Study On Feature Extraction And Classification Algorithm Of Thyroid Nodules

Thyroid nodule is a kind of mass that is caused by the rapid growth of the cells in the thyroid gland under the influence of inducement.Clinical ultrasound diagnosis technology has become the first choice for the examination of the disease.Because there is a significant difference between the ultrasonographic characteristics of benign and malignant thyroid nodules,So it is feasible to distinguish the benign and malignant nodules by ultrasound image.In this paper,the thyroid nodule ultrasound image is used as the research object,and relevant research is carried out on nodule region segmentation,feature extraction and identification,thus providing effective diagnostic decision support for doctors.The main work of this paper is as follows:1.Regional segmentation of nodules:As ul-trasound images have the characteristics of large noise,low contrast and uneven gray,the segmentation algrithm based on improved local image fitting(LIF)model and Chan Vese(CV)model is proposed in this paper.The traditional LIF model is easy to fall into local minimum during evolution process.So in order to avoid the local optimal problem during evolution,we introduce global gradient energy information and combine the advantages of global segmentation of CV model.The model not only can segment the images with nonuniform gray distribution,but also weaken the sensitivity of the active contour to the initial position.The experimental results show that the proposed algorithm can not only overcome the influence of noise,but also achieve accurate segmentation of uneven gray images.2.Feature extraction:According to the TI-RADS standard of thyroid tumor diagnosis,this paper describes the characteristics of thyroid nodules from the aspects of edge,shape,calcification,attenuation,echo and so on.As the internal and boundary echo distribution of the nodules can be described by the texture features,an algorithm for the feature extraction of thyroid nodules is proposed,which combines edge,shape,texture,attenuation and calcification.In the aspect of texture feature extraction,the shortcomings of the traditional CLBP model are improved.By introducing the local variance information,the improved CLBP is more fine to characterize the local texture information of the image,Thus,the recognition rate is improved.3.SVM classifier optimization:In order to further improve the recognition rate,a SVM parameter optimization algorithm based on particle swarm optimization(PSO)and mesh search algorithm is proposed.First of all,the suitable parameter group is determined quickly by using the large distance mesh search algorithm.Then the range determined by this parameter group is regarded as the best search range of particle swarm optimization algorithm,which effectively combines the advantages of grid search algorithm and particle swarm optimization algorithm.The SVM parameters are optimized efficiently and the classification accuracy is higher.The experimental results show that the proposed algorithm can achieve a higher recognition rate compared with other common classification algorithms.The proposed SVM joint optimization algorithm has obvious advantages over the traditional single optimization algorithm,shortens the training time and improves the classification efficiency.