Study On Key Techniques Of Computer-aided Diagnosis System For Pulmonary Nodule

The morbidity and mortality of lung cancer caused by air pollution and the disordered life style have been consistently increased, which threaten our lives and health. Lung cancer potentially manifests itself as pulmonary nodules in an early stage. Computed tomography(CT) is one of the most important methods in the inspection of pulmonary nodules. However, the diagnosis is time-consuming extremely due to hundreds of axial images and the fatiguing process is prone to increasing the risk of missed diagnosis and misdiagnosis by radiologists. To improve the efficiency and the accuracy of diagnosis, a computer-aided diagnosis(CAD) system is designed to detect pulmonary nodule and display the 3D volume visualization of the chest.The CAD system for pulmonary nodule detection consists of four steps:(1) To quickly extract 3D lung parenchyma and suspicious nodule regions, the lung parenchyma segmentation based on automatic region growing and the extraction for regions of interesting based on the multiscale dot enhancement filtering are proposed. After extracting initial lung parenchyma region by adaptive threshold method, the central seed points of two lungs are computed to realize region growing. Morphological close operation is used to fill all holes and smooth edges for initial lung parenchyma. After segmenting the lung parenchyma, multiscale dot enhancement filtering is employed to enhance the nodule image that is smoothed in different scales before. The enhanced image is segmented with an appropriate threshold value for the separation of suspicious regions from other structures. Then, a morphological open operation and Seed-filling method are used to eliminate some small objects and label connected regions. The remaining objects are considered to be the regions containing nodule candidates.(2) To improve the accuracy of pulmonary nodule detection, a solution to collect pathological diagnosis information in XML documents is proposed with the help of MSXML parser, and a random forests algorithm is exploited to recognize pulmonary nodules. Five normative data tables are designed to store pathologic diagnosis information and related medical image information collected by the Microsoft's MSXML parser. Subsequently, the 50000 training samples are randomly selected from collecting data source. Twenty features such as the intensity features and shape features are extracted at each sample. After reducing dimensions by principal component analysis, random forests classifier is trained and employed to classify the pulmonary nodules. Quantitative evaluations in term of recall and accuracy measures show the superiority of random forests algorithm compared with Bayes, SVM and Boosting classifiers and the rate of false positives is simultaneously low.(3) To render realistic 3D images and display the spatial relationship of pulmonary nodules, the volume visualization of pulmonary nodules based on ray casting model, illumination calculation model and transfer function are designed in virtue of the OpenGL, Cg and GPU. In order to simulate surgeries and remove irrelevant background, interactive arbitrary shape volume clipping is proposed. Finally, the lung vessels in lung parenchyma are segmented with an appropriate threshold value. Lung vessels of color rendering are realized by the hierarchical calculating method and k-means cluster method.The CAD system in this paper can segment lung parenchyma and suspicious nodule regions a the pre-segmentation of pulmonary nodules. The performance of result is achieved, as indicated by an accuracy of 89.81% and a false positive rate of 8.76%.