Based on Parameter Optimization of SVM to Establish an Auxiliary Diagnosis Model of Benign and Malignant Pulmonary Nodules (Preprint)

BACKGROUND Solitary pulmonary nodule (SPN) is a common disease in clinic but it is difficult to diagnose[1]. Since most patients have no symptoms when nodules are found, doctors' judgment of nodules is mainly based on their clinical experience, which is highly subjective.Therefore, it is necessary to establish an accurate and objective method for the diagnosis of benign and malignant pulmonary nodules. OBJECTIVE The SVM parameters were optimized by the intelligent algorithm, and the auxiliary diagnosis model of benign and malignant solitary pulmonary nodules combining CT images and serological indicators was constructed, and its test efficiency was evaluated. METHODS CT images and serum indexes of 1030 patients (515 cases of lung cancer and 515 cases of benign pulmonary nodules) diagnosed in our hospital between July 2015 and December 2018 were collected. The CT images of pulmonary nodules were characterized by artificial dimension reduction for feature extraction and assignment,At the same time, the serological indexes were tested; Logistic regression analysis was used to screen CT features and serum indexes of lung cancer; Grid, PSO and GS were used to find the optimal parameters C and g of SVM, and an auxiliary diagnosis model of benign and malignant solitary pulmonary nodules was constructed. RESULTS A total of 9 quantitative image features were extracted from the lung lesion regions segmented from the CT images to describe the phenotypic features of the tumor and their values were successfully assigned. 8 related serological indicators were detected, totaling 17 indicators.The main features of lung cancer including nodule site, edge condition, burr sign, foliation sign, cyfra21-1, scc-ag, CA153 and CA125 were obtained through Logistic regression analysis.Based on the above 8 screening indexes and 17 overall indexes, SVM modeling was carried out after optimization by three intelligent algorithms. The prediction results of the three algorithms in the SVM model with 8 indexes included were as follows: the prediction accuracy of the SVM model under optimization by grid search algorithm was 100%.The accuracy of SVM model was 99.5146% under gga and PSO, and 98.544% under default parameters.The three algorithms were consistent in the prediction results of the SVM model with 17 indexes, and the accuracy reached 100%, while the model accuracy under the default parameters was 88.350%. CONCLUSIONS The accuracy of SVM model can be improved by searching the optimal parameters of SVM with intelligent algorithm.8 relevant indexes screened by the logistic system are selected, and the prediction of the SVM model under optimization by the grid search algorithm can select the least inclusion indexes and guarantee the accuracy, which is the best choice.