| Objective : For more faithful and verifiable recognition of thyroid nodules,the BP neural network model and Logistic regression classification model were respectively constructed by using the ultrasonographic characteristics of thyroid nodules and their combined serological characteristics to compare the diagnostic value of thyroid nodules and ultrasonographic characteristics,and the action of the two models was compared to find the optimal diagnostic model.For clinical diagnosis of disease to provide assistance.Method:All told there were 93 patients with thyroid nodules who had hospitalized in the nail breast sector in the past two years and underwent routine US,puncture biopsy and preoperative endocrine test were collected.In line with pathological biopsy,these patients were split into cancerous set(52 cases)and harmless set(41 cases).The ultrasound features and serological indicators of all patients were processed and translated into machine-learned language.SPSS26.0 statistical software was applied to dispose the data and the P value.Med Calc statistical software was used to describe the ROC curve of serological indicators and calculate the value of AUC.Indexes with P value greater than 0.5 and AUC value greater than 0.7 were selected to construct the BP neural network model and Logistic regression model.The model construction was split into two steps: The first step is to construct a model that only contains ultrasonic features,and the second step is to construct a model that combines ultrasonic features with thyroid serological indexes.Hosmer test coefficient should be obtained in the Logistic regression classification model,and Hosmer test quotient greater than 0.05 should be included.By comparing the reliability,susceptibility,precision of the experimental results,the optimal model is selected.Results:Statistical analysis showed that the US characteristics of thyroid were statistically meaningful(P<0.05).Refering to the AUC value,TSH,TPOAB,TGAB and TG could be included in the experiment.BP neural network model trained by matlab2016 a program software was used for testing,which only included US feature of thyroid nodule.The susceptibility,reliability,precision of model 1 were 90.9%,75.0% and 84.2%.Logistic regression analysis revealed that the reliability,susceptibility,precision of the model 1 were 66.6%,534% and 60.6%.The diagnostic efficiency of BP neural network model was 23.6% higher than that of Logistic regression model.In addition,Hosmer test coefficient of Logistic regression model is0.039,less than 0.05,indicating that the model is underfitting.Through the program software matlab2016 a,BP neural network model II containing ultrasonic characteristics of thyroid nodule and thyroid serological indexes was trained.The susceptibility,reliability,precision of the results of thyroid cancerous set and harmless set were 100%,87.5% and 94.7%.The model including thyroid nodule US feature and serological indexes was obtained by Logistic regression analysis algorithm.The susceptibility,reliability,precision were 82.6%,77.5%,80.3%.The diagnostic efficiency of BP neural network model was 14.4% higher than that of Logistic regression model.The Hosmer test quotient of Logistic regression model was 0.865,greater than 0.05,indicating good model fitting.The diagnostic efficiency of BP neural network mode 2 was 10.5% higher than that of mode 1.Conclusion:The diagnostic coincidence rate of thyroid nodule ultrasound features combined with serological indexes was higher,and the diagnostic efficacy of BP neural network model constructed was significantly better than that of Logistic regression model,which could be applied as a powerful auxiliary tool for clinical diagnosis.Figure[7] Table[10] Reference [45]... |
PDF Full Text Request