The Study On Image Segmentation And Feature Recognition In Chest Radiographs

The chest has ever been described as the mirror of health and disease, because an enormous amount of information about the condition of the patient can be extracted from a chest film, and therefore, the chest radiographs is greatly important in clinic. However, the image quality of chest radiographs is rather limited, such as low resolution and lots of superimposed anatomical structures, which make almost none of the organizations has specific boundaries. So the technology of computer-aided diagnosis (CAD) is highly required.This paper is based on the clinical application of CAD. Several key process and analysis of images are investigated here, including lung segmentation, rib segmentation and nodule detection.A comprehensive conclusion of lung segmentation methods is carried out. Several existing methods are implemented, including rule-based detection of lung contours, geometric models and Snake models. After the defects of those methods are concluded, a compositive method is proposed in this paper. A strategy of segmenting step by step is adopted in this method. First, according to the expanding ability to the dark region of soft erosion operator, image filtering is carried out once by it, so that a high sensitivity of segmentation can be achieved. Second, a clustering algorithm is adopted so as to get finer segment. At last, small regions are sifted and grouped based on rules to get the final result. Compared with other methods, both higher sensitivity and accuracy are achieved by this method. In order to extract the lung contour, a model based method——Active Shape Model (ASM) is investigated in this paper, and the extraction of lung contour is implemented by ASM. After analyzing the defects of classical ASM and multi-resolution ASM, a modified ASM is developed in this paper, that is, the sampled contours are filled to be the desired outputs of ANN at the same time an ASM is being trained, and so the gray appearance model is based on the output of neural net. Compared with other methods, the segmentation sensitivity, accuracy and execution speed are all optimized by modified ASM.The rib segmentation is investigated in depth. Several existing methods are also implemented, including K-means, Gaussian surface threshold, iterated contextual pixel classification and Hough transform. The usage of Hough transform is modified in this paper. The Hough transform is carried out after thinning the image, so one rib is expressed by only a single line. Then the morphological dilation is implemented to obtain the actual ribs. This method avoids distinguishing the upper and lower ribs that is necessarily required when using Hough transform directly. The experiment results indicate that this method achieves higher accuracy and is more understandable.All kinds of lung nodule detection algorithms are investigated. An integrated nodule detection system based on rib information is developed in this paper. The system treats the ROIs differently with the consideration whether there are ribs crossing the centres of ROIs. The original method that includes only a single time of classification is improved and the complexity of training is simplified. According to the simulation, the number of false positives in proposed system is a little smaller than other systems.