Heterogeneity Analysis And Overall Survival Prediction Of Glioblastoma Based On Multimodal MRI

BackgroundGlioblastoma(GBM)is the most common primary malignant brain tumor.The standard of care is surgical resection followed by postoperative radiotherapy and chemotherapy.The median survival time of GBM is about 15 months.High heterogeneity is an important factor leading to the poor prognosis of GBM,so how to mine the heterogeneity information related to prognosis is an important issue in clinical research..At present,the evaluation of GBM heterogeneity mainly depends on invasive pathology with potential sampling error and it is impossible to comprehensively evaluate the peritumoral edema.In contrast,magnetic resonance imaging(MRI)can noninvasively and comprehensively evaluate the heterogeneity information of the whole region of the lesion(enhanced tumor,non-enhanced tumor and edema).In our study,multimodal MRI is used to mine the heterogeneity information related to the prognosis of GBM from the two dimensions of"vision"and"sub-vision",comprehensively analyze and construct the prediction model of overall survival(OS).Purpose1.To evaluate and compare the stability of VASARI feature sets before and aftersimplification,and explore the relationship between simplified VASARI features and the OS of GBM patients.2.To compare the prognostic stratification value of GBM structural subregions(enhanced tumor(ET),non-enhanced tumor(NET)and edema(ED))based on conventional MRI.3.To explore the heterogeneous pattern of ED and clarify its prognostic stratificationvalue based on conventional MRI.4.To explore the heterogeneous pattern of the whole tumor(WT)and ED,andcompare the prognostic stratification value of WT structural subregions(WT＿s ROIs),WT functional subregions(WT＿f ROIs)and ED functional subregions(ED＿f ROIs).MethodsSection 1Task 1-Evaluation of consistency of VASARI feature set before and after simplification:A total of 200 conventional MR images of GBM patients confirmed by pathology from our hospital and TCIA database were collected retrospectively.Four neuroradiologists(two residents and two attending physicians)independently completed the evaluation of VASARI features before and after simplification.Cohen Kappa method was used to analyze the consistency.Task 2-Evaluation of the relationship between simplified VASARI feature set and OS:MR images of 222 patients with IDH wild-type GBM confirmed by pathology were included retrospectively.Two neuroradiologists completed the simplified VASARI features extraction independently.After balancing the clinical confounding factors by propensity score matching(PSM),the survival differences between the simplified VASARI feature groups were evaluated by Kaplan-Meier log rank test.Task 3-Construction of OS prediction model:OS prediction model was constructed with clinical,pathological and simplified VASARI features based on Cox proportional hazards model.Section 2The conventional MR images and clinicopathological information of 129 GBM patients were collected retrospectively.Based on the segmentation results of experts in the database,radiomics features of ET,NET and ED were extracted respectively.The minimum absolute contraction and selection operator algorithm combined with Cox proportional hazards regression model(LASSO-Cox)were used for feature selection.The following radiomics signatures were constructed:Rad Score＿ET、Rad Score＿NET、Rad Score＿ED and Rad Score＿Con(all extracted features of ET,NET and ED were entered into LASSO-Cox).Kaplan-Meier log rank test was adopted to evaluate the relationship between radiomics signatures and OS.Multiple models based on the combination of clinicopathological information and radiomics signatures were constructed to evaluate the prognostic stratification value of different structural subregions.Section 3The conventional MR images and clinicopathological information of 181 GBM patients in public database and our hospital were collected retrospectively.The automatic segmentation of structural subregions(ET,NET and ED)is completed by using nn U-Net.ED heterogeneity subregions are generated by K-means clustering.Radiomics signatures were constructed based on ED and ED subregions(the same method as that in section 2).The high-risk subregions of ED were selected by regional comparison and inter-model comparison.Finally,OS prediction model was constructed with radiomics signature of high-risk ED subregion and clinicopathological information.Section 4The conventional MRI,diffusion weighted imaging(DWI),arterial spin labeling imaging(ASL),clinicopathological information of 167 GBM patients in our hospital were collected retrospectively.Based on conventional MRI,nn U-Net was adopted to automatically segment the whole tumor to different structural subregions(ET,NET and ED).Based on the functional MRI parameter maps(apparent diffusion coefficient map(ADC)and cerebral blood flow map(CBF)),the whole tumor and ED were clustered by K-means to obtain the corresponding functional subregions.Radiomics features were extracted from the whole tumor structural subregions(WT＿s ROIs),whole tumor functional subregions(WT＿f ROIs)and ED functional subregions(ED＿f ROIs),and LASSO-Cox was performed for feature selection and radiomics signatures(Rad Scores)construction.Univariate and multivariate Cox regression analysis was used to clarify the relationship between Rad Scores and OS and the incremental prognostic value of the prediction models.Finally,Rad Scores of different subregions were added to clinical model to construct multiple OS prediction models,and the optimal model is selected through model efficiency evaluation.ResultsSection 1Task 1:Consistency assessment of VASARI feature set:Inter-observer consistency assessment of VASARI feature set before simplification:In the attending physician group,the number of“high”,“medium”and“low”consistency features were 11,11 and 7,respectively."Low"consistency features included F3,F7,F14,F22,F26,F27 and F28.In the resident group,the number of“high”,“medium”and“low”consistency features were 7,11 and 10,respectively."Low"consistency features included F3,F6,F7,F9,F13,F14,F22,F26,F27 and F28.Intra-observer consistency assessment of VASARI feature set before simplification:In the attending physician group,the number of“high”,“medium”and“low”consistency features were 14,13 and 2,respectively."Low"consistency features included F27 and F28.In the resident group,the number of“high”,“medium”and“low”consistency features were7,11 and 11,respectively."Low"consistency features included F3,F5,F6,F7,F9,F13,F14,F22,F26,F27 and F28.VASARI feature set simplification and consistency re-evaluation:Fifteen features were simplified,namely F1,F2,F3,F5,F6,F7,F9,F13,F14,F15,F22,F23,F26,F27 and F28.After simplifying the"low"consistency features,both the intra-observer and inter-observer consistencies in the attending physician and resident group were significantly improved.Task 2:Relationship between simplified VASARI features and OS:Before propensity score matching(PSM),a total of 9 features were related to OS based on Kaplan-Meier log rank test,namely SVZ involvement,F2_M,F8,F12,F15_M,F19,F21,F22 _M and F23 _M.After PSM,only 5 features were related to OS,namely SVZ involvement,F2 _M,F15 _M,F19 and F23 _M.Task 3:OS prediction model construction:The following variables were included in the model:gender,KPS score before and one month after operation,therapy,extent of resection,MGMT promoter methylation status and SVZ involvement.The C-index of model in the training and test cohort were 0.669(95%CI,0.551-0.807)and 0.701(95%CI,0.602-0.800),respectively.Section 2Based on the training cohort,Rad Score＿ET,Rad Score＿NET and Rad Score＿ED were closely related to the OS of GBM patients(all P<0.05).The OS prediction efficiency of Rad Score＿ED(C-index=0.654)is higher than those of Rad Score＿ET(C-index=0.632)and Rad Score＿NET(C-index=0.632)based on univariate Cox regression analysis.On the basis of clinical risk factors(age,gender,KPS score,and therapy),adding a radiomics signature can improve the OS prediction performance of the model.Moreover,compared with the whole-tumor radiomics signature(Rad Score＿Con),the model constructed by the integrated multi-regional radiomics signature(Rad Score＿ET+Rad Score＿ED+Rad Score＿NET)achieved the best predictive performance,and the net benefit for patients was higher than the clinical model.Section 3The ED was clustered into two subregions based on the global matrix by using K-means clustering,and the radiomics strategy was performed to construct three radiomics signatures,namely Rad Score＿ED,Rad Score＿ED＿Label 1 and Rad Score＿ED＿Label 2.At the regional level,Rad Score＿ED＿Label 1 achieved the highest OS prediction performance,with C-indexes 0.647(95%CI:0.606-0.688)in training cohort and 0.625(95%CI:0.588-0.662)in overall cohort,respectively.At the model level,compared with Rad Score＿ED and Rad Score＿ED＿Label 2,Rad Score＿ED＿Label 1 had the highest incremental prognostic value,suggesting that ED＿Label 1 was high-risk subregion.The OS prediction model was constructed by Rad Score＿ED＿Label 1 and clinicopathological variables,with C-index of0.739(95%CI:0.698-0.780)in training cohort and 0.702(95%CI:0.582-0.822)in test cohort,respectively.Section 4Based on the standardized CBF and ADC maps,the whole tumor and ED were grouped into three functional subregions by K-means clustering.Based on the training cohort,univariate and multivariate Cox analysis indicated that the whole tumor structural subregion(WT＿s ROI＿Label 1(ET),WT＿s ROI＿Label 3(ED)),the whole tumor functional subregion(WT＿f ROI＿Label 3),and the ED functional subregion(ED＿f ROI＿Label 3)is an independent prognostic factor affecting the OS of patients.After adding Radscore of each subregions to the clinical model,Radscore＿ED＿f ROI＿Label 3 achieved the highest incremental prognostic value(C-index of model=0.866,95%CI:0.817-0.915).Finally,the model constructed by combining the clinicopathological variables with Rad Score＿WT＿s ROI＿Label 1(ET)and Rad Score＿ED＿f ROI＿Label 3 achieved the highest predictive performance,with C-index of 0.871(95%CI:0.834-0.908)in training cohort,0.881(95%CI:0.846-0.916)in test cohort,respectively.Conclusion1.Compared with the VASARI feature set,the stability of the simplified VASARIfeature set is significantly improved.After balancing for confounding factors using PSM,F2_M,F15_M,F19,F23_M and SVZ involvement were associated with OS of patients.2.The prognostic stratification value of Rad Score＿ED was better than those ofRad Score＿NET and Rad Score＿ET,suggesting that ED is an important prognostic subregion.3.The results of cluster analysis based on conventional MRI suggested that ED doeshave heterogeneity,and prognostic value of Rad Score＿ED＿high-risk subregion is higher than that of whole ED.4.Multiple heterogeneous functional subregions based on functional MRI clusteringwere obtained,and their prognostic stratification values were better than those of structural subregions.Moreover,ED high-risk functional subregion had the highest incremental prognostic value.