The Application Of MRI Based Radiomics And Deep Learning In The Preoperative Diagnosis And Evaluation Of Glioma

Background and PurposeGlioma is the most common primary tumor of the central nervous system.Patients with gliomas usually have a poor prognosis and this brings a huge economic and social burden for our country.Even for cases with the same tumor histological type and WHO grade,as well as the same treatment,they may have a significantly different prognosis.Based on the histopathological features,the addition of molecular information helps gliomas to be classified more accurately.Including glioma,MRI is the basic examination of brain tumors.However,the MR imaging manifestations of glioma are diverse.Sometimes it is difficult to distinguish glioma from other brain tumors or diseases by MR imaging.Even gliomas with different types or WHO grades may present similar MR imaging manifestations,which make they are difficult to distinguish from each other.Also,traditional interpretation of conventional MR imaging has some difficulty in predicting molecular markers of glioma.Accurate preoperative diagnosis and evaluation,including the accurate differentiation of gliomas from other brain tumors,the differentiation of different histological types of gliomas,and evaluation of the molecular markers of glioma,is useful for individualized treatment and the therapeutic effects prediction.Radiomics and deep learning are good at mining imaging features and revealing the underlying pathophysiological information of tumors.Therefore,MR images based radiomics and deep learning,may be useful for the accurate diagnosis and evaluation of glioma before operation.For the differentiation of glioma from other tumors,it is usually very difficult to distinguish a supratentorial glioblastoma from a single brain metastasis,as both of them may present similar MRI findings.However,it is of great value to distinguish the two types of tumors,as the clinical decision is different for each other,especially for patients without previous malignancy.As previous studies had found that there are tumor cell infiltration and tumor angiogenesis in the peri-enhanced edema region for glioblastoma,but not for brain metastasis,so,the using of radiomic method to extract features from the peri-enhanced edema region on enhanced MR images may be very valuable to differentiate the two tumors.For high grade glioma of cerebellar,except for metastasis,hemangioblastoma is another tumor that needs to be differentiated due to its diverse manifestations.Glioblastoma is a kind of highly malignant tumor,and it accounts for about half of all gliomas.EGFR gene amplification and mutation are the most common genetic changes in glioblastoma.The EGFR gene amplification state has an important influence on the selection of treatment methods and the prognosis of glioblastoma.Since EGFR gene amplification is associated with tumor cell proliferation,invasion,angiogenesis and some other events.It may be possible to find more information about these events from enhanced MR images by radiomics.And thus,it may be possible to predict the EGFR gene amplification state of glioblastoma.The prognosis of oligodendroglioma is usually better than that of astrocytoma with the same WHO grade.So,the accurate diagnosis of oligodendroglioma has important clinical value for lower grade gliomas.The codeletion of chromosome arms lp/19q is a molecular marker of oligoglioma.It is useful for the diagnosis of oligodendroglioma,the differentiation from other gliomas and the evaluation of prognosis.Based on the excellent performance of deep learning got in previous image classification work,the using of deep learning method with convolutional neural network may be also satisfying for the prediction of codeletion of chromosome arms lp/19q of lower grade gliomas.Building models or classifiers using the extracted features is usually considered as the last step of radiomics workflow,but the performance of established models or classifiers may be not always satisfactory.How to further improve the performance of models or classifiers to meet the clinical needs is a question.Since there are many different algorithms,and the performance of each models or classifiers built based on the algorithms may be not all the same,whether the combined use of multiple models or classifiers,which is similar to the multi-disciplinary expert consultation mode in clinical setting,can got a better results is interesting and worth studying.Deep learning usually needs a large labeled training data,which is very difficult for medical scene.With transfer learning method based on deep neural network model trained by large-scale image data,satisfying results were obtained in many research,such as skin cancer classification,fracture diagnosis and so on.Whether this transfer learning method can get a similar performance for molecular markers prediction of glioma is unknown.At present,the extracted features or built models in radiomic studies are mainly specific to diseases or classification results.The features and models used in each study may have some differences.Also,radiomics are usually studied in common or frequently occurring diseases.Whether the radiomic features and models have generalizability between different diseases or between different classification results unknown,and how radiomics be used in rare or uncommon disease is worth studying.Last but not the least,neither radiomics nor deep learning is omnipotent.It requires the active participation of doctors in the using process.Therefore,in this study,we used radiomics or deep learning methods to differentiate single supratentorial brain metastasis from glioblastoma(Experiment 1),differentiate high grade cerebellar glioma from hemangioblastoma(Experiment 2),predict the EGFR gene amplification status of glioblastoma(Experiment 3),and predict the codeletion of chromosome arms lp/19q of lower grade glioma(Experiment 4).What's more,in the study,we tried to explore the value of the combined use of multiple classifiers(Experiment 1),the feasibility of using a borrowed radiomic model for rare disease(Experiment 2),and the use of deep neural network based transfer learning method for predicting the molecular markers of glioma(Experiment 4),and analyze the application and the value of subsequent analysis by doctors(experiment 1 and experiment 4).Materials and methodsExperiment 1:A total of 180 cases of supratentorial single brain tumor(brain metastasis,n=90;glioblastoma,n=90)were included.All of the tumors were pathologically identified,and the enhanced MR examination was performed preoperatively.The data were randomly divided into training data set and test data set according to the ratio of 8:2.Quantitative radiomic features from the peri-enhanced edema region were extracted.After feature selecting and processing,5 classifiers were built,and then,these classifiers were combined by using the same weight voting method and different weight logistic regression analysis method.The ability of 5 base classifiers and also the combined use of the classifiers to distinguish single supratentorial brain metastasis from glioblastoma were analyzed.Accuracy,sensitivity,specificity and Youden index were used to evaluate the performance of base classifiers and the combined use of the classifiers.Experiment 2:Thirty patients with pathologically identified cerebellar tumor(high grade glioma,n=15;hemangioblastoma,n=15)were included.Each patient had their preoperative MRI performed.The two kinds of tumors were differentiated using a ready-made classification model(borrowed model)that built for differentiating glioblastoma from pilocytic astrocytoma.Meanwhile,according to the similar enhancement mechanism to high grade glioma,15 single metastasis of cerebellar with their preoperative MR images were also included for validation,and the same model was used to differentiate metastasis from hemangioblastoma.The label of glioblastoma in the model was replaced by high grade glioma or metastasis,and the label of pilocytic astrocytoma was replaced by hangioblastoma.Accuracy,sensitivity and specificity were used to evaluate the ability of the borrowed model to distinguish high grade cerebellar glioma from hemangioblastoma.Experiment 3:Fifty glioblastoma(GBM)patients with epidermal growth factor receptor(EGFR)gene amplification status detected were included.Conventional MRI,mainly including T1WI,T1WI and contrast enhanced T1WI,were performed preoperatively.The patients were divided into the training data set and the test data set with a ratio of about 7:3.High throughput radiomic features of the tumor,including enhanced part and necrotic part,on enhanced MR images were extracted by semi-automatic software.According to the feature stability(evaluated by intraclass correlation efficient(ICC)value)and Lasso regression algorithm,feature selection were carried out in the training data set data.The selected features were then used to build three models,including the logistic regression model,support vector machine model and neural network model.The value of area under the ROC curve(AUC)in both training data set and test data set were used to evaluate the performance of the models.Experiment 4:One hundred and forty cases with lower-grade glioma,and codeletion of chromosome arms lp/19q were detected by molecular pathology,in our center were included.All of the patients had their conventional MRI performed preoperatively,with the sequences mainly including T1WI,T1WI and contrast enhanced T1 WI.Seven hundred and twelve T2W images of the tumor were cropped and randomly divided into training data set,validation data set and test data set with a ratio of approximately 8:1:1.The top layer of GoogleNet Inception v3 model was retrained using images from the training data set,and data augmentation method were used during training the model.The number of iterations was 5000,and the initial learning rate was 0.01.Another one hundred and fifty-nine lower-grade glioma cases from TCIA public dataset were used as external test data set.In order to further improve the prediction accuracy,the tumor location features of cases that were misdiagnosed in the test data set were retrospectively analyzed.The common valuable features got by doctors were validated by cases in TCIA external test data set.ResultsExperiment 1:A total of 271 features were extracted from each tumor,and 8 features were selected after feature selection.In the training data set,the accuracy,sensitivity,specificity and Youden index of the five classifiers were 0.67?0.80,0.60?0.82,0.63?0.86 and 0.34?0.60 respectively.In the test data set,the accuracy,sensitivity,specificity and Youden index of the five classifiers were 0.61-0.69,0.39?0.72,0.50?0.83,and 0.22?0.39.When using the same weight voting method for the five classifiers,the accuracy,sensitivity,specificity and Youden index in the training set were 0.73,0.76,0.69 and 0.45 respectively.And in the test data set,they were 0.61,0.56,0.67 and 0.23 respectively.For cases,when the five classifiers reached agreement,the accuracy,sensitivity,specificity and Youden index in the training data set were 0.86,0.80,0.91 and 0.71 respectively,and the corresponding accuracy,sensitivity,specificity and Youden index in the test set were 0.77,0.75,0.78 and 0.53 respectively.When using different weights,the combined performance of the five classifiers was not outstanding.Experiment 2:The accuracy,sensitivity and specificity of the borrowed model were 0.73,1.0 and 0.47 respectively for differentiating high grade glioma from hemangioblastoma.As a validation,the same results were obtained for the differentiation of metastasis from hemangioblastoma using the borrowed model.Experiment 3:A total of 102 radiomic features were extracted from each tumor on enhanced T1W images.After feature selection,three features were chosen,including a shape-based feature original_shape_SurfaceVolumeRatio,a Gray Level Cooccurence Matrix(GLCM)feature original glcm_MaximumProbability and a Gray Level Dependece Matrix(GLDM)feature original_gldm_LargeDependenceHighGrayLevel Emphasis.The AUC values of logistic regression model,support vector machine model and neural network model in the training data set were 0.85,0.84 and 0.84 respectively,and those in the test data set were 0.84,0.84 and 0.82 respectively.Experiment 4:Using the deep convolutional neural network,the accuracy in training data set,validation data set,test data and TCIA external test data set was 0.94,0.68,0.77 and 0.73,respectively.Among the 14 cases(15 images)that had been wrongly predicted in the test data,3 cases with the tumors mainly located in the subcortical region,all of these cases had no codeletion of chromosome arms 1p/19q but were misdiagnosed as codeletion of chromosome arms 1p/19q.And similar,5 cases with tumors mainly located in the subcortical region among the 14 cases that had beenmisdiagnosed as codeletion of chromosome arms 1p/19q in the TCIA external test data set.Combined with tumor location analysis,the accuracy of test data set and TCIA external test data set reached 0.81 and 0.75.ConclusionsIn summary,our study showed that:First,eight features derived from the peri-enhancing edema region had moderate value in differentiating supratentorial solitary brain MET from GBM with multiple classifiers.Combined use of the classifiers could bring an extra benefit for increasing the classification performance of cases,especially when all classifiers reached an agreement.Second,it is a feasible way to differentiate high grade glioma from hemangioblastoma by using the borrowed GBM-PA model.Third,it is valuable to predict the EGFR gene amplification state of glioblastoma by using three features deriving from tumor on the contrast enhanced MR images.Fourth,the use of GoogleNet inception v3 deep convolutional neural network has a moderate value in predicting the chromosome lp/19q combined deletion status of lower-grade glioma.The retrospective analysis by doctors is helpful to improve the prediction accuracy.