A Prediction Model For Probability Of Malignancy In Solitary Pulmonary Nodules Detected On ¹⁸F-FDG PET/CT:Establishment And Validation

Objectives:To analyze the imaging feature of the nodules and clinical information of the patients,to establish a prediction model assess the probability of malignancy in solitary pulmonary nodules which detected on PET/CT scanning,and compare our model’s performance with the other three models to validate the accuracy.Materials and methods:The data of the PET/CT center in our hospital from January2013 to July 2017 were analyzed retrospectively,362 patients with SPN were included in the study,and they were divided into establishment group（group A）and validation group（group B）randomly.The data in group A were analyzed by Nonparametric rank sum test and single-factor Logistic regression analysis.Multivariate non-conditional Logistic regression analysis was performed to establish the model and validated the performance of our model used receiver operating characteristic（ROC）and k-fold cross-validation.Subsequently,data of group B were taken into the model and the remaining three models to compare the performances of these models.Data of group A was used to construct a CT and a PET prediction model,compare the area under the ROC curve of the PET/CT model with the CT and the PET.Results:There were significant differences in age of the patients,lesions’size and SUV_max,and presence of lobulation,spiculation,pleural indentation,vessel connection,calcification,vacuole sign,emphysema（p<0.05）.Multivariate non-conditional Logistic regression analysis showed that the risk factors for malignant nodules included age,SUV_max,size,lobulation,calcification and vacuole sign.The each OR values（95%Confidence intervals,CI）were 1.040（1.003-1.077）,1.566（1.234-1.987）,1.154（1.074-1.241）,4.597（1.919-11.010）,0.208（0.077-0.561）,2.851（1.137-7.149）.Logistic regression model:P=1/(1+e^-x),x=-5.590+0.039×age+0.448×SUV_max+0.144×size+1.525×lobulation-1.570×calcification+1.048×vacuole sign.The ROC analysis showed an area under the curve（AUC）of 0.914（0.870-0.949）and 0.897,0.789 for its sensitivity and specificity.The prediction accuracy and training accuracy were 0.864±0.053 and 0.898±0.009respectively.In group B,the AUC value for our model was 0.929（0.850-1.000）,and for the Herder model,PKUPH model and Brock model were 0.800（0.685-0.915）,0.655（0.482-0.827）and 0.839（0.721-0.958）respectively.The diagnostic efficiency of the model established in our study were higher than the other three models.The CT models were established as follows:P=1/(1+e^-x),x=-3.374+1.515×lobulation+0.965×pleural traction-1.226×spiculation+0.201×size.The AUC value for PET/CT model was0.929（0.850-1.000）,and for the CT and PET were 0.859（0.718-1.000）and 0.799（0.670-0.927）respectively.The diagnostic efficiency of the PET/CT model were higher than the other two methods.Conclusions:Our study showed that the risk factors for malignant nodules included age,SUV_max,size,lobulation,calcification and vacuole sigh.In predicted and analyzed the same cases,our model provides a higher diagnostic efficacy than formers.It may help clinics make more accurate decisions.