Multiparametric MRI For The Risk Stratification Of Endometrial Carcinoma

PART Ⅰ A study with different MRI sequences to evaluate cervical stromal invasion in endometrial carcinomaObjective:To compare the value of T2 weighted image(T2WI)combined with diffusion weighted image(DWI)and T2WI combined with contrast enhanced T1 weighted image(CE-T1WI)of multiparametric magneticresonance imaging(mpMRI)in assessing cervical stromal invasion(CSI)of endometrial carcinoma(EMC)Materials and methods:One hundred and thirty consecutive patients with EMC confirmed by surgery and pathologywere enrolled in this retrospective study.This study was approved by the Institutional Review Board and the requirement for informed consent was waived due to retrospective nature.Two radiologists reviewedmpMRI respectively.CSI was evaluated by T2WI combined with DWI sequences,and thenT2WI combined withCE-TIWI sequences one month later.The associations of age,deepmyometrialinvasion(DMI),lympho-vascular space invasion(LVSI),grade and lymphadenopathywith CSI were compared by using the t test and chi-square test(p<0.05).Reliability was measured by Kappa test.The accuracy,sensitivity,specificity,positive predictive value and negative predictive value of T2WI combined with DWI and T2WI combined with CE-T1WI in diagnosing CSI were calculated,respectively.Diagnostic performance was evaluated with thearea under the curve(AUC)of the receiver operating characteristic(ROC)and difference was compared by DeLong test.Results:Histopathological examinations revealed CSI in 31(23.8%)of the 130patients.The difference between age of patients with and without CSI was not statistically significant(54.35±9.56 vs 55.72±9.49 years,t=0.856,p=0.396).DMI and grade were associated with CSI(theχ2 value were 13.31 and 6.32 with p values of less than 0.001 and 0.012,respectively).While LVSI and lymphadenopathywere not associated with CSI(the χ2 values were 0.81 and 2.06 with p valuesof 0.367and 0.151,respectively).The consistency between T2WI combined with DWI and T2WI combined with CE-T1WI was 0.831 and 0.772,respectively.The accuracy,sensitivity,specificity,positive predictive value and negative predictive value of T2WI combined with DWI for evaluating CSIwas 92.3%,80.6%,96.0%,86.2%and 94.1%for Reader 1,and 92.3%,87.1%,93.9%,81.8%and 95.9%for Reader 2.On the basis of T2WI combined with CE-T1WI,reader 1 achieved accuracy,sensitivity,specificity,positive predictive value and negative predictive for CSIof 85.4%,67.7%,90.9%,79.0%and 90.0%,respectively.Reader 2 accomplished 83.1%,64.5%,88.9%,64.5%and 88.9%.The ROC-AUC of T2WI combined with DWIfor CSI was higher than that of T2WI combined with CE-T1WI(reader 1:0.883[95%CI:0.815-0.933]vs 0.793[95%CI:0.713-0.859],p=0.006;reader 2:0.905[95%CI:0.841-0.950]vs 0.767[95%CI:0.685-0.837],p=0.001).Conclusion:The diagnostic performance of T2WI combined with DWI is superior to T2WI combined with CE-T1WIin assessing CSI in patients with EMC.PART Ⅱ Predicting the grade of endometrial carcinoma using multiparametric MRI-basedradiomics machine learning modelObjective:To predict the histological grade of EMCusing multiparametric MRI-based radiomics machine learning modelMaterials and methods:One hundred and forty-four consecutive patients with EMC confirmed by surgery and pathology were enrolled in this retrospective study.This study was approved by the Institutional Review Board and the requirement for informed consent was waived due to retrospective nature.Radiomics features were extracted from mpMRI(including T2WI,CE-T1WI and ADC mapping).The associationsof age,DMI and lymphadenopathywith grade were assessed using the t test or χ2 test.LASSO regression was used to determine independent risk factors of radiomicsfeatures for grade.Four machine learning models(Logistic regression[LR],naive Bayes[NB],support vector machine[SVM]and random forest[RF])were developed to predict grade of EMC based on clinicopathological metrics and radiomics features.Every model wassubdivided into a training cohort and a test cohort with a ratio of 7:3 andrepeated randomly for one hundred and one times Performance of models was measured using AUC of ROCResults:There were 33(22.9%)patients with high grade EMC and 111(77.1%)with low grade EMC.Difference of age between patients with high grade and lowgrade EMC was not statistically significant(p=0.947),while both DMI and lymphadenopathywere correlated with grade(p<0.001).Five independent risk factors for gradewere identified by LASSO regression analysis from radiomics features.The median AUC values of RF,LR,SVM andNB machine learning models based on DMI,lymphadenopathy and the five radiomics features were 0.882(IQR:0.111),0.714(IQR:0.118),0.677(IQR:0.093)and 0.718(IQR:0.124),respectively.The difference was statistically significant and pare comparisons revealed that the diagnostic performance of RF model was high than the remaining three models with statistical significance.The best model was RF among the four modelsConclusion:The mpMRI-based radiomics machine learning models based can achieve high diagnostic performance to predict gradePART Ⅲ Predicting lympho-vascular space invasion inendometrial carcinoma using multiparametric MRI-basedradiomics machine learning modelObjective:To determine the feasibility of using multiparametric MRI-based radiomics machine learning model to predict LVSI in patients with EMCMaterials and methods:One hundred and forty-four consecutive patients with surgery and pathological confirmed EMC were enrolled in this retrospective study This study was approved by the Institutional Review Board and the requirement for informed consent was waived due to retrospective nature.Radiomics features were extracted from mpMRI(including T2WI,CE-T1WI and ADC mapping).The associationsof age,DMI,grade and lymphadenopathywithLVSI were assessed using the t test or χ2 test.LASSO regression was used to determine independent risk factors of radiomicsfeatures for LVSI.Four machine learning models(LR,NB,SVMand RF)were developed to predict LVSI based on clinicopathological metrics and radiomics features.Every model wassubdivided into a training cohort and a test cohort with a ratio of 7:3 andrepeated randomly for one hundred and one times.Performance of models was measured using AUC of ROCResults:Histopathologicalexaminations identified LVSIin 32(22.2%)of the 144 patients.Age,DMI,lymphadenopathy and grade were statistically associated with LVSI.Five independent risk factors were identified by LASSO regression analysis from radiomics features.The median AUC values of RF,LR,SVM andNB machine learning based on age,DMI,lymphadenopathy,grade and the five radiomics features were 0.840(IQR:0.123),0.718(IQR:0.108),0.524(IQR:0.066)and 0.601(IQR 0.127),respectively.The difference was statistically significant and pare comparisons revealed that the diagnostic performance of RF model was high than the remaining three models with statistical significanceConclusion:Multiparametric MRI-based radiomics machine learning modelcan preoperatively predict LVSI in EMC with high diagnostic performance,which can be a useful image biomarker for the risk stratification of EMC to provide decision support for patients’ precision medicine.