Construction Of Individualized Prediction Models For Lymph Node Metastasis In Patients With Papillary Thyroid Carcinoma And Study On Its Metastasis Mechanism

ObjectiveTo evaluate the characteristics of ultrasound and dual-energy CT images of primary foci in papillary thyroid carcinoma（PTC）patients with cervical lymph node metastasis.To construct an individualized prediction model to evaluate the diagnostic efficacy and explore the mechanism of lymph node metastasis in PTC patients.First,to evaluate the feasibility of dual-energy CT in predicting cervical lymph node metastasis with a diameter less than 0.5 cm in patients with PTC.Second,combined the ultrasound and dual-energy CT features of the primary tumor to construct a nomogram with a risk stratification system to explore its predictive efficacy for the occurrence of central cervical lymph node metastasis（CLNM）in patients with PTC,and further explain from the perspective of pathophysiology.Third,we have introduced Python language to construct eXtreme Gradient Boosting（XGBoost）model,and discussed its predictive efficacy for CLNM in PTC patients.SHapley Additive exPlanation（SHAP）and local interpretable model-agnostic（LIME）were used for overall and local visual interpretation of CLNM.Materials and MethodsPart One:359 lymph nodes in 52 patients with PTC who were admitted to Tianjin First Central Hospital from May 2016 to June 2018 were retrospective included.The64 multi-detector row CT scanner（SOMATOM Definition Flash,Siemens Healthcare,Forchheim,Germany）was used to performed dual-energy dual-phase CT scans on the neck.The SIEMENS Syngovia workstation was used for image analysis.The Liver VNC function was used to obtain the iodine maps in the arterial and the venous phases by automatic computer processing.Image quantitative analysis:all lymph nodes were double-blindly assessed by two radiologists with more than 10 years of experience in the CT diagnosis of head and neck diseases.The measurement parameters included the lymph nodes’diameter,iodine concentration（IC）,normalized iodine concentration（NIC）,and the energy spectrum curve(λ_HU)in the artery and the venous phases.The Kolmogorov-Smirnov test method was used to test for the normal distribution of continuous variables of quantitative parameters.If the quantitative parameter fitted the normal distribution,mean±standard deviation was used to describe it and t-test was used to compare the difference between metastatic and non-metastatic lymph nodes.If it did not fit the normal distribution,the median（inter quartile range,IQR）and Mann-Whitney U tests were used.The diagnostic ability of the quantitative parameters was evaluated by receiver operating characteristic（ROC）curve analysis,and the optimal threshold was selected using the Youden index.The area under the curve（AUC）of parameters was compared using the Z-test.The generalized estimating equation（GEE）based on dual-energy CT quantitative parameters was used to evaluate the relationship between the selected variables and lymph node metastasis.Part Two:525 patients with PTC who were admitted to Tianjin First Central Hospital from January 2017 to December 2019 were retrospectively included.Review the ultrasound and dual-energy CT characteristics of the primary tumor.Multivariate binary logistic regression analysis was used to screen the independent risk factors for CLNM.R language was used to construct a nomogram,and risk stratification was carried out according to the maximum slope of the locally weighted regression（LOESS）curve.After the prediction model was established,the case data of the patients who were admitted to Tianjin First Central Hospital in 2020 was retrospectively analyzed as a temporal validation（external validation cohort I）.To further verify the performance of the prediction model,we also conducted a retrospective cohort study.We collected data of 107 patients in Binzhou Medical University Hospital in 2019 for geographic validation（external validation cohort II）.Calibration curve and ROC curve were used to evaluate the performance of the nomogram.The DeLong method was used to compare the AUCs of two models（model 1:based on ultrasound parameters alone,model 2:combined ultrasound and dual-energy CT parameters）.R language software was used to calculate net reclassification index（NRI）and integrated discriminatory improvement index（IDI）.Decision curve analysis（DCA）was used to evaluate the clinical application value of the nomogram,to find out the distribution weight of different parameters to predict the possibility of CLNM.Part Three:The ultrasound and dual-energy CT image data of 1,122 PTC patients who admitted to Tianjin First Central Hospital from 2016 to 2020 was analyzed retrospectively.Among them,545 patients had CLNM and 577 patients had no CLNM.The 1,122 patients were randomly divided into two groups in an 8:2 ratio.The training data set（n=897）included 443 patients with CLNM and 454 patients without CLNM,and the test data set（n=225）included 102 patients with CLNM and123 patients without CLNM.TheXGBoost model was constructed using a Python database.A validation data set,consisting of 104 PTC patients（45 patients with CLNM and 59 patients without CLNM）,was collected from Binzhou Medical University Hospital during 2020 to conduct a retrospective cohort study.Visual interpretation of the model as an overall and locally through SHAP and LIME was carried out.The prediction performance of the model was compared with the traditional machine learning models.ResultsPart One:359 lymph nodes from 52 PTC patients were retrospectively included, including 139 metastatic lymph nodes and 220 non-metastatic lymph nodes.The diameter,IC,NIC,andλ_HUin the arterial phase,and IC,NIC,andλ_HUin the venous phase of metastatic lymph nodes were all higher than non-metastatic lymph nodes（p value:0.000-0.007）.GEE analysis showed that the lymph node with a diameter<0.5cm that met the following conditions was considered as a metastatic lymph node:located at the level VIa or VIb,diameter≥0.32 cm,IC in the arterial phase≥2.1mg/mL,and IC in the venous phase≥2.4 mg/mL.Among them,the predictive value of IC in the arterial phase of the lymph node was highest,with an AUC of 0.775.Part Two:Multivariate binary logistic regression analysis showed that diameter,taller-than-wide,calcification,the ratio of capsular abutment over the lesion perimeter,IC in the arterial and the venous phases were independent risk factors for predicting CLNM.The nomogram constructed based on the above six independent risk factors had an AUC of 0.922.A risk stratification system was established based on the maximum slope of the LOESS curve as the cut-off value,including the low-risk group（0-50 points,including 198 patients）,the intermediate-risk group（51-100points,including 228 patients）,and the high-risk group（>100 points,including 99patients）.Two external validations were performed on the forecast prediction model.In the temporal validation,the AUC reached 0.912,and the geographic validation cohort adopted a retrospective-prospective method,with an AUC of 0.861.The DeLong method was used to compare the performance of two prediction models,and found that model 2 was better than model 1,indicating that dual-energy CT was an effective supplement to ultrasound in PTC patients（AUC increased by 12%）.Part Three:The parameters included in theXGBoost model included:sex,age,Hashimoto’s thyroiditis,location,composition,diameter,echogenicity,shape,boundary,margin,capsular invasion,calcification,vascularization,IC in the arterial phase,IC in the venous phase,NIC in the arterial phase,and NIC in the venous phase.The prediction accuracy of theXGBoost model for CLNM was 83.05%,with 425patients accurately predicting CLNM（+）,327 patients accurately predicting CLNM（-）,and 145 patients with overestimated or underestimated the probability of CLNM.And the AUCs of the training,test,and validation data sets were 0.918,0.901,and0.890,respectively.The SHAP summary plots showed that the six most important parameters of theXGBoost model to predict CLNM were capsular invasion,diameter,sex,IC in the venous phase,taller-than-wide,and calcification.And capsular invasion,diameter,IC in the venous phase,taller-than-wide,and calcification had positive predictive effects,while sex had negative predictive effect.The SHAP interaction plots showed that the interaction between capsular invasion vs.diameter,capsular invasion vs.sex,diameter vs.IC in the venous phase,diameter vs.calcification played positive roles in predicting the occurrence probability of CLNM.Conclusions1.Quantitative parameters of dual-energy CT could assist clinicians in predicting metastatic lymph nodes less than 0.5 cm in PTC patients.2.The nomogram constructed based on the ultrasound and dual-energy CT features of the same primary lesion had good predictive performance for CLNM,and could get the distribution weight of different parameters to predict CLNM,in which taller-than-wide had the largest weight,and odds ratio（OR）was 3.802.The risk stratification system based on LOESS curve was helpful for clinicians’simple application.3.XGBoost model based on ultrasound and dual-energy CT images of the same primary focus had excellent predictive performance for the possibility of CLNM.SHAP and LIME were used to solve the“black-box”problem of artificial intelligence（AI）,and the overall and local visual interpretation of the occurrence probability of CLNM in PTC patients was performed.XGBoost provided objective explanations including positive and negative effects for CLNM’s prediction by comparing with traditional prediction models.