Research And Application Of Lung Cancer Tissues Classification Algorithm Based On Dielectric Properties

Dielectric properties which is inherent property of biological tissues can indirectly reflect the physiological changes of tissues.The dielectric properties change significantly when tissue becomes cancerous.The dielectric parameters(eg.permittivity and conductivity)measured by the open-ended coaxial probe which has the advantages of non-invasive,simple and timesaving,and is expected to be used in the real-time detection of malignant tumors in surgical operations,so as to realize the accurate resection of malignant tumors,reduce the damage to healthy tissues and the risk of surgery.The study of tissue classification based on dielectric properties is helpful to realize the automatic decision of clinical tumor diagnosis and promote the process of clinical application.Lung cancer is one of the tumors with high morbidity and mortality,and its five-year survival rate is low,which seriously endangers human life and health.In view of the lack of relevant studies on the classification of lung cancer tissues based on dielectric properties,this paper proposes an adaptive probabilistic neural network for the classification study.The dielectric parameters of samples which include 200 cases of lung tissue and 219 cases of lymphonodi pulmonales were measured by open-ended coaxial probe in the frequency range of 50MHZ to 4GHz.Synthetic Minority Over-sampling TEchnique(SMOTE)is used to solve the problem of class-imbalance.Statistical Dependency(SD)algorithm to score the sample characteristics,and then conduct the classification study by adaptive probability neural network which respectively include based on validation set and based on clustering algorithm.In addition,the traditional Probabilistic Neural Network(PNN),BackPropagation(BP)neural network,Radial Basis Function(RBF)neural network,linear discriminant analysis,Support Vector Machine(SVM),k-Nearest Neighbor(kNN)algorithm were also used to conduct classification experiments.The discriminant accuracy,sensitivity and specificity were calculated through 20 times hold-out(20-hold-out),and experimental results of different methods were compared.Finally,the man-machine interactive Graphical User Interface(GUI)is designed to intuitively show the identification results to users and facilitate the experiment.The experimental results showed,for lung tissue,the permittivity at the frequency points of 984,2724,2723,1001,982,2771,2722,2729,2726,2728MHz as sample characteristics,the adaptive probability neural network achieved the highest identification accuracy of 93.63%,corresponding to sensitivity of 94.32%,and specificity of 92.78%.For lymphonodi pulmonales,the conductivity at the frequency points of 3959,3958,3960,3978,3510,3889,3888,3976MHZ as sample characteristics,the adaptive probabilistic neural network has achieved the highest identification accuracy of 92.92%,corresponding to sensitivity of 94.72%,specificity of 91.11%.The number of each layer neurons of the proposed adaptive probability neural network can be easily determined,the network parameters can be adaptively change,and performance of adaptive probability neural network outperforms other algorithm in terms of accuracy in lung cancer tissue classification based on dielectric properties.The paper provides a new method for automatic decision of intraoperative tumor diagnosis based on dielectric properties and has potential clinical application value.