| BACKGROUND Parotid tumors(PT)account for the largest proportion of human salivary gland tumors,approximately 73%,and are treated mainly by surgical resection.The frequency of postoperative complications and the patients’ quality of survival are closely related to the surgical method.Therefore,it is extremely important to accurately identify the benign and malignant parotid tumors before surgery.Nowadays,the preoperative diagnosis of parotid tumors is largely accomplished by the use of medical imaging,particularly multi-slice spiral computed tomography(MSCT).However,traditional imaging diagnosis is mainly based on the physician’s personal experience to analyze the tumor images and make a judgment on the nature of the tumor,which is prone to misdiagnosis.Artificial intelligence-based radiomics has emerged as a result of the advancement of technology.It is capable of extracting a large number of quantitative features from medical images of tumors,and these features contain a wealth of information reflecting the internal pathophysiological conditions of lesions.This method is more repeatable and can objectively and accurately diagnose the type of tumors.OBJECTIVE To investigate the value of three diagnostic models based on qualitative features of MSCT-enhanced scans,quantitative radiomics features and the combination of the two in identifying benign and malignant parotid tumors,and to provide new ideas and methods for clinicians to determine benign and malignant preoperatively.METHODS Preoperative MSCT-enhanced scan images and basic clinical data were gathered from 185 patients who underwent parotid mass resection at the First Hospital of Lanzhou University between January 2017 and January 2022.These patients were then at random separated into a training cohort(130 patients)and a validation cohort(55 patients).First of all,the qualitative features were studied,five qualitative features were recorded after visually analyzing the CT images of parotid tumors: whether the tumor shape was regular,whether the boundary was clear,whether the density of the tumor was homogeneous,whether the tumor showed cystic changes and where the tumor was located.Univariate and multifactorial logistic regression analyses were performed on the five features and two clinical basic information of age and sex in the training cohort,and diagnostic model A was established.The diagnostic performance of model A was then tested with the validation cohort,and its discrimination was evaluated by plotting the receiver operating characteristic(ROC)curve and calculating the area under the curve(AUC);its goodness of fit and clinical validity was evaluated by plotting the calibration curve and the decision curve analysis(DCA).The study of radiomics quantitative features came next,the 3D Slicer software was used to outline the 2D region of interest(ROI)of the tumor at each level of the CT images to generate 3D ROI.Using Python,highthroughout radiomics features were extracted from the 3D ROI.These features in the training cohort were then downscaled and screened using the Least absolute shrinkage and selection operator(LASSO)and 10-fold cross-validation approach.The Radscore for each patient was calculated using the equation created by multiplying the screened features by their corresponding regression coefficients,and the Rad-score was then utilized to create the radiomics model B.Model B and Model A were then merged to create the combined diagnostic model C.The discrimination,goodness-offit and clinical validity of the three models were compared.Finally,the best diagnostic model was chosen,and a Web version dynamic nomogram of which was created for clinical usage.RESULTS The AUC,accuracy,sensitivity,specificity and F1 values of model A were 0.856(95%CI: 0.775～0.938),83.9%,71.4%,87.3% and 0.656,respectively,in the training cohort;0.845(95%CI: 0.717～0.974),85.5%,61.5%,61.5%,92.9% and0.667,respectively,in the validation cohort.The AUC,sensitivity,specificity and F1 values of model B were 0.864(95%CI: 0.790～0.938),73.9%,85.7%,70.6% and0.585,respectively,in the training cohort;0.780(95%CI: 0.646～0.915),61.9%,84.6%,54.8% and 0.512,respectively,in the validation cohort.The AUC,accuracy,sensitivity,specificity and F1 values of model C were 0.911(95%CI: 0.845～0.978),86.9%,89.3%,86.3% and 0.746,respectively,in the training cohort;0.910(95%CI:0.827～0.994),87.3%,84.6%,88.1% and 0.759,respectively,in the validation cohort.The calibration curve and DCA curve showed that model C had the best fit and the highest clinical use value,and model A was the next best.CONCLUSIONS The diagnostic model A based on the qualitative features of MSCT-enhanced scan images had a lower ability to diagnose malignant tumors and a higher ability to diagnose benign tumors,i.e.,a higher rate of missed diagnosis and a lower rate of misdiagnosis.The ability of model B based on radiomics quantitative features to diagnose benign tumors was lower than that of model A,but the ability to diagnose malignant tumors was higher than that of the latter,i.e.,the missed diagnosis rate was lower and the misdiagnosis rate was higher.Model C,which combined qualitative and radiomics features,had the highest accuracy and lowest leakage rate compared with the other two models,and had the highest net clinical benefit for patients when used clinically.The Web version dynamic nomogram established based on this model could assist clinicians to judge the benignity and malignancy of parotid tumors more accurately before surgery so as to select the appropriate surgical method. |
PDF Full Text Request