Predicting Myometrial Invasion In Endometrial Cancer Based On Whole-uterine MRI Radiomics

Objective:Whole uterus segmentation was performed,and features were extracted based on sagittal T2-weighted imaging(T2WI)and axial diffusion-weighted imaging(DWI).To evaluate whether these whole-uterine MRI radiomics features can predict myometrial invasion depth in endometrial cancer(EC).Materials and methods:This is a retrospective study.One hundred and sixty-three EC patients confirmed by pathology underwent 3.0 T MR imaging before operation from February 2011 to September 2019,including 44 patients with deep myometrial invasion(DMI).The average age was 57.34±8.71.The sagittal T2-weighted images,axial diffusion-weighted images and both of them were independently diagnosed by two senior radiologists,and the judgement were made through a third radiologist when the review opinions were not uniform.The whole uterus was segmented on each slice as an region of interest(ROI)by two radiologists independently on Deepwise scientific research platform.And the radiomics features were extracted based on the generated three-dimensional(3D)volume of interest(VOI).Then the features were selected through multiple feature dimension reduction methods,such as the F-test and L1 regularization.And the logistic regression(LR)classifier algorithm were used to establish the radiomic models,which were verified by ten-times five-fold cross validation.T2WI model,DWI model and combined model,were established based on sagittal T2-weighted images,axial diffusion-weighted images,and the combination of both,respectively.The same method was used to segment the primary EC lesions in sagittal T2-weighted images and axial diffusion-weighted images as 3D VOI.A kappa consistency test was used to compare the subjective diagnostic results of the two radiologists.The McNemar test was used to compare the sensitivity and specificity of the subjective diagnosis among different sequences.Intra-and interclass correlation coefficients(ICCs),obtained through reliability analysis,were used to evaluate the feature consistency of the two radiologists’ ROI segmentation.The areas under the receiver operating characteristic curve(AUCs)were assessed by the DeLong test to compare differences among the above three models.Based on the segmentation of the whole uterus or the lesion,the chi-square test was used to compare the feature composition ratios of ICC≥ 0.75 and≥ 0.90,to evaluate the stability and repeatability of the extracted features.Based on different sequences(sagittal T2WI images vs.T2WI model;axial DWI images vs.DWI model;both of them vs.combined model),the chi-square test was used to compare the diagnostic efficacy of the radiologists’ subjective diagnosis and the corresponding LR models.Results:The kappa values of the two radiologists’ subjective judgements on sagittal T2-weighted images,axial diffusion-weighted images and combined images were 0.686,0.687 and 0.726,respectively(P<0.001).After the judgment of the third radiologist,the sensitivity,specificity and accuracy of the subjective diagnosis were 0.59,0.93,0.84 vs.0.70,0.91,0.85 vs.0.68,0.90,0.84,respectively,according to the three groups of MR images.There were no significant differences in the sensitivity and specificity among T2WI,DWI and their combination(T2WI vs.DWI,P=0.134;T2WI vs.both,P=0.096;DWI vs.both,P=1.000).A total of 1132 features were extracted from each VOI of sagittal T2-weighted images,as well as axial DW images,whether delineating the whole uterus or the lesion.For the constituent ratio of features with ICC≥ 0.75 and≥ 0.90,constituent ratios for delineating the whole uterus were larger than those for delineating the lesion(P<0.05).The T2WI,DWI and combined models had AUCs of 0.76,0.80 and 0.85 in the validation set,with no significant differences in AUCs among the models(P>0.05).The single-sequence LR models had lower specificities and accuracies than the corresponding subjective diagnostic results(P<0.05),while the sensitivities were higher(P>0.05).The combined model included 24 radiomics features,and the accuracy,sensitivity and specificity were 0.83,0.77 and 0.85 for DMI.There were no significant differences compared with subjective diagnosis(P>0.05).Conclusion:Whole-uterine MRI radiomics features based on sagittal T2WI and axial DWI show potential in predicting myometrial invasion in EC.