Research On Lung Nodule Automatic Detection

Lung cancer is the most common visceral malignancy, it ranks as the highest incidence and mortality rates of cancer in China recent years. The early manifestations of lung cancer are spherical or spherical nodules, whose diameters less than3cm. The key to improving patient survival is timely detection and treatment. CT images with higher resolution tissues, is the most effective means of imaging to detect and diagnose pulmonary diseases. It has an important value for detection and diagnosis of lung cancer, is widely used in screening for lung nodules. Due to the huge number of CT images, the complexity of the structure of the lung and the shapes and sizes of nodules, workload of doctors’ is very great. Doctors are prone to fatigue and may even result in misdiagnosis. Computer Aided Detection (CAD) system can detect lung nodules automatically, and give the location and associated quantitative information of lesions by processing and analysis the images of the patients’. CAD system has a very important significance. It not only greatly reduces the workload of doctors, but also helps doctors diagnose and treat and improves the efficiency.Depending on the appearance of pulmonary nodules in CT images, which can be divided into solid nodules and ground glass opacity (GGO) nodules. Solid nodules showed high brightness solid spherical object, GGO nodules are presented as fuzzy opacity. As the appearance of two types of nodules is different in CT images, the detection methods are very different. Currently, most researchers bias solid nodules, research on GGO nodule detection is less. This paper presents two types of nodules were appropriate detection methods, and nodular type of discrimination conducted a preliminary study. For two types of nodes, this paper proposed the corresponding detection methods individually and researched the type discrimination of nodules preliminarily.First, the method of solid nodule detection.Hessian matrix method is very representative for the detection of solid nodules. It combines the methods of gray, shape and structure. The basic idea is that the different structures can be distinguished from each other based on local anisotropic variations of voxel intensities in CT image. For different structures, the gray scale distribution patterns of voxels are different. Solid nodules are highlight spherical or almost spherical structures, whose intensity distribution is radial attenuation from the center, can be approximated three dimensional Gaussian distribution. Blood vessels have similar structures with tubular segments that is approximately a cylinder with a Gaussian intensity profile at the section plane, the voxel intensity variations of a vessel are expected to be low along the local orientation of the tube and high in the plane perpendicular to this orientation. Pleura is approximated to planar structure, the intensity changes slowly in the pleura, but it changes greatly extending the perpendicular direction.Hessian matrix is composited by the second derivative of the images, with a high sensitivity. It can well reflect the grayscale changes, is widely used in nodule detection. However, the Hessian matrix is obtained by calculating the grayscale changes of voxel and adjacent ones, using Hessian matrix eigen-values alone can’t reflect larger regional grayscale changes or distribution. When different tissues have similar intensity distribution, such methods will be misjudged and difficult to eliminate false positives generated by vessel junctions.In order to solve this problem of Hessian matrix methods for nodule detection, this paper proposed a novel method, which introduced a three dimensional adaptive window. First, calculate the Hessian matrix of every voxel, analyze the intensity distribution of voxels by analyzing the Hessian matrix eigen-value. Second, use eigen-values to design structure coefficient that describes a degree of voxel belonging to certain tissues. Then construct three-dimensional adaptive window based on structure coefficient, this made the area for analysis increased from adjacent voxels to a larger local area. Finally, analyze local structural features of tissues within the windows, and use discrimination function to remove blood vessels and vessel junctions. The experiments on49nodules of21CT scans show that the method can effectively reduce false positives produced by blood vessel junctions.Second, the method of GGO nodule detection.Most current GGO detection methods had used gray threshold or filter.Gray threshold methods extracted suspected GGO regions and removed irrelevant tissues by setting threshold, based on the experience or histogram analysis. As the gray-value of GGO between the lung parenchyma and vessel, with large span, the thresholds set by the experience are difficult to apply to different images. If the range of thresholds is too large, the result of suspected GGO regions will contain a lot of false positives, which will hamper further analysis. If the range of thresholds is too small, it is likely to miss GGO nodules. The thresholds determined by image histogram analysis are more robust and suitable for different images.The filter methods enhanced GGO regions directly or suppressed noise and other tissues indirectly. It is difficult to design templates of spatial filter for GGO nodules, because they are, unlike solid nodules, no rules on grayscale distribution and shape of the structures.In this paper, the method of GGO nodule detection determined the thresholds of GGO regions were adaptively by Gaussian Mixture Model fitting histogram. Gaussian high pass filter with low cut-off frequency was used to stretch grayscale difference between GGO regions and pulmonary parenchyma, suppress pulmonary parenchyma and eliminate most of the patch shadows. First, employ Gaussian Mixture Model fitting curve to analyze the histogram of images, adaptively set upper and lower thresholds of GGO regions. Second, employ a Gaussian high pass filter to enhance GGO regions and suppress pulmonary parenchyma to avoid the interference of patchy shadows. Then use the result of Gaussian high pass filter and the upper and lower thresholds to extract candidate GGO regions. Finally, use a circular averaging filter and regional characteristics to remove false positives. The experiments on26nodules of21CT scans show that this method can detect most of the GGO nodules.In addition, GGO nodules have higher malignant than solid nodules, the malignancy of pure GGO nodules reaches up to59-73%, and mixed GGO nodules’is even higher than pure GGO nodules’. Therefore, after the computer aided detection, assisting doctors for diagnosis of benign and malignant by discriminate types of nodules also has research value. Comprehensive testing was composed by solid nodule detection method and GGO nodule detection method for the same CT images. Nodules detected by the solid method were considered as solid ones. Nodules detected by the GGO method were considered as GGO ones. Nodules detected by both methods were considered as mixed ones. The nodules in different types were given the corresponding scores and compared with the ratings of experts of LIDC database. By contrast with the experts prove that the scoring of this method is consistent with the experts’.From the status at home and abroad in the field of computer-aided detection of lung nodules, combining with anatomical knowledge and lung CT imaging features, using LIDC database for the study, this paper achieved the detection of solid nodules and GGO nodules, and studied the type discrimination and score of nodules. It is innovative and valuable.