmentation algorithms for lung filed segmentation on chest X-ray, such as pixel-based classification, active shape model, non-rigid registration. Segmentation of lung field has been obtaining higher accuracy and speed. But still there are some difficulties:easily drawn into a local minimum at the edge of clavicles and ribs cage, the shoulder areas and effects of low contrast with surrounding organs, tiny diaphragm angle is very difficult to accurately segment and so on.The main contributions of this dissertation include:(1) Automatic segmentation of lung field on chest X-ray based on dense features matching. Accurate X-ray automatic segmentation of lung field is important for the computer-aided diagnostic system of lung diseases. Inspired by PatchMatch algorithm which has widely used on segmentation of natural images. The main idea of PatchMatch is, for each image patch in an image, search its most similar image patch from another image, and then reconstruct the input image with these similar image patches. PatchMatch algorithm is a simple and effective searching algorithm, using the consistency of the image, that is, between adjacent patches, the displacement vectors between their nearest neighbors may be the same, so the nearest neighbors can be propagated to adjacent areas. Meanwhile, at the initialization stage, each patch is assigned a nearest neighbor randomly from template image, there is high possibility that at least one patch obtain a reasonable nearest neighbor. PatchMatch algorithms can be divided into three steps:random initialization, propagation, random search. Random initialization refers to assign a random nearest neighbor from template image B for each patch in image A, a nearest neighbors field is obtained; propagation can propagate good nearest neighbor to surrounding region, gradually optimizing the nearest neighbors field; random search refers to assign a random nearest neighbors within searching window centered at the current nearest neighbors, optimized nearest neighbors again, and avoided local minimal at some extent. The idea of this paper is, for an input chest radiographs, the dense Scale Invariant Feature Transform descriptors and raw image patches are extracted as the local features for each pixel. Then, the nearest neighbors of the local features are quickly searched by dense matching directly from the whole feature dataset of reference images. The dense matching includes three steps:limited random initialization, propagation of nearest neighbor field, and limited random search. The last two steps are performed iteratively several times. The label image patches for each pixel are extracted according to the nearest neighbor field and weighted by the matching similarity. Finally, the weighted label patches are rearranged as the label class probability image of the input chest radiographs which can be threshold as the segmentation of lung fields. The Jaccard index of the proposed method on the public Japanese Society of Radiological Technology dataset can achieve 95.5%. High accuracy and robustness can be obtained by adopting dense matching of local features and label fusion to segment lung fields in chest radiographs, which surpasses that of the best state-of-the-art segmentation method for lung fields in chest radiographs.(2) Virtual dual-energy subtraction of digital radiography based on dense features matching. Dual-energy subtraction is a special X-ray imaging, and can generate images with organizational characteristics. Dual-energy subtraction refers to use of high and low levels of kilovolt-level voltage in a very short period of time to generate two exposures to the same imaging area, and obtain two chest X-ray images, after subtraction soft tissue and bone images are obtained. Two exposures within very short time interval and with patient holding breath after taking a deep breath, will effectively avoid motion artifacts due to respiratory movement. Studies have shown that dual-energy subtraction technique compared to the general Digital radiograph, has better showing performance on the structures of the chest, and help to improve the diagnosis of lung diseases. However, most of the conventional X-ray machine is not able to provide the ability to dual-energy subtraction, and due to the higher technical requirements such as tubes and the generator, only a handful of manufacturers have the ability to produce dual-energy subtraction X-ray machine. In addition, dual-energy subtraction technique will result in higher radiation dose than conventional DR. From a technical and safety considerations, and in order to overcome technical problems of hardware and reducing radiation doses patients receive, the concept of virtual dual energy subtraction has been proposed. The idea of this paper is:For an input chest digital radiograph, firstly retrieve a number of similar radiographs from the reference images group including radiographs, the corresponding dual-energy subtraction bone images and soft tissue images. Performing dense features matching between the input radiograph and radiographs from the reference images group, according to nearest neighbors fields extract similar bone patches from the corresponding bone images from the reference images group, finally adopt label fusion method to construct bone image of the input radiograph with similar bone patches, soft tissue image can be obtained after subtract bone image from input radiograph, completing the separation of bone/soft tissue of radiograph. |
PDF Full Text Request