To Build A Preoperative MRI Staging Prediction Modelfor Pelvic Endometriosisbased On Bayesian Network

Part 1 A compared study between narrative reports and structured MRI reportsfor pelvic endometriosis diagnosis Purpose: To compare diagnostic accuracy between narrative and structured MRI report for pelvic endometriosis.Method: We retrospectively studied clinical and radiology data of 200 patients with pelvic endometriosis confirmed by pathology.All patients were admitted in our hospital during January 2016 to January 2019,and all underwent pelvic MRI before surgery.The pelvic was devided into three compartments,anterior,middle and posterior.17 MRI features of pelvic endometriosis were extracted by experienced gynecology and radiology experts,according to key factors of the rASRM and Enzian score for endometriosis.Structured report template was created by the above features.Narrative reports were recalled from the PACS system.Then comparision was made between the narrative reports and structured reports employing Mc Nemar test to evaluate the sensitivity and specificity,focus on MRI features and compliance with pahtology.If p was less than 0.05,then it was considered statistically significant.Result:(1)In 200 patients,there were 1,129 lesions,including 512 lesions in the posterior compartment(45.35%),610 lesions in the middle compartment(54.03%),and 7lesions in the anterior compartment(0.62%).The most frequentlyinvolved sites,including uterosacral ligaments(26.66%),the ovaries(25.24%),the rectouterine pouch(15.85%),the posterior wall of the uterus(12.75%),the rectosigmoid colon(10.27%),the rectovaginal space(4.78%)and the vagina(1.68%).(2)Compared with the narrative report group,structured report group got higher sensitivity(p<0.001),but had similar specificity.(3)Compared with the narrative report group,the sensitivity of posterior compartment was significantly higher in structured report group(p<0.001),while there was no different in specificity(p=0.115).As for sensitivitystructured report group was significantly higher in posterior wall of the uterus(p<0.001),rectovaginal space(p<0.001),vagina(p=0.002),rectosigmoid colon(p=0.001),and rectouterine pouch(p=0.001).Specificity was lower for the posterior wall of the uterus(p=0.021).Specificity was no different for rectovaginal space,vagina,rectosigmoid colon and rectouterine pouch(p=0.180,0.648,0.839,0.001).(4)Compared with narrative report group,structured report group had higher sensitivity(p<0.001),and similar specificity(p = 0.332)as for middle compartment,the left and right uterosacral ligaments’ s sensitivity increased(p< 0.001),specificity was no different(p> 0.05).(5)There was no different in sensitivity or specificity between the narrative report group and the structured report group as for anterior compartment(p> 0.05).Conclusion:For the diagnosis of pelvic endometriosis,sructured report has significantly higher sensitive than narrative report.Moreover,structured report has more potential value for treatment planning for endometriosis.Part 2 To build a preoperative MRI staging prediction model for pelvic endometriosis based on Bayesian networkPurpose : To establish preoperative MRI staging modelfor pelvic endometriosis patients based on Bayesian network(BN),aming toidentify the severepatients earlyand ease the future treatment planning.Method:Of 200 patients with pelvic endometriosis were staged surgically correlation with the revised American Society for Reproductive Medicine score(r ASRM),and underwent pelvic MRI before surgery.All patients were divided into the mild-moderate group(n=31)and the severe group(n=169).The Bayesian network staging model were established as follow:(1)18 independent variables(age and 17 features of endometriosis)were extracted;(2)Preprocessing independent variables(data discretization);(3)Choosing an Algorithm(Tree Augmented Naive Bayes,TANB);(4)Training and fitting the Model;(5)The model was evaluated by the k-fold cross validation method(Receiver operating characteristic curve(ROC),area under curve(AUC),sensitivity,specificity and accuracy were choosen for evaluation indexes);(6)Using fitted model for predictions with preoperative structured MRI reports;(6)Evaluating the predicting effects of the model(sensitivity,specificity and accuracy).One-sample t test was used to analyze independently in two groups for age,course of disease,intraoperative blood loss,and operative times.Chi-squared test or Fisher’s exact probability was used to analyze independently for operation methods in two groups.Wilcoxon rank sum test was used to analyze independentlyin MRI features between the two groups.If p was less than 0.05,then it was considered statistically significant.Result:There was no significant difference in age,course of disease,intraoperative blood loss or operation methods between the two groups(all p values were > 0.05).The operation times of severe patients were longer than that of mild and moderate patients(p=0.014).The posterior wall of the uterus,rectosigmoid colon,rectouterine pouch,left ovary,right ovary,left and right uterosacral ligaments were more likely involved in severepatients(All p< 0.05).The Strength of the influence of independent variables for staging results were ranked by the bayesian network model,rectouterine pouch(0.677478),right uterosacral ligamentst(0.433668),posterior wall of the uterus(0.377358),left ovary(0.368732)and right ovary(0.298518)were top five.After cross-validation of the model for pelvic endometriosis,the staging accuracy of the model was confirmed to be 90.00%,and the AUC was 0.895973.The staging accuracy remained up to 90.00% when the preoperative model was tested with structured MRIreports.Conclusion:The Bayesian networkmodel for endometriosis established using TANB algorithm was cross-validated,and tested by preoperative MRI structured reports.The staging accuracy was 90.00%,which proved that the model had great application value.