| Background and purpose:Glioma is the most common primary intracranial tumor and accounts for 75% of adult malignant primary brain tumors,and the 5 year overall survival rates of glioma patients are less than 35%,remaining one of the most difficult tumors to treat.Identification of glioma molecular typing can help determine the treatment strategies and predict the prognosis.With the rapid development of molecular biology technology,the understanding of glioma has also gradually reached the genetic level.Studies demonstrated that combination of pathology classification with molecular typing are more accurate to predict glioma prognosis.In recent years,classification of glioma using molecular pathological characteristics has been gradually developed and widely used in clinical practice.For example,isocitrate dehydrogenase(IDH)and co-deletion of short arm chromosome 1 and long arm of chromosome 19(1p/19 q co-deletion)were used to classify WHO II and III grade diffuse gliomas and can reflect chemotherapy sensitivity,prognosis and the risk of recurrence;IDH1 is also an independent prognostic factor of glioblastoma;O6-methylguanine-DNA methyltransferase(MGMT),EGFR,p53,PI3 K,Rb,RAF,etc,can predict prognosis and drug response.The molecular typing can provide the basis of target therapy for glioma.So far,malignant degree and molecular typing of glioma can be identified using approaches such as surgery or biopsy,which are invasive and inconvenience.Thus,it is important to explore and evaluate non-invasive manner.Studies have confirmed that the ability of tumor to induce angiogenesis correlates with the ability of cells to invade.Magnetic resonance perfusion weighted imaging(PWI)can be used to measure local tissue hemodynamic parameters to evaluate tissue perfusion status and vascular permeability.Therefore,PWI technique plays an important role in determining the malignant degree of glioma.Recent studies have demonstrated that MRI plays important roles in molecular typing of glioma,but MRI based on single parameter used to assess molecular typing of glioma are unstable and incomplete.The multiparametric characteristics of magnetic resonance perfusion can comprehensively evaluate the morphological and functional characteristics of tumor microvasculature,so that it can enhance to the accuracy of identification for molecular typing of glioma.In addition,the traditional image measurement methods provide limited information,and the new generation of artificial intelligence(AI)technology based on big data has brought new chances to the medical industry.The research method of disease-centered imaging genomics is to combine machine learning with imaging to display gene changes so that achieve the goal of precision medicine.The low spatial resolution limits the development of radiomics based on functional magnetic resonance imaging.However,with the continuous upgrading of AI software,the imaging features of functional magnetic resonance are gradually excavated,and a more optimized model needs to be established.This study first intends to analyze the MRI perfusion and histologic microvascular parameters of glioma with different molecular subtypes,and to study the pathological basis of multiparametric perfusion imaging and its role in differentiating different molecular subtypes of glioma and evaluating the prognosis of patients.Subsequently,we will build an imaging model based on magnetic resonance functional imaging and habitat imaging to evaluate the IDH1 molecular typing of glioma,and strive to improve the efficacy of traditional MRI sequences,so that more accurately judge the prognosis of glioma patients and guide the formulation of clinical treatment.Materials and Methods:Part ?: Evaluation of microvessel characteristics of glioma using multiparametric magnetic resonance perfusion imaging and its application in molecular typing(1)Cases with suspect gliomas were enrolled to perform the prospective study.The patients prior to operation were scanned by one or two(the interval between the two methods was from 24 to 72 hours)MRI perfusion approaches,including dynamic susceptibility contrast-enhanced(DSC)imaging,vessel size imaging(VSI)and dynamic contrast-enhanced(DCE)imaging.After operation,IDH1 mutation,1p/19 q co-deletion and MGMT methylation status were detected in the cases with confirmed gliomas,and finally MR imaging and molecular typing data of 161 patients with diffuse infiltrating gliomas were collected.(2)The paraffin samples from 30 patients who received both traditional DSC perfusion scan and VSI scan and 29 patients who received DCE perfusion scan were prepared into immunohistochemical sections to detect CD34 expression.The microvascular diameter,microvascular area(MVA)and microvascular density(MVD)were measured in vascular proliferation rich area in each patients.(3)By two neuroradiologists using different workstations and post-processing software to analyze perfusion scanning images: using hot spot method to measure the VSImax and VSImean values of patients;use Extended Tofts,a two-compartment model to calculate the four quantitative parameters of DCE : Volume transfer constant(Ktrans)?volume fraction of extra-vascular extracellular space(Ve)?Kep and Vp;The relative cerebral blood volume(r CBV)values were obtained by measuring the maximum CBV values of tumor regions and contralateral normal brain tissues.Intraclass correlation coefficient(ICC)analysis is used to evaluate the consistency of values measured between different observers.(4)Pearson correlation was used to analyze the correlation between various perfusion scanning parameters(VSI,r CBV,Ktrans,Kep,Ve,Vp)and histological microvascular parameters(microvascular diameter,MVA?MVD)to find the pathological basis of each perfusion parameter.(5)According to the 2016 World health organization(WHO)classification,we divided all patients into two groups: diffuse infiltrating lower grade gliomas(LGG)and glioblastoma multiforme(GBM).Mann-Whitney U test was used to analyze the difference of VSI value,quantitative parameter value of DCE and r CBV value between IDH1 mutant and wild-type gliomas,1p/19 q co-deletion and 1p/19 q intact gliomas,and MGMT methylation and unmethylation gliomas,respectively.(6)ROC analysis was used to test the efficacy of perfusion parameters on glioma typing.The maximum value of Jordan index(sensitivity + specificity-1)was used as the cutoff value.The correlation of clinical and imaging indexes with IDH1 was analyzed using multivariate logistic regression.Chapter 2: The role of multiparametric magnetic resonance perfusion imaging in evaluating the prognosis of LGG patients(1)The MR perfusion imaging(VSI or DCE scan),clinicopathological data(including sex,age,tumor location,adjuvant therapy,resection degree,histopathological type),IDH1 mutation status,1p/19 q combined deletion status,patient's Progression-Free survival(PFS)and overall survival(OS)were retrospectively analyzed and recorded in 60 cases with WHO II and III grade gliomas.(2)The LGGs were divided into VSI-high and VSI-low groups(dividing by cutoff values of differential IDH1 typing),grade WHO II and grade WHO III groups,IDH1 mutant and wild-type groups.Kaplan-Merier curve was used to analyze PFS and OS(follow-up time was 4 years),and the difference was detected using Log-rank test.(3)Univariate Cox regression was used to assess risk factors for PFS of patients,including gender,age,tumor location,adjuvant therapy,degree of resection,histopathological type,IDH1 mutation status,1p/19 q co-deletion,VSI values,and DCE parameter values.On the basis of univariate Cox analysis,we established four models of multivariate Cox analysis to further find independent prognostic factors in LGG patients.Since there are a lot of censored data in the OS,we did not analyze the OS of each factor using Cox regression.Part ?: Prediction of IDH classification of GBM based on radiomics and habitat imaging(1)The preoperative imaging data and IDH1 mutation information of 92 patients diagnosed as GBM were retrospectively included.(2)Extraction of radiomics feature: spatial registration of different sequences using MATLAB software,and then the registered image was imported into the AKMITK Work software(GE pharmaceutical industry)for image processing.When segmentation,we used T2 WI,FLAIR,T1 WI and contrast-enhanced imaging to divide the tumor into whole regions.The enhancement area and the edema area were marked separately,and the image features of each segmentation on the T2 WI,CBV and ADC were extracted,including 396 features consisted of histogram,morphological features(sphericity,compactness,etc.),Gray-Level Co-Occurrence Matrix(GLCM),Gray-Level Run-length Matrix(GLRLM),etc.(3)Radiomics feature analysis: two feature selection methods,m RMR and LASSO,were used for data dimension reduction and screening of radiomics feature,and the radscore was established using the minimum penalty coefficient corresponds to the number of features.(4)Model construction and validation: based on radiomics features from different areas of multiparameter,we established four models,including: T2 WI model,CBV model,ADC model,and MRI multi-parameter multi-area combination model.Clinical factors of age,gender,and lesion site were included in the four models.Finally,the performance of the model was evaluated by ROC analysis and Hosmer-Lemeshow test.Results:Part ?:(1)VSImax and VSImean values showed a good inter-group observer consistency(ICC=0.955 and 0.923).Through Person correlation analysis,there was a significant positive correlation between VSImean and microvascular diameter,VSImean and MVA,VSImax and microvascular diameter,VSImax and MVA(0.01),among which VSImean was most correlated with microvascular diameter(r =0.8432).r CBV values also showed good inter-group observer consistency(ICC=0.862),and r CBV was most correlated with microvascular area(r = 0.6579).There was no significant correlation of both VSI value and r CBV with MVD.There was also good inter-observer consistency among different parameters of DCE(a range of ICC value,0.71-0.809).We found that significant positive correlations of both Ktrans and Vp with MVA(P < 0.05),of Ktrans,Ve and Vp with microvascular diameter and MVD(P < 0.05),among which Vp was most correlated with MVA(r = 0.89),followed by Vp with MVD(r = 0.6386)and Ktrans with MVA(r = 0.6013).(2)We found that the VSImax and VSImean values of IDH1 mutant LGGs were significantly lower than those of IDH1 wild-type LGGs(P < 0.01),and the VSImax and VSImean values of IDH1 mutant grade WHO II gliomas were also significantly lower than those of IDH1 wild-type grade WHO II gliomas.No significant difference was found between IDH1 mutant and wild-type WHO III grade gliomas.Moreover,the VSImax and VSImean values of grade III gliomas were significantly higher than those of grade II gliomas.There was no significant difference of VSI values between wild-type GBM and mutant GBM.(3)ROC curves were established and an area under the curve(AUC)of 0.7305 for the VSImax values in distinguishing IDH1 mutation from wild-type LGGs,with 62.79%sensitivity,82.35% specificity and cut-off value of 112.8?m;the AUC of 0.7401 for the VSI mean values with 65.12% sensitivity,82.35% specificity and cut-off value of 78.5?m.A stepwise logistic regression analysis showed that age,tumor location,the VSImean values were associated with IDH1,and when they were combined using the formula,the AUC was elevated to 0.7798 in differentiating IDH1 mutant from wild-type LGGs.However,our data indicated that gender and the VSImax values were not independent predictors of IDH1 mutation.(4)We found that the VSImax and VSImean values in grade II gliomas with 1p/19 q co-deletion and IDH1 mutation were lower than those in 1p/19 q intact and IDH1 mutant grade II gliomas.No significant difference of VSI values was observed between LGGs or GBMs of MGMT methylation and unmethylation.(5)We found that IDH1 wild-type LGGs had a higher Ve than IDH1 mutant LGGs(P =0.0486),but Ktrans,Kep and Vp had no difference between IDH1 wild-type and mutatn LGGs;The Ktrans values of IDH1 wild-type GBMs were higher than those of IDH1 mutant GBMs(P = 0.015),but Kep,Ve and Vp values had no difference;both 1p/19 q co-deletion LGGs and grade II gliomas had higher Vp values than 1p/19 q intact ones(P = 0.0079,0.004);the Kep values in GBM with MGMT methylation was significantly lower than in GBM with MGMT unmethylation,whereas the Ve and Vp values were higher(P = 0.008,0.05).However,there was no significant difference in the Ktrans values between the two groups.(6)The r CBV values of IDH1 wild-type GBMs were larger than those of IDH1 mutant GBMs,and GBMs with MGMT methylation had lower r CBV values compared with GBMs with MGMT unmethylation,but there was no significant difference(P = 0.053/0.072).Part ?:(1)After more than four years of follow-up,prognostic information of 51 in 60 LGG patients undergoing VSI perfusion scan was collected.Kaplan-Meier curve analysis showed that the average OS of the VSImax-high and VSImean-high groups was significantly shorter than those of the VSImax-low and VSImean-low groups;the average PFS of the VSImax-high and VSImean-high groups was significantly shorter than those of the VSImax-low and VSImean-low groups,respectively.PFS and OS of patients with IDH1wild-type LGGs were also significantly shorter than those of patients with IDH1 mutant LGGs.In addition,PFS and OS of patients with WHO grade III gliomas are shorter than those of patients with WHO grade II gliomas.However,there was no significant difference of PFS and OS between patients with IDH1 wild-type WHO II and WHO III gliomas.(2)Univariate Cox analysis revealed that VSImax values,VSImean values,IDH1 mutation status,WHO grade,age,adjuvant treatment modalities,and multiple lesions or lobes were associated with PFS,but gender,histopathological type,1p/19 q co-deletion and resection extent were not significantly associated with PFS.On the basis of univariate Cox analysis,we developed four multivariate Cox risk proportion models and found that multiple lesions or lobes,IDH1 mutation status,and VSImean values were independent risk factors of PFS in different models.(3)We obtained prognostic information of 22 in 47 patients who underwent DCE perfusion scans.Univariate Cox analysis showed that Ktrans and Ve correlated with PFS(P =0.0051 and 0.0047,respectively),and Hazard Ratio(HR)values were 39.607 and 8.9779,respectively.Meanwhile,Ktrans correlated with OS(P = 0.019),and HR value was 36.608.Part ?:(1)A total of 92 GBM patients were randomly divided into training set(n = 66)and validation set(n = 26)according to proportion of 7:3.(2)ICC values of different observers ranged from 0.81 to 0.92,suggesting a good consistence in the observers.When the clinical indicators were analyzed using Mann-Whitney U test,we found that only age had significant difference between IDH1 wild-type and mutant groups.Diagnostic efficacy of the clinical indicators were all lower than that of radiomis.Multiple logistic regression was established based on combination of radscore with clinical indicators of statistical significance.(3)The number of the most predictive feature subsets obtained by LASSO analysis in T2 WI,CBV,ADC and multi-parametric modles was 5,6,2 and 5,respectively.Radscore of IDH1 wild-type and mutant GBMs in each model was significantly different.(4)ROC analysis showed that the AUC values of radscore of CBV model in training and test groups were higher than those of T2 WI model and ADC model,both of which were 0.955.The AUC value of the multi-parameter multi-area model was 0.962 in the training group and0.955 in the test group,and the diagnostic efficacy of this model was higher than that of the single sequence model.The AUC values of clinical indicators in the training and test groups in each model ranged from 0.773 to 0.864.(5)In T2 WI model,combination of radiomics with age got the highest accuracy(0.84 in training set and 0.92 in test set)in predicating IDH1 mutation.In CBV model,the highest accuracy(0.9375 in training set and 0.92 in test set)was observed in radiomics feature.In ADC model,combination of radiomics feature with clinical indicators reached the highest accuracy(0.84 in training set and 0.85 in test set).In muti-parameter multi-area model,the accuracy in radiomics is the same as that in combination of radiomics with clinical indicators(0.969 in training set and 0.923 in test set).(6)Hosmer-Lemeshow test showed that the P values were more than 0.05 for AUC values of the test groups in each model,suggesting that each model established had good degree of fitting.Conclusion:In this study,we analyzed the effect of multi-peremetric perfusion imaging on glioma molecular typing and patient prognosis,and demonstrated that VSI,DCE perfusion and conventional DSC perfusion as non-invasive markers can predict glioma IDH1 mutation,1p/19 q deletion and MGMT methylation status,and VSImean,Ktrans and Ve were associated with the outcome of LGGs.The VSImean value was an independent prognostic factor of LGG patients.Bases on CBV radiomics models and multi-sequence multi-region radiomics models,the traditional imaging models were further optimized,and showed good efficacy in predicting GBM IDH1 mutation status.These may provide a favorable basis for personalized diagnosis and treatment of glioma,further promoting the goal of precision medicine. |
PDF Full Text Request