Multimodal Magnetic Resonance Imaging And Radiomics Study Of Diffuse Midline Glioma With H3 K27M Mutation

Objective:Diffuse midline glioma,H3 K27M-mutant,a newly defined tumor subtype in the2016 WHO classification of tumors of the central nervous system,is a life-threatening disorder for human beings with a very poor prognosis,especially for children and adolescent patients.At present,few imaging studies were focused on it.However,limited reports have shown that its conventional imaging features were heterogeneous,thus it was difficult to differentiate this tumor from its wild-type counterpart through traditional review of imaging data.The purpose of this study was to explore the difference between the two groups from the following three aspects and to make a noninvasive prediction of the H3 K27 M mutation status in diffuse midline gliomas before surgical treatment.Our ultimate goal is to facilitate the precision diagnosis and treatment of the tumor.Firstly,the preoperative MRI features of diffuse midline gliomas were quantitatively analyzed by using a standardized and unified VASARI feature set,and the stability of the features were evaluated,in order to find VASARI features of diagnostic value for this tumor.Secondly,we used multimodal functional MRI to explore the quantitative imaging features,and established a LASSO regression model to select the most valuable MRI features as well as clinical variables to identify H3 K27 M mutation.Lastly,we explored the feasibility of automatic machine learning(Auto-ML)models based on radiomics features extracted from conventional MRI to noninvasively predict the status of H3 K27 M mutation in diffuse midline gliomas.Materials and Methods:In the study to evaluate the characteristics of VASARI,we retrospectively collected patients with diffuse midline gliomas who were admitted in our hospital from December 2016 to August 2020 and underwent gene sequencing to identify H3 K27 M mutation.All patients had preoperative magnetic resonance imaging(MRI),including T1-weighted imaging(T1WI),T2-weighted imaging(T2WI),fluid attenuated inversion recovery(FLAIR)and contrast-enhanced T1WI(T1WI-CE),and some subjects also had diffusion-weighted imaging(DWI)data.At the same time,the clinical data such as the preoperative Karnofsky functional status(KPS)score and the course of disease were also collected.Two radiologists reviewed the data and collected the preoperative VASARI features including 27 items.The study evaluated the stability of VASARI features and used these features to predict the status of H3 K27 M mutation.In the study of the quantitative characteristics of multimodal MRI,the MRI sequences included conventional MRI sequences,DWI,diffusion tensor imaging(DTI),perfusion weighted imaging(PWI,dynamic susceptibility contrast imaging was employed in this study)and magnetic resonance spectroscopy(MRS).We retrospectively collected patients with diffuse midline gliomas who were admitted in our hospital from December 2016 to May 2020.All patients underwent the beforementioned MRI and H3 K27 M related gene sequencing.The clinical characteristics and KPS score at the time of diagnosis were also collected.The basic region of interest(ROI)of the tumor was delineated based on FLAIR images,and the ROI was used to extract the quantitative MRI features of the whole tumor.The FMRIB software library tool(FSL)was used to register the basic ROI to other image series,and then collected the quantitative parameters and histogram parameters of the MR features were collected.LCModel was used to calculate the absolute and relative concentrations of metabolites based on MRS data.Finally,the most valuable clinical and MRI features were screened out by LASSO regression,and a prediction model of H3 K27 M mutation was established.In the research based on radiomics and Auto-ML,we used Py Radiomics to extract radiomics features from FLAIR images,and then used the Auto-ML to predict the mutation status of H3 K27 M.Before Auto-ML analysis,the entire data set were randomly divided into 75% training set and 25% testing set.We applied the Tree-based Process Optimization Tool(TPOT)to automatically optimize machine learning parameters and the screening of important imaging features.In the performance evaluation,the area under the curve(AUC)and the average accuracy were used to compare the performance of the 10 independent TPOT models which were generated based on training and testing sets,so as to obtain the optimal model.In addition,we collected additional 22 subjects(mutant group VS wild-type group = 10 VS 12)to form an independent set to verify the effectiveness of the optimal model.In this study,the conventional statistical description and analysis,the establishment of regression model and the calculation of evaluation index were based on the open-source R language platform,and the feature extraction of radiomics features and the establishment and comparison of Auto-ML models were based on the Python language platform.Results:In the evaluation of VASARI characteristics,there were 60 subjects with H3K27 M mutation(the mutant group)with a mean age 33.96 years and 80 subjects without H3 K27 M mutation(the wild-type group)with an average age of 23.58 years.The age of onset in the H3 K27 M mutation group was younger than that in the wildtype group,and the difference was statistically significant(P=0.001).The functional status of these subjects in both groups decreased to varying degrees(KPS:mutant group VS wild-type group =76.47 VS 79.74,P=0.029),and the average course of disease in the mutant group was slightly shorter than that in the wild-type group(mean:2.59 VS 4.95 months,P=0.05).Dizziness with or without vomiting was more common in the mutation group,while headache was more common in the wild-type group.VASARI feature analysis showed that the consistency coefficients for the vast majority(88%)of features were high,all of which were greater than 0.8,and the average consistency coefficient of all features was 0.899.Most of the tumors in both groups demonstrated significant enhancement(63% in the mutation group and 61% in the wild-type group),and the proportion of enhancement in the mutation group was slightly smaller(nearly 60% of enhancement were less than 1/3 of the tumors).The mutation group had a slightly shorter maximum diameter at the largest cross-sectional image,while the wild-type group demonstrated more edema and invasion of functional areas.Univariate analysis showed that age,KPS score,F13(definition of the nonenhancing margin)and F21(deep white matter invasion)were of great value in predicting H3 K27 M mutation(e.g.,F13-AUC= 0.704,F21-AUC=0.739;and P <0.05).A total of 55 patients with diffuse midline gliomas were included in the study of multimodal MR quantitative characteristics(mutant group VS wild-type group =23VS 32).Among the adult patients,the average age of the mutant group was younger than the wild-type group(37.41 VS 48.96 years,P=0.005),the average KPS score was lower and the course of disease was shorter,but these differences did not reach significance.Among the children and adolescent patients there were no significant difference of age,KPS score or course of disease.Survival analysis for patients with available data demonstrated that the mean overall survival of the mutant group was shorter than that of the wild-type group(5 vs 19 months)but the difference of OS was near to the significance(P=0.067).Multimodal MRI quantitative feature analysis showed that most of the histogram parameters of relative apparent diffusion coefficient(r ADC)in the mutation group were lower than those in the wild-type group for both adult and children & adolescent patients.MRS analysis showed that the Ins/t Cr values of mutant tumors were also lower than those of wild-type tumors(P<0.05).The prediction model of H3 K27 M mutation status showed that age,Ins/t Cr and r ADC＿15th were the most valuable and effective predictors of H3 K27 M mutation status,and the AUC of the model was 0.898 and the score of F1 was 0.732.In the Auto-ML study,we retrospectively included 100 patients with diffuse midline gliomas,including 40 patients with H3 K27 M mutation(mean 23.60 years)and 60 wild-type patients(mean 31.57 years).We used TPOT to establish 10 prediction models.The prediction accuracy of the training queue was between 0.788 and 0.867,and it was between 0.6 and 0.84 in the testing queue.The AUC of the final Auto-ML model(model 7)for the testing queue was 0.903,and the accuracy was 0.911.In the independent verification set,the AUC of model 7 was 0.85,and accuracy was0.855.Conclusion:Diffuse midline gliomas with H3 K27 M mutation tends to affect children and young adults.This study shows that the difference between H3 K27 M mutant and wild-type midline gliomas could be identified and the status of H3 K27 M mutation would be non-invasively predicted by VASARI features,multimodal MRI features or radiomics plus Auto-ML.Among them,the Auto-ML method based on radiomics shows more potential because of its high prediction accuracy,the availability of data,and the robustness against observer’s bias.