Study Of Segmentation, Texture Feature Extraction And Classification Methods Of Chest Radiographs

X-ray diagnosis is one of the most important radiography diagnosis methods. Since many human organs exist in thorax, more than 40 percent of radiography imaging diagnosis is chest radiograph diagnosis. While the original chest radiographs have many disadvantages: low contrast, organ overlapping and fuzzy boundaries in the radiograph. These problems badly affect the clinic value of the chest radiographs. Thus the study of chest radiograph processing method has great medical value and application meaning. This paper mainly studies four aspects of chest radiograph processing technology:preprocessing and enhancement, segmentation, texture feature extraction and classification of chest radiographs.The preprocessing part consists of histogram equalization, median filtering and morphological filtering. The aim is to improve the global contrast and minimize the noise of chest radiograph. The Gassian-Laplacian pyramid enhancement method is adapt in the enhancement part and the experiment result shows that the chest radiograph quality after enhancing improved greatly.Next in segmentation part the lung contour is extracted. Here threshold method, derivative method and active shape model method are introduced and applied. The Otsu method is selected to get the best threshold in maximizing the inter-class variance of the image histogram. The experiment shows that this method acquired an enhanced chest radiograph. Derivative method first approximately find the area occupied by the lung contour, then find the lung contour points according to the maxima and minima of local first order derivative of gray scale. This method can relatively precisely estimate the lung contour, but was sensitive to noise and the division of approximate area of lung contour is a key premise operation. Active shape models method first labels the training set, aligns the training set, then gets the lung contour local gray scale statistics information of training set. When doing segmentation, an initial contour of the object can be gotten from training set, and iterated to adjust the initial object contour by calculating the gray scale statistics information's distance between the initial contour in the image and the training set, finally find the real lung contour. This method performs well in the lung contour segmentation.The texture feature extracting part is based on the experiment result of image enhancement and segmentation. Here in this paper fractal dimension and wavelet are chose to be the texture feature of the image. Before extracting the fractal feature, some image transform is done to the original image to get series of sub-images. These sub-images can reflect the details and direction information of the original image. As for wavelet feature, the two dimension wavelet transform is performed four times to get a series of sub-images, then calculating the gray scale mean and variance of the sub-images as the wavelet feature. The fractal and wavelet feature can reflect the image details in some extent. The aim of chest radiograph classification is to distinct the normal radiographs to the lung cancer ones. Support Vector Machine(SVM) is tried to do this job by using the fractal and wavelet feature, then cross-validation is applied to validate the precision of classification method. The fractal feature performs better in the experiment. What's more, a chest radiograph processing system is realized, and the system can be applied to do the enhancement, histogram equalization and threshold segmentation et al to chest radiograph.The chest radiograph processing methods adapted in this paper perform well and each part can be applied to practice. As the final aim, the image classification works well here, but the algorithm still has room for improvement.