| Part I The value of MRI VASARI feature analysis and Radiomics in grading for diffuse gliomasBackgroundGliomas are the most common malignant primary brain tumors in adults,of which diffuse gliomas have the highest incidence,showing extensive infiltrating growth into the central nervous system parenchyma.In 2016,the World Health Organization(WHO)classified diffuse gliomas as grade Ⅱ,Ⅲ and Ⅳ,of which grade Ⅱ-Ⅲ gliomas as lower-grade gliomas(LGG)and grade Ⅳ glioma as high-grade glioma(HGG).Due to large differences in treatment strategies between different grade of gliomas,preoperative classification of gliomas is critical.Pathological diagnosis after biopsy or surgery is the main gold standard.However,the inevitable sampling errors and invasive procedures may pose some risks to glioma patients.Moreover,this histological examination is usually time-consuming,making it difficult to grade gliomas in a timely manner.Magnetic resonance imaging(MRI),with its high image resolution and soft tissue contrast,has become a standard diagnostic tool for evaluating brain tumors,and plays an important role in the preoperative diagnosis,treatment planning,efficacy evaluation and follow-up of gliomas.However,due to the ambiguity of qualitative imaging analysis and the lack of reproducible and verified objective indicators,tumor grading remains certain challenges.Although advanced MRI technologies such as diffusion tensor imaging,susceptibility weighted imaging,magnetic resonance spectroscopy and perfusion weighted imaging can provide us with more information about water molecular diffusion,microbleeds and neovascularization,local tissue metabolism and hemodynamics in tumor,playing a better auxiliary role in tumor grading.However,multi-modal MRI scan will take a long time and is complex in image post-processing,making them unsuitable as conventional scan sequence.The Visually Accessible REMBRANDT Images(VASARI)feature set is based on conventional MRI T1 weighted image(T1WI),T2 weighted image(T2WI),T2 fluid attenuated inversion recovery(T2FLAIR),diffusion weighted image(DWI)and contrast enhanced T1 weighted image(CE-T1WI),extracting the imaging features of glioma.The 30 qualitative imaging features of gliomas were quantified in a standardized manner,which was helpful to avoid the errors of subjective factors in previous studies.At the same time,the extraction process of quantitative features is standardized,so that the extracted features are more stable and easily to understand and repeatable.With the widespread application of radiomics,there will be opportunities to help us make better use of quantitative information from millions of voxels on MRI images to improve the accuracy of histological grading of diffuse gliomas.ObjectiveIn this study,VASARI features and radiomics features of preoperative multi-parametric MRI images were used to achieve a non-invasive,simple and reproducible method to predict the grade of diffuse gliomas before surgery effectively.Material and Methods1.The clinical information and preoperative MRI images of 135 patients with diffuse gliomas confirmed by surgical pathology were retrospectively analyzed.Patients were divided into a training group(n=108)and a validation group(n=27)at a ratio of 8:2.Two neuroradiologists extracted the preoperative VASARI features of all patients and then performed the quantitative analysis of imaging features.Mann-Whitney U test,Chi-square test or Fisher’s exact test were used to evaluate the statistical differences in imaging features of lower-grade gliomas and high-grade gliomas.The VASARI features associated with glioma grade were selected.The least absolute shrinkage and selection operator(LASSO)method was used to further screen VASARI features closely related to the grade of diffuse gliomas.Support vector machine(SVM)classifier was used for training data in training group,and all single VASARI feature models and a multiple feature combined model were established respectively and the models were verified in the validation group.2.The preoperative MRI images of all the above patients were analyzed,including T1WI,T2WI,T2FLAIR,DWI and CE-T1WI.The radiomics features were extracted from the above MRI images.Variance threshold method,univariate selection method and LASSO methods were applied to reduce the feature dimension.Then the radiomics features most related to the grade of diffuse gliomas were gradually screened.The SVM classifier was used to train the data in training group,and all single sequence models and a multi-sequence combined radiomics model were established respectively.Then the models verification were performed on the validation group.3.The VASARI features and radiomics features were combined to construct combined model.ROC curve was used to evaluate the predictive performance of each model.Results1.Among all VASARI features,19 VASARI features were correlated with glioma grade by univariate analysis.After dimension reduction by LASSO regression,6 VASARI features were closely related to the grade of glioma.Their names and coefficients were F24(satellites),0.218;F11(thickness of enhancing margin),0.098;F6(proportion non-enhancing component),-0.072;F16(hemorrhage),0.060;F9(multifocal or multicentric),0.035;F29(leasion size),0.004.Based on these 6 VASARI features,a prediction model and a combined model Comb1 for glioma grading were constructed respectively.For the single VASARI feature model,F11 had the highest AUC value of 0.873(95%CI:0.815-0.932)in the training group,and the sensitivity,specificity and accuracy were 84.85%,81.82%and 83.12%,respectively.In the validation group,the AUC value was 0.848(95%CI:0.789-0.907)and the sensitivity,specificity and accuracy were 86.67%,77.78%and 81.82%,respectively.For the combined model,the AUC value was 0.889(95%CI:0.836-0.942),and the sensitivity,specificity and accuracy were 93.94%,84.10%and 87.01%in the training group.In the validation group,the AUC value was 0.861(95%CI:0.804-0.918),and the sensitivity,specificity and accuracy were 93.33%,83.33%and 81.82%,respectively.2.A total of 1409 radiomics features were extracted from the region of interest(ROI)on images of each MRI sequence,and a total of 7045 features were extracted from all the 5 MRI sequences of each patient.After variance threshold method,univariate selection method and LASSO regression for dimensionality reduction,the radiomics features closely related to glioma grade were screened for each sequence.When the SVM classifier was used for training,the AUC value of the prediction model based on CE-T1WI sequence was the highest(0.841,95%CI:0.786-0.896),and the sensitivity,specificity and accuracy were 82.35%,78.72%and 74.07%,respectively.In the validation group,the AUC value was 0.795(95%CI:0.742-0.849),and the sensitivity,specificity and accuracy were 92.86%,60.00%and 68.97%,respectively.The AUC value of Comb2 model in the training group was 0.927(95%CI:0.872-0.983),and the sensitivity,specificity and accuracy were 86.11%,84.44%and 80.25%,respectively.In the validation group,the AUC value was 0.882(95%CI:0.826-0.939),and the sensitivity,specificity and accuracy were 75.00%,94.12%and 82.76%,respectively.3.For the Comb3 model,which was constructed by VASARI features and radiomics features,had the best prediction performance,with an AUC value of 0.937(95%CI:0.8800.995),a sensitivity of 91.67%,a specificity of 88.57%,and an accuracy of 88.14%in the training group.In the validation group,the AUC value was 0.900(95%CI:0.839-0.960),and the sensitivity,specificity and accuracy were 91.67%、85.19%and 82.35%,respectively.Conclusions1.VASARI feature model,radiomics model and VASARI feature combined radiomics feature model based on multi-parametric MRI can effectively predict the grade of diffuse gliomas.2.The VASARI feature and radiomics feature combined model has the best prediction performance among all the models.Part Ⅱ The value of Radiomics in predicting ATRX mutation status in lower-grade gliomasBackgroundIn recent years,molecular markers have become increasingly important in providing auxiliary diagnosis and definitive diagnostic information for gliomas,which can help us better understand the biological characteristics of this malignant tumor and improve the diagnostic level.With the development of precision medicine,molecular detection for gliomas has become an important part of guiding clinical decision-making.In 2016,the World Health Organization(WHO)introduced alpha thalassemia/mental retardation syndrome x-linked gene(ATRX)mutation as a new molecular diagnostic criteria for central nervous system tumors,and explored the significance of mutation/wild-type isocitrate dehydrogenase 1(IDH1)for glioma classification diagnosis,ushering in a new era of brain tumor classification and treatment.ATRX mutation is an important genetic status for the classification of lower-grade gliomas(LGG)and prediction of prognosis.At present,invasive methods such as biopsy and surgical resection are still the main methods to evaluate ATRX mutation status.However,the accuracy of these methods may not be high due to intra-tumor heterogeneity and insufficient tumor samples,which may affect the choice of treatment strategy.Magnetic resonance imaging(MRI),as a routine non-invasive diagnostic tool,may play a role in the gene-level classification of gliomas.Radiomics can provide a large amount of biological information about tumors,and can be used as an important auxiliary means to predict the genotype of gliomas.ObjectiveThe aim of this study is to develop a preoperative multi-parametric MRI-based radiomics model for non-invasive prediction of ATRX mutation status in LGG,and to find the optimal sequence combination and radiomics feature for predicting ATRX mutation status.Material and MethodsThe clinical records and MRI images of patients with LGG who underwent surgical treatment in our hospital were retrospectively analyzed.Patients were randomly divided into training and validation groups at a ratio of 8:2.ATRX mutation was denoted as ATRX-,and ATRX non-mutation was denoted as ATRX+.Clinical characteristics including age,gender,histological grade,histological type,tumor location and tumor side were compared between ATRX-and ATRX+patients in the training group.The region of interest(ROI)was delineated on MRI T1WI,T2WI,T2FLAIR,CE-T1WI and ADC images and then the radiomics features were extracted.Variance threshold method,univariate selection method and least absolute shrinkage and selection operator(LASSO)method were used to select radiomics features.The optimal radiomics model and radiomics features for predicting ATRX mutation status in LGG were found out through combining MRI sequences with different numbers.The clinical factors with statistically significant differences between ATRX-and ATRX+patients were included in the analysis to further construct a model combined radiomics with clinical factors.For the training group,the SVM classifier was used to establish the prediction models,and then the models were verified in the validation group.The performance of each model was evaluated by the area under curve(AUC)value and sensitivity,specificity,accuracy of the receiver operating characteristic(ROC)curve.ResultsA total of 129 LGG patients were included in the study,including 41 ATRX-patients,of which 33 in the training group and 8 in the validation group.In the training group,ATRX-and ATRX+patients were significantly different in age(P<0.05).There were no significant differences in gender,histological grade,histological type,tumor location and tumor side.1409 features were extracted from each MRI sequence;therefore,a total of 7045 features were extracted on the 5 sequence images.After feature dimension reduction,the number of features selected for each sequence were 7 features from T1WI,6 features from T2WI,9 features from T2FLAIR,14 features from CE-T1WI and 4 features from ADC map.Among different combinations of MRI sequences,T1+T2+T2FLAIR+CE-T1 has the best diagnostic efficiency for predicting ATRX mutation status in LGG.In the training group,the AUC value was 0.907(95%CI:0.851-0.962),the sensitivity,specificity and accuracy were 90.00%,85.71%and 81.08%,respectively.In the validation group,the AUC value was 0.824(95%CI:0.768-0.878),and the sensitivity,specificity and accuracy were 76.47%,87.50%and 80.95%,respectively.There were 14 optimal radiomics features,among which the size zone non uniformity from gray level size zone matrix on T2FLAIR was the most significantly related feature for predicting ATRX mutation status in lower-grade gliomas.Among the single sequence models,CE-T1 model had the best diagnostic performance.In the training group,the AUC value was 0.712(95%confidence interval:0.653-0.770),the sensitivity,specificity and accuracy were 90.00%,50.00%and 81.08%,respectively.In the validation group,the AUC value was 0.743(95%confidence interval:0.686-0.795),and the sensitivity,specificity and accuracy were 67.65%,87.50%and 81.95%,respectively.Among the two sequence combinations,the T2+T2FLAIR model had the best diagnostic performance.In the training group,the AUC value was 0.781(95%confidence interval:0.724-0.836),the sensitivity,specificity and accuracy were 90.00%,64.29%and 82.43%,respectively.In the validation group,the AUC value was 0.779(95%confidence interval:0.726-0.832),the sensitivity,specificity and accuracy were 91.18%,62.50%and 80.95%,respectively.Among the three sequence combinations,the T1+T2+T2FLAIR model had the best diagnostic performance.In the training group,the AUC value was 0.831(95%confidence interval:0.7790.882),the sensitivity,specificity and accuracy were 93.33%,71.43%and 82.43%,respectively.In the validation group,the AUC value was 0.842(95%CI:0.788-0.894),and the sensitivity,specificity and accuracy were 73.53%,87.50%and 80.95%,respectively.For the five sequence combined model,the AUC was 0.904(95%confidence interval:0.850-0.956),and the sensitivity,specificity and accuracy were 90.00%,85.71%and 81.08%,respectively in the training group.In the validation group,the AUC was 0.816(95%confidence interval:0.7590.873),and the sensitivity,specificity and accuracy were 73.53%,87.50%and 80.95%,respectively.When age was used as a clinical variable to further construct a comprehensive model combined radiomics features with age,the performance of the model was not improved.The AUC of the optimal radiomics combined age model(T1+T2+T2FLAIR+CE-T1+age)in the training group was 0.906(95%confidence interval:0.846-0.965),and the sensitivity,specificity and accuracy were 90.00%,85.71%and 81.08%,respectively.In the validation group,the AUC was 0.820(95%CI:0.759-0.880),and the sensitivity,specificity and accuracy were 76.47%,87.50%and 80.95%,respectively.Conclusions1.The preoperative multi-sequence MRI radiomics analysis of LGG patients can effectively predict ATRX status.2.Among the radiomics models established by different MRI sequence combinations,the prediction performance of multi-sequence radiomics model was higher than that of the single sequence.3.The T1+T2+T2FLAIR+CE-T1 sequence combined model has the best diagnostic performance for predicting ATRX mutation status in LGG,and size zone non uniformity from T2FLAIR gray level size zone matrix was the most significantly correlated radiomics feature for predicting ATRX mutation status in lower grade gliomas. |
PDF Full Text Request