Researches On Key Technologies Of Computer Aided Diagnosis For Solitary Pulmonary Nodule Based On CT Images

Lung cancer is the most common visceral malignant tumor which has the lowest survival rate after diagnosis;it becomes the No.1killer among various cancers. Solitary pulmonary nodule is the most common and the most important imaging finding of lung cancer. The detection of lung nodules is quit crucial for the early detection and early diagnosis of lung cancer. In order to reduce undetected lung nodules and possibilities of error detection,improve sensitivity,specificity and diagnostic accuracy of nodule detection, we must apply CAD (computer aided diagnosis) techniques to conduct automatic image anlysis and recognition on pulmonary nodules and assist radiologists to conduct evaluations and diagnosis.The key technologies of automatic detection on pulmonary nodules include:segmentation and detection, other image processing, data mining, research and application of machine learning. Since medical images have the characteristics such as uneven gray,various individual differences and easily subject to artifacts and noises, so it is quite difficult for detection algorithm of pulmonary nodules achieve higher sensitivity and accuracy.In this paper, it focuses on the medical characteristics of lung CT images,conducts in-depth researches and various experiments from the main directions of making image segmentation of lung parenchyma, detecting and extracting suspected nodules features and conducting recognition on lung nodules.The main work and innovations of this paper are as follows:(1) Quick and accurate segmentation of lung parenchymaLung CAD method requires to make lung parenchyma segmentation on the image. This paper based on imaging and anatomical characteristics of chest CT images, makes use of the approaches of combined thresholding and region growing to divide lung parenchyma. It based on two-dimensional Ostu method of quantum particle swarm to select threshold.It applies chest CT image to remove the background and chest, achieves lung parenchyma segmentation of two dimensional CT image. Makes use of region growing to achieve lung parenchyma segmentation of three-dimension. Improve segmentation accuracy of lung parenchyma and keep segmentation speed.(2) The detection of pleural adhesions nodulesWhen there are pleural adhesions pulmonary nodules, we need to correct the lung boundary and detect adhesion pulmonary nodules. It proproses adhension pulmonary nodules detection method which is based on line scan curvature analysis which can effectively solve the easily-missed lung parenchyma in high-density, easy leakage of pulmonary nodules on the edge of lung parenchyma and other problems. It proposes pleural adhensions nodules detection algorithm which is based on convex hull and ray projection which can accurately detect mediastinal adhesions pulmonary nodules and correct the damaged lung parenchyma. (3) The segmentation of suspected pulmonary nodulesThe automatic segmentation of suspected pulmonary nodules is the premise of making accurate extraction of nodule characteristics. This paper regards suspected pulmonary nodules as regions of interest(Regions of interest),puts forward an algorithm which integrates spatial information and fuzzy C means clustering with feature weight to realize segmentation on regions of interest. It brings the spatial information into fuzzy C means clustering which can effectively remove nosies,it makes clustering by taking image gray histogram as weight to improve the clustering speed and segmentation accuracy.(4) Feature extractionBased on the medical phenomena of pulmonary nodules and pathological basis of lung diseases,it makes analysis and extraction of basic gray scale,morphologic features and texture features in pulmonary nodules. Conduct measurements and calculation on the feature values of pulmonary nodules and thus realize the feature extraction and quantification of ROI region.(5) Feature selection based on rough setIn order to conduct dimension reduction on the feature space of high-dimension,it introduces feature selection methods which are based on rough sets. The reduction method of rough set attribute can effectively reduce the feature dimension and obtain the important feature of lung nodule identification.(6) The recognition of pulmonary nodules based on sample attribute weight support vector machineProposes the method of sample attribute weight support vector machine to identify pulmonary nodules.It makes use of rough sets to determine the importance of each attribute as weights and adds weights on the attributes to achieve the classification of pulmonary nodules, It increases the accuracy of lung nodule recognition. From the comparisons between the pulmonary nodules identification algorithm which is based on fuzzy C means proposing by author and the algorithm which ia based on improving Manhalanobis distances, it proves that the sample attribute weighted SVM classifier is slightly superior than others on sensitivity specificity and accuracy.In conclusion, this paper makes certain progress in improving segmentation accuracy of lung parenchyma and pulmonary nodules, enhancing sensitivity and specificity of solitary pulmonary nodule detection, reducing the misdiagnosis rate and wrong diagnosis rate, but there are still some deficiencies which need make further improvements on theoretical basis and practical applications.