Research On Diagnosis Of Solitary Pulmonary Nodules Based On Biomimetic Pattern Recognition

Lung cancer is one of the most common malignant diseases. The incidence of this disease has obviously ascended in recent decade. Early detection and therapy is the most effective way to prevent and cure the pulmonary diseases. At present, CT scan is an important tool for diagnosis of pulmonary diseases. But, with the widespread use of CT scan, a large number of CT images will increase the physicians'workload. And suspicious medical signs may be underestimated because of the difference in physicians'apperceiving. These all may increase the probability of misdiagnosis. With the development of pattern recognition, machine learning and digital image processing techniques, the Computer Aided Diagnosis has advantage for detection and diagnosis of lung disease. It enhances the efficiency of medical diagnosis and reduces physicians'burden.First, the theory of Artificial Neural Network (ANN), Bionic Pattern Recognition (BPR) and Support Vector Machine (SVM) are introduced. Second, a new method for Solitary Pulmonary Nodules (SPNs) detection based on BPR is proposed. And then six classifiers are realized for SPNs detection and classification for benign and malignant SPNs respectively based on SVM, BPR, BP neural network. And the experiment results of detection and distinction are compared and analyzed. Finally, two multi-class methods are realized respectively based on SVM and BP neural network. The main aspects of research include:1. The neural network model for the SPNs detection based on the BPR principles is proposed. The BPR makes recognition from the views of "matter cognition" instead of "matter classification". It analyzes and cognizes the high dimensional geometrical distribution consisting of the sample sets in the high dimensional feature space. It provides the theoretic basis of building the neural network based on high dimensional theory.2. Six classifiers are studied and realized for SPNs detection and classification for benign and malignant SPNs respectively based on SVM, BPR and BP neural network. And the experiment results of detection and distinction are analyzed by Receiver Operating Characteristic (ROC) curves. The analytic results show that the classification results of SPNs detection and classification for benign and malignant SPNs based on BPR are better than the classification results based on SVM and BP neural network.3. The multi-class methods for SPNs diagnosis based on BPR and BP neural network are designed and realized respectively. In the SPNs diagnosis, two multi-class methods based on BPR are used. One is named as combined multi-class, which is implemented by combining two bi-class methods for SPNs detection and classification for benign and malignant SPNs. The other is designed by using direct multi-class method. And the experiment results of the above three methods are compared and analyzed. The results show that the SPNs diagnosis used combined multi-class based on BPR is better than the others.