Research Of Segmentation And Classification Algorithm For Solitary Pulmonary Nodules In PET-CT Imaging

As the PET-CT chest imaging has been used in clinical practice more and more widely, the PET-CT technology has become an important means for pathology research and diagnosis of pulmonary nodules, and function evaluation of pulmonary. However, the PET-CT scan generates large amount of images, resulting in an increase in doctors’ workload and misdiagnosis rate. The computer-aided diagnosis system is implemented using a combination of computer techniques and imaging diagnostic methods. It can help detect and diagnose lesions in medical imaging and can improve the sensitivity and specificity compared with the subjective, unstable manual reading. At early stages, lung cancer mostly appears as solitary pulmonary nodules(SPNs). Early detection and treatment for SPNs has important significance to improving survival rates. Therefore, computer-aided diagnosis methods for SPNs in PET-CT imaging become research hotspots.In the computer-aided diagnosis system of lung disease based on medical imaging, the segmentation algorithm is the foundation of SPNs detection, and the classification algorithm is the key to diagnose SPNs. This paper studies and develops from the segmentation and classification algorithms of the SPNs.In the research of the SPNs segmentation, this paper proposes a sub-region clustering-based method. The adopted clustering algorithm is a self-generating neural network(SGNN) optimized by the particle swarm optimization(PSO). The proposed segmentation algorithm is mainly aimed at the subsolid SPNs containing the constituents such as ground glass and cavitation. Due to the non-uniform internal density and the low boundary contrast of these SPNs, the segmentation effect of these SPNs in the traditional algorithms is not ideal. In the proposed method, the rough segmentation of image is performed firstly through a region growing method. After that, the PSO-SGNN-based clustering algorithm is used to cluster regions. Finally, metabolic features are utilized to identify and segment the nodular region. Experimental results show that, this algorithm has satisfied segmentation precision for both solid and subsolid nodules.In the study of SPNs classification, the features of the SPNs was first extracted and quantized. Then, based on the advantage of SGNN in solving the clustering and classification problems, a classifier based on a hybrid optimized SGNN is built to classify SPNs. In the process of classifier training, we design an integrated optimization method of network, merging pruning and M-L optimization, and adjust the network structure in a supervision way. In the testing phase, the feature-based distance measurement and sample automatic connection are both used to classify benign and malignant SPNs. Experimental results on the datasets show that, the classifier based on the improved SGNN can improve the performance of the benign and malignant diagnosis, and can classify SPNs more effectively and rapidly.