Research Of Feature Selection Algorithm For Solitary Pulmonary Nodules In PET-CT Imaging

Lung cancer presents the highest incidence and mortality rate across theworld. Early detection, early diagnosis and early treatment is one of theimportant measures to reduce its mortality. In the early stages, lung cancerappears predominantly as solitary pulmonary nodules (SPNs). The emergence ofPET-CT provides a very reliable technique for the early detection and diagnosisof SPNs. Using the computer-aided diagnosis (CAD) system to early detect anddiagnose SPNs has important research value and application prospect, as it canreduce the workload of doctors, reduce the rate of misdiagnosis and misseddiagnosis, and improve the efficiency and accuracy of the detection anddiagnosis of pulmonary nodules.The CAD system based on PET-CT imaging generally contains two mainstage, detection of lung nodules and diagnosis of benign and malignant nodules.On the basis of the detection of SPNs, this paper researched on how toefficiently diagnose lung nodules. The diagnosis system of SPNs mainly includePET-CT image preprocessing, PET-CT image registration, lung parenchymasegmentation, pulmonary nodule segmentation, feature extraction and classifierconstruction. To construct a high-performance classifier, the optimization of thefeature set of pulmonary nodules is still a problem to be solved, namely how tochoose a low redundancy feature subset that has high correlation with benignand malignancy. Therefore, this paper carried out a series of study, mainlyincluding the following sections:(1) Constructing of the feature set of SPNs on PET-CT imaging. From the angles of the imaging diagnosis and computer-aided diagnosis of lung nodules,the annotation feature sets based on imaging diagnosis and nodules feature setsbased on feature extraction are respectively constructed, which established theexperimental basis for the research of feature selection algorithm.(2) Studying the feature selection method based on information theory. Thisstudy theoretically compared the classical feature selection methods based onmutual information measure metrics and found that the method based on jointmutual information (JMI) measurement criterion has better stability and higheraccuracy. Then the nodules hybrid feature selection algorithm based on JMI onPET-CT imaging was put forward. The effectiveness of the proposed algorithmwas verified by experiments. Under the predefined classifier, this method cannot only get less number of features in the subsets, but also improve theclassification accuracy in solitary pulmonary nodules diagnosis.(3) Studying the feature selection method based on gray correlationanalysis. On the basis of research and analysis of the grey correlation analysismodel based on different correlation, and according to the different selectingway of the first and subsequent features of the feature subset, this paperdesigned and put forward four different feature selection algorithm of SPNs onPET-CT imaging. Experimental results show that this algorithm got betterperformance on both the number of features and classification accuracy,verifying the effectiveness of the grey correlation analysis applied to featureselection.