| Part 1ObjectiveMeningiomas represent the most common intracranial neoplasms.Knowing meningioma grade before surgery is of considerable clinical importance,as it could aid in treatment decision-making and operation planning.Traditional magnetic resonance imaging(tMRI)is a routine examination for patients with meningiomas,and it could provide valuable information on tumor's size,shape,location,heterogeneity,etc.In this study,we reviewed preoperative tMRI scans of meningioma patients,and explored the possibility of determining meningioma grade based on these data.The aim of the first part of the dissertation was to assess the value of several traditional radiological features in grading meningioma.MethodsWe searched our departmental database for consecutive patients with histopathological confirmation of meningiomas between January 2015 and December 2016.After applying the excluding criteria,a total of 131 patients were enrolled in the study(including 21 patients with high-grade meningiomas,and 110 patients with low-grade meningiomas).We obtained the clinical information and preoperative digital imaging and communications in medicine(DICOM)data for these patients.We assessed the diagnostic performance of two clinical features(age and sex)and five radiological features(tumor-brain interface,peritumoral edema,tumor enhancement,capsular enhancement,and tumor shape)for the differentiation of high-grade and low-grade meningiomas.Statistical analysis was performed with univariate and multivariate analysis.ResultsMean patient age was 52.89±9.00 years(ranged from 24 to 78 years);39 patients were male,and 92 were female.The female to male sex ratio was 2.36.Mean period between preoperative MRI examinations and operations was 8.32±3.55 days(ranged from 1 to 26 days).Our analysis showed that three radiological features(i.e.,tumor shape,tumor enhancement,and tumor-brain interface)were significantly different between the two groups(p values were<0.0001,<0.0001 and =0.0004,respectively).According to the results of univariate and multivariate analysis,"irregular tumor shape","heterogeneous tumor enhancement",and "negative capsular enhancement" were significantly correlated with high-grade meningiomas.The logistic regression model also incorporated the above three factors,with the coefficients being 2.04,1.56,and 1.38,respectively.The model's AUC,sensitivity,specificity,and diagnostic accuracy were 0.88,47.62%,95.45%,and 87.79%,respectively.ConclusionTumor shape,tumor enhancement,and tumor-brain interface were three radiological features that could aid in grading meningioma.Part 2ObjectiveAs the results of the previous part of the study showed,tumor shape and tumor enhancement were two radiological features that could aid in grading meningioma.Texture and shape analysis was widely used in medical research,and could be used to evaluate tumor's heterogeneity and shape quantitatively.In the second part of the dissertation,we explored the diagnostic value of texture and shape analysis in the differentiation of high-grade and low-grade meningiomas.MethodsWe searched our departmental database for consecutive patients with histopathological confirmation of meningiomas between January 2015 and December 2016.After applying the excluding criteria,a total of 131 patients were enrolled in the study(including 21 patients with high-grade meningiomas,and 110 patients with low-grade meningiomas).We obtained the clinical information and preoperative DICOM data for these patients.Firstly,for each patient,352 radiological features(including 279 texture features and 73 shape features)were calculated;Then we selected the three most differentiating texture features(including Horzl_RLNonUni,S(2,2)SumOfSqs,and WavEnHL_s-3)and shape features(including GeoFv,GeoW4,and GeoW5b).Mann-Whitney test,ROC curve(and the corresponding AUC value),sensitivity,specificity,and diagnostic accuracy were used to assess these features.ResultsMean patient age was 52.89±9.00 years(ranged from 24 to 78 years);39 patients were male,and 92 were female.The female to male sex ratio was 2.36.Mean period between preoperative MRI examinations and operations was 8.32±3.55 days(ranged from 1 to 26 days).As the results showed,the six features were significantly different between the two groups(p values were =0.0001,<0.0001,=0.0009,=0.0006,<0.0001,and<0.0001,respectively).ROC curves deviated from the diagonal toward the top left corner;the corresponding AUC values were significantly greater than 0.50(ranged from 0.73 to 0.88).The diagnostic accuracy ranged from 69.47%to 88.55%,in which three were in the range of 80-90%and two were in the range of 70-80%.From the sensitivity perspective,GeoW5b demonstrated the best performance(90.48%),followed by Horzl_RLNonUni and GeoW4(both were 76.19%);GeoFv gave the worst value(47.62%).From the specificity perspective,GeoFv unexpectedly performed the best(96.36%),followed by S(2,2)SumOfSqs(90.00%);Horzl_RLNonUni and WavEnHL_s-3 gave the worst values(both were 69.09%).ConclusionTexture and shape analysis could provide relatively comprehensive and objective analysis of tumor's heterogeneity and shape,and thus were useful in the preoperative determination of meningioma grade.Part 3ObjectiveIn the previous part of the study,we obtained three texture features and three shape features;one limitation was that these features were assessed individually,their combined effect was not analyzed.In the third part of the dissertation,we investigated the possibility of combining machine learning algorithms with texture and shape analysis for grading meningioma.MethodsWe searched our departmental database for consecutive patients with histopathological confirmation of meningiomas between January 2015 and December 2016.After applying the excluding criteria,a total of 131 patients were enrolled in the study(including 21 patients with high-grade meningiomas,and 110 patients with low-grade meningiomas).We obtained the clinical information and preoperative DICOM data for these patients.We sequentially used the three texture features,the three shape features,and all the six radiological features to train three machine learning classifiers,and obtained a total of 9 meningioma grade prediction models.The three machine learning classifiers used were logistic regression(LR),naive bayes(NB),and support vector machine(SVM).We evaluated the models'performance with AUC value,sensitivity,specificity,and diagnostic accuracy.ResultsMean patient age was 52.89±9.00 years(ranged from 24 to 78 years);39 patients were male,and 92 were female.The female to male sex ratio was 2.36.Mean period between preoperative MRI examinations and operations was 8.32±3.55 days(ranged from 1 to 26 days).The results showed that the predicting performance of these models were generally satisfactory,with the AUC values in the range of 0.80-0.91 and the diagnostic accuracy in the range of 0.77-0.89.Models built with all six radiological features performed better than models built with the other two feature sets.From the perspective of comparing classifiers,NB and SVM models performed better than LR models;the specificities of the three were comparable,but NB and SVM models yielded better sensitivities(0.762,0.857 vs 0.667).SVM models performed better than NB models,with the sensitivity,specificity,diagnostic accuracy,and AUC value being 0.857,0.873,0.870,and 0.865,respectively.ConclusionBased on the results of the above research,we think preoperative tMRI is applicable in grading meningioma.Computer-assisted quantitative analysis has compelling advantages in terms of speed,automation,objectiveness,and precision,all of which are beneficial for clinical application.As the results of the last two parts of the dissertation have showed,the performance of models obtained by training machine learning classifiers with texture and shape features in differentiating meningioma grade is generally satisfactory.We hope our current study could serve as a basis and reference for further related studies,and we would be glad to see the method proposed here be adopted in clinical practice in the future. |
PDF Full Text Request