| Objective: To explore the diagnostic value of radiomics in noninvasively predicting the status of lymph vascular space invasion(LVSI)in cervical cancer with sagittal T2 WI MRI images.A predictive radiomics model was constructed to assist clinical decision-making.Methods: 160 cervical cancer patients pathologically confirmed being with or without LVSI were enrolled in this retrospective study and randomly divided into a training and test cohort with a stratified sampling ratio of 2: 1.All patients underwent the MRI examination within two weeks before the operation.Predictive performance of the whole tumor,peritumoral areas and intratumoral subregions were separately evaluated for the LVSI status in cervical cancer patients.The tumor regions(ROIs)segmented by radiologists were radially expanded at intervals of 1 mm(up to a dilation distance of 8 mm).The cervical tumors were partitioned into spatially distinct intratumoral subregions at the patient and population level,using the K-means clustering algorithm based on T2 WI MRI images and local entropy maps.To obtain the most predictive feature subsets,a three-step feature selection method consisting of Mann-Whitney U test,LASSO regression and m RMR algorithm was applied to all radiomics feature sets extracted from the whole tumor,peritumoral areas and intratumoral subregions.Logistic regression models were built based on each feature subset to analyze and compare the predictive performance of different regions.Feature subsets of the peritumoral area with the optimal dilation distance and intratumoral subregions were fused and then further selected by the m RMR.A fusion model was then constructed based on the selected fusion features to improve the predictive performance of the LVSI status.AUC,sensitivity and specificity were used to assess the predictive performance of the model.A radiomics nomogram was constructed based on the fusion model for future clinical application.Calibration curves were plotted to evaluate the fit goodness of the fusion model.Decision curve analysis was performed to evaluate the clinical application value of the fusion model.Results: The optimal dilation distance of the peritumoral area is 6mm.The peritumoral area with 6mm dilation distance outside the tumor(denoted as Peri-6)showed the best performance with AUCs of 0.777 and 0.734 in the training and test cohorts,respectively.The cervical tumor was partitioned into two intratumoral subregions(denoted as edge subregion and core subregion).The AUCs of the logistic models constructed based on the edge subregion,core subregion,and the whole tumor region were 0.702,0.685,and 0.601 in the test cohort,respectively.The fusion model constructed based on the Peri-6,edge subregion and core subregion yielded AUCs of0.841 and 0.795 in the training and test cohorts,respectively.The calibration curve and the decision curve analysis separately demonstrated the good fit and the clinical usefulness of the fusion model.Conclusions: The fusion model established based on the peritumoral area and intratumoral subregions can effectively predict the LVSI status in cervical cancer before the surgery.It may assist clinicians in decision-making to achieve precision medicine. |
