Machine Learning-Based Radiomics And Qualitative Features From Multimodal Magnetic Resonance Imaging Predict Molecular Subtypes Of Lower-Grade Glioma

BackgroundDiffuse lower-grade glioma(LGG,World Health Organization[WHO]grade Ⅱand Ⅲ)is an invasive tumor that originates from glial cells or precursor cells,with significant genetic heterogeneity and genetic variability.Compared with WHO gradeⅣ glioblastoma(GBM),the course of LGG is relatively slow.Although there are standardized treatment methods,including neurosurgical tumor resection,radiotherapy and chemotherapy,tumor recurrence and malignant progression are inevitable due to its high invasiveness.The survival time of LGG patients ranged from 1 to 15 years,and there were great differences when stratified by tumor type.In the 2016 WHO classification of central nervous system tumors,molecular markers were included in the diagnosis of glioma for the first time.According to isocitrate dehydrogenase(IDH)mutation and 1p/19q codeletion status,diffuse LGG is divided into three molecular subtypes,namely(1)IDH wild-type astrocytoma,(2)IDH mutation without 1p/19q codeletion astrocytoma,and(3)IDH mutation with 1p/19q non-codeletion oligodendrocytoma.IDH wild-type LGG has the worst prognosis,while the LGG with IDH mutation and 1p/19q codeletion has the best prognosis and is more sensitive to radiotherapy and chemotherapy.At present,the diagnosis of LGG mainly depends on the results of preliminary imaging examination and surgical biopsy.Although biopsy is the gold standard for pathological diagnosis and histological grading of LGG,the accuracy of biopsy is affected by tumor heterogeneity and sampling site.In addition,as an invasive examination,biopsy may increase the risk of death.Therefore,we urgently need a new method for noninvasive diagnosis of LGG molecular pathology before operation.In recent years,as a new high-throughput quantitative imaging technology,radiomics can extract a large number of high-dimensional data from Computed Tomography(CT),Magnetic Resonance Imaging(MRI)and other image data,overcome the limitations of traditional evaluation of image features only by vision,provide a new non-invasive method for diagnosis and prediction of tumors,and can more comprehensively evaluate the characteristics of tumor areas,thereby providing more valuable information for clinical practice.Recent studies have shown that the imaging characteristics are closely related to the histological grade,molecular pathology and prognosis of glioma.However,based on multimodal MRI,there are few reports on the combination of traditional qualitative imaging features and radiomics features to predict LGG.The establishment of a new prediction model of LGG molecular subtypes has important clinical and social significance for the early diagnosis,individualized treatment and comprehensive management of LGG population.ObjectivesTo explore the value of machine learning model based on multimodal MRI radiomics,qualitative imaging and clinical characteristics of patients with lower-grade gliomas in preoperative diagnosis of molecular subtypes of lower-grade gliomas.MethodsA total of 335 patients with WHO grade Ⅱ/Ⅲ lower-grade gliomas were included in this study.The clinical and imaging data of patients were collected for molecular pathological detection(IDH mutation and 1p/19q codeletion).There were 94 cases of IDH wild-type(IDHwt),110 cases of IDH mutant with 1p/19q non-codeletion(IDHmut-noncodel),and 131 cases of IDH mutant with 1p/19q codeletion(IDHmutcodel).All patients were divided into training set(n=269)and test set(n=66)in a ratio of 4:1.A total of 5929 imaging features were extracted from preoperative multimodal MRI including T1WI,CE-T1WI,T2WI,FLAIR and DWI(ADC).Based on the robust,non-redundant and relevant radiomics features selected by boruta algorithm,a radiomics random forest prediction model is established.At the same time,a radiomics model with only conventional sequence and no ADC map was established.The preoperative MRI of all patients were scored according to the Visually Accessible Rembrandt Images(VASARI)annotations and T2-FLAIR mismatch sign.The features screened by boruta were selected to establish the qualitative imaging prediction model.In addition,a prediction model of clinical characteristics was established by using clinical characteristics(gender and age).A combined prediction model of three molecular subtypes of lower-grade gliomas was established by combining radiomics features,qualitative imaging features(VASARI features and T2-FLAIR mismatch sign)and clinical features.The receiver operating characteristic(ROC)curve was used to evaluate the performance of the prediction model,and Delong analysis was used to compare the differences between the models.ResultsThe 17-feature radiomics model achieved area under the curve(AUC)values of 0.8121,0.7384 and 0.7905 for IDHwt,IDHmut-noncodel,and IDHmut-codel,respectively in the training set,0.6557,0.6830 and 0.7579 respectively in the testing set.The qualitative imaging model based on 9 VASARI features and T2-FLAIR mismatch sign achieved AUC values of 0.7657,0.7938 and 0.7368 for IDHwt,IDHmut-noncodel,and IDHmut-codel,respectively in the training set,0.7488,0.7599 and 0.7892 respectively in the testing set.The clinical model achieved AUC values of 0.6229,0.5717 and 0.5028 respectively for IDHwt,IDHmut-noncodel,and IDHmutcodel,respectively in the training set,0.6551,0.4841 and 0.5873 respectively in the testing set.The radiomics model without ADC map achieved AUC values of 0.7729,0.6612 and 0.7573 respectively for IDHwt,IDHmut-noncodel,and IDHmut-codel,respectively in the training set,0.7234,0.5144 and 0.6533 respectively in the testing set.The combined model constructed by the combination of the selected radiomcis features,qualitative imaging features and clinical features achieved AUC values of 0.8414,0.8190 and 0.8193 respectively for IDHwt,IDHmut-noncodel,and IDHmutcodel,respectively in the training set,0.8623,0.8056 and 0.8036 respectively in the testing set.The performance of the radiomics model based on five sequences is better than the radiomics model without ADC map using only conventional sequences(Delong P<0.05),and the performance of the combined model is better than the individual radiomics model,qualitative imaging model and clinical model(Delong P<0.05).ConclusionsBased on machine learning,this study analyzed and integrated the radiomics features,qualitative imaging features and clinical features of multimodal MRI,and constructed a joint model with good predictive performance for the molecular subtypes of LGG.This model can provide a reference for preoperative diagnosis of LGG,and it has the potential as a non-invasive tool to replace invasive tissue sampling,so as to achieve the purpose of precision medicine and guide the individualized management of LGG patients.