Study On Predicting Cervical Central Lymph Node Metastasis Based On Ultrasound Radiomics-Transfer Learning Model For Papillary Thyroid Cancer

Objective(s):To explore the diagnostic efficacy of models based on ultrasound radiomics,transfer learning and ultrasound radiomics-transfer learning combination for cervical central lymph node metastasis of patients with thyroid papillary cancer.Methods:1.The ultrasound images of 191 patients with thyroid papillary cancer diagnosed by preoperative fine needle biopsy or postoperative histopathology in the Second Affiliated Hospital of Kunming Medical University from January 2020 to October2022 were retrospectively analyzed.All patients underwent total thyroidectomy/subtotal thyroidectomy+preventive lymph node dissection in the central region.The postoperative pathological diagnosis confirms whether there were metastasis in the central lymph nodes.2.The accuracy,sensitivity and specificity of the preoperative sonographer diagnosis of cervical central lymph node metastasis with papillary thyroid cancer patients were calculated according to the comparison between the preoperative sonographer diagnosis of cervical central lymph node metastasis and the postoperative histopathological findings.3.All patients underwent strict inclusion and exclusion criteria,collecting the size,location,aspect ratio,boundary,margin,internal echo,thyroid capsule invasion and microcalcification of the nodules on ultrasound scans of the enrolled patients,as well as clinical data such as age,gender,and history of Hashimoto’s thyroiditis.Using t-tests and chi-square tests,statistically significant differences were found with P <0.05 to identify risk factors associated with the cervical central lymph node metastasis.4.The risk factors with statistically significant differences in clinical and ultrasound images were included in the analysis,based on five Support Vector Machine,K-Nearest Neighbor,Extreme Random Tree,Random Forest and XGBoost bosting methods machine learning methods and establishing clinical-feature prediction models.The best performing model was selected by comparing the accuracy and AUC values of the five models.5.Establish the ultrasound radiomics model.The 3D Slicer 4.11 software was applied to manually outline the ROI of the optimal section of the lesion.Use the Py Radiomics platform based on Python was used to extract radiomics features from each region of interest,and the ICCs,U-test,Spearman correlation coefficient and LASSO algorithms were used for radiomics feature screening.Five machine learning algorithms were used for model construction.The models were evaluated and screened using their accuracy and AUC values.6.Construct the model of transfer learning.Features were extracted using Medical Net pre-trained Res Net34 algorithm,and then the features were filtered using the same method as radiomics.The same five machine learning algorithms were used for modeling,and the prediction efficacy was evaluated using the AUC value and accuracy of each model.7.Construction the model of ultrasound radiomics-transfer learning model.By combining image features from radiomics and transfer learning,five machine learning algorithms were used to model the combined feature models,and the model with the best prediction performance was screened based on the accuracy and AUC value of each model.8.Compare the AUC value,accuracy,sensitivity and specificity of the clinical ultrasound feature model and the image feature based model,including the ultrasound radiomics model,the transfer learning model and the ultrasound radiomics-transfer learning model,and evaluate the prediction efficiency of the cervical central lymph node metastasis.Evaluate the clinical practicality of the four models using clinical decision curves.9.Using the multivariate Logistic regression method,based on the optimal ultrasound image feature model and the clinical-ultrasound feature model,a nomogram of a prediction model for cervical central lymph node metastasis is established,making the model visualized,and the calibration degree of the model is evaluated by drawing a calibration curve,and the clinical practicality of the model is evaluated by decision curve analysis.Results:1.Among the 191 papillary thyroid cancerpatients,the preoperative sonographer indicated metastasis in the cervical central lymph nodes in 32 cases and no metastasis in 159 cases;the postoperative pathological histological diagnosis showed that 90 cases had the cervical central lymph nodes metastasis and 101 cases had no metastasis.2.According to the comparison between the results of preoperative sonographer diagnosis of cervical lymph node metastasis in patients with papillary thyroid cancer and postoperative histopathological diagnosis,the accuracy,sensitivity,and specificity of ultrasound diagnosis of cervical central lymph node metastasis in 191 patients with papillary thyroid cancer were 62.3%,27.8% and 93.1%,respectively.3.Based on the independent sample t-test and chi-square test,three risk factors with P< 0.05 were screened: age,maximum diameter of nodules and internal echo.The metastasis of cervical central lymph nodes in patients with papillary thyroid cancer is related to the patient’s age,while younger patients(age<40 years old)are more prone to cervical central lymph node metastasis,and age is negatively correlated with the central lymph node metastasis.The maximum diameter and internal echo of the nodule are risk factors for ultrasound characteristics of lymph node metastasis.Nodules with a maximum diameter greater than 1cm and heterogeneous echoes inside the nodule are more likely to experience cervical central lymph node metastasis,while nodules with a maximum diameter less than 1cm and homogeneous echoes have a lower probability of metastasis.4.The AUC values of the clinical ultrasound feature model based on Support Vector Machine,K-Nearest Neighbor,Random Forest,Extreme Random Tree and XGBoost bosting methods was 0.754,0.790,0.635,0.651 and 0.729,respectively.The accuracy was 63.2%,73.6%,63.1%,60.5% and 52.6%,respectively.The model based on K-Nearest Neighbors has the highest AUC value and accuracy,and the model has the best predictive performance.5.In 191 images of thyroid nodules,radiomics extracted 439 features from each region of interest,and finally retained five features with non-zero coefficients through feature filtering.The AUC values of the radiomics model based on Support Vector Machine,K-Nearest Neighbor,Random Forest,Extreme Random Tree and XGBoost bosting methods was 0.775,0.572,0.708,0.611,and 0.723,respectively.The accuracy was 71.1%,60.5%,57.9%,60.5% and 71.1%,respectively.The AUC of the radiomics model built by Support Vector Machine is the best,therefore,Support Vector Machine is chosen as the machine learning method for building the model.6.Transfer learning extracted 100 image features from each thyroid nodule image;and constructs transfer learning models based on the above five machine learning methods.The prediction efficiency AUC of each model was 0.822,0.667,0.844,0.783,0.681,respectively;The accuracy was 76.3%,57.9%,73.7%,68.4% and55.2%,respectively;The model with the best performance is built by a support vector machin.7.Combined with the characteristics of ultrasound radiomics and transfer learning,a model was established to predict cervical central lymph node metastasis of patients with thyroid papillary cancer.The AUC of the model based on Support Vector Machine,K-Nearest Neighbor,Random Forest,Extreme Random Tree and XGBoost bosting methods was 0.872,0.800,0.694,0.704,0.619,respectively.The accuracy is73.6%,76.3%,68.4%,65.7% and 52.6%,respectively.Therefore,the AUC value of the ultrasound radiomics-transfer learning model using the Support Vector Machine algorithm is the highest.8.Ultrasound image feature models built based on Support Vector Machine: the best performance of the models for ultrasound radiomics,transfer learning and ultrasound radiomics-transfer learning.The Support Vector Machine is selected as the machine learning method for building the models,and the performance of the four feature models is compared with the clinical-ultrasound feature model,ultrasound radiomics model,transfer learning model and ultrasound radiomics-transfer learning model with the test set AUCs was 0.754,0.775,0.822 and 0.872,respectively.And the ultrasound radiomics-transfer learning model has the best prediction performance.The decision curve analysis of all four models are higher than the two reference curves,indicating that all four models have good clinical utility.9.The nomogram established based on the clinical-ultrasound features including age,maximum diameter of nodules and internal echo.The ultrasound radiomics-transfer learning feature model shows that the model has positive predictive ability(the test set AUC of 0.875)and the calibration curve reflects that the calibration of the nomogram is good.Conclusions:1.Compared with the efficacy of physicians’ preoperative diagnosis of cervical central lymph node metastasis,clinical-ultrasound feature model,radiomics model,transfer learning model and radiomics-transfer learning model are all valuable in predicting cervical central lymph node metastasis in the papillary thyroid cancer patients,which can improve the sensitivity of cervical central lymph node metastasis and have some clinical value.2.The clinical-ultrasound feature model,ultrasound radiomics model,transfer learning model and ultrasound radiomics-transfer learning model are constructed using the Support Vector Machine,and each group of data has the best index overall,which can be used as the best prediction model of the four feature models.Compared with the four feature models,the predictive efficacy of the ultrasound radiomics-transfer learning model was higher than that of the radiomics model and transfer learning model,respectively,and higher than that of the clinical-ultrasound feature model,which could improve the accuracy of preoperative assessment of cervical central lymph node metastasis in patients with papillary thyroid cancer.3.Combining the clinical-ultrasound feature model and the ultrasound radiomics-transfer learning model feature nomogram,the nomogram establishes in this paper is a non-invasive tool to assess the risk of lymph node metastasis in disease,which is beneficial for clinicians to make optimal preoperative decisions for patients and lay the foundation for subsequent related studies.