Radiomics Features From Magnetic Resonance Imaging And Clinical Features Predict Glioma Groups Based On A Machine-Learning Approach

BackgroundGliomas are the most common malignant tumors of the brain,mostly occurring in the cerebral hemispheres,and can be morphologically classified into hairy cell-shaped astrocytomas,astrocytomas,oligodendrogliomas,oligodendro-astrocytomas,and glioblastomas.The World Health Organization classifies gliomas into grades Ⅰ,Ⅱ,Ⅲ,and Ⅳ based on the morphology and malignancy of the tumor.grade Ⅰ(hairy cell astrocytoma)is clinically and pathologically distinct from other gliomas.grade Ⅱ,Ⅲ,and Ⅳ gliomas are called diffuse gliomas.grade Ⅱ and Ⅲ gliomas are also called lower-grade gliomas(LGGs),and grade Ⅳ gliomas are called glioblastomas(GBW).Gliomas,LGGs),and grade Ⅳ gliomas are called glioblastomas(GBW).Due to the aggressive and infiltrative nature of tumor growth,even if the tumor is completely removed microscopically,the residual tumor tissue after surgery will still recur after a period of time or progress to higher grade tumors.Surgery combined with radiotherapy can extend the life span of patients to some extent,but some gliomas are sensitive to radiotherapy,while others do not respond well to radiotherapy,and the prolonged survival of patients varies,with the OS of WHO grade Ⅱ,Ⅲ and Ⅳ astrocytomas being approximately 6-8 years,2 years and 15 months,respectively.Therefore,a more accurate staging of gliomas is of great importance in predicting the prognosis of gliomas.Numerous studies have recently advanced the staging of gliomas to the molecular genetic level.Molecular pathology has become increasingly important for the staging of gliomas.Of these molecular alterations,three are particularly noteworthy because they occur early in glioma formation,are prevalent in gliomas,or are strongly associated with overall survival.The first identified alteration is the co-deletion of chromosomal arms 1p and 19q(1p/19q co-deletion),which is associated with the histological type of oligodendrocytes and sensitivity to alkylating agent chemotherapy.The second is IDH1 or IDH2(these genes are very similar to each other and are hereafter collectively referred to as IDH),which is not restricted to a specific histopathological type of glioma,but is associated with a distinct tumor cell metabolism.The third is the promoter of TERT,which encodes telomerase.Curiously,mutations in the TERT promoter,which leads to enhanced telomerase activity and telomere lengthening,can be seen in both the most aggressive human glioma(grade Ⅳastrocytoma)and the least aggressive diffuse glioma(grade Ⅱ oligodendroglioma).This suggests that telomere maintenance may be a necessary prerequisite for brain cancer formation.2016 WHO classification of CNS tumors identifies the importance of this 1p/19q co-deletion and IDH mutation.2021 WHO classification of CNS tumors includes TERT promoter mutations in the diagnostic criteria for glioblastoma.Related studies have shown that glioma staging based on 1p/19q co-deletions,IDH mutations and TERT promoter mutations can more specifically differentiate gliomas in terms of prognosis,acquired somatic changes and germline variants.However,at present,obtaining this molecular pathology information is mainly achieved by immunohistochemistry and genetic testing after obtaining tissue specimens through surgery or biopsy.These procedures are usually invasive and carry a high risk(e.g.,postoperative infection,seizures,etc.),so the need to develop a noninvasive test to predict patient glioma staging is increasingly evident.Recent studies have shown that imaging histology has made great progress in predicting glioma staging and has become a frontier and hot topic in current research.In this study,we developed a machine learning model based on radiomics and clinical features to predict the molecular subtypes of diffuse gliomas,aiming to make new discoveries and advances in the prediction of glioma staging.ObjectivesTo investigate the prognostic value of MRI features in LGG and the classification of LGG combined with molecular pathology.MethodsThe clinical and MRI data of 540 patients with diffuse glioma who underwent surgery at the First Affiliated Hospital of Zhengzhou University from June 2011 to June 2021 and met the criteria were retrospectively analyzed.Diffuse gliomas were classified into 5 subtypes according to IDH,1p19q and TERT:IDH and TERT mutations,IDH mutations only,TERT mutations only,triple-negative and triplepositive.Patients were randomly assigned to the training cohort(n=408)and the validation cohort(n=132)in a 3:1 ratio.The PyRadimoc package based on Python was used to extract radiomic features from multiparametric MRI(T1,T1c,T2,FLAIR,ADC).Meaningful radiomic features were filtered using the Caret package and Boruta package in R language,and then combined with clinical features to construct a random forest-based machine learning model.ROC curves were used to calculate the predictive performance of all individuals and models.ResultsA total of 5985 radiomic features were extracted from each patient.After screening,26 radiomic features were selected.In the training cohort,the AUC values of our model were 0.835(triple positive),0.527(TERT and IDH mutations),0.810(IDH mutations only),0.661(triple negative),and 0.808(TERT promoter mutations only).In the validation cohort,0.802(triple positive),0.857(TERT and IDH mutations),0.800(IDH mutations only),0.621(triple negative),0.819(TERT promoter mutations only).ConclusionsOur model showed good predictive performance for three groups of gliomas with triple positive,IDH mutation only and TERT promoter mutation only.Our study provides a new method for predicting the staging and prognosis of gliomas.Thus,we obtained that magnetic resonance imaging features are an important indicator for determining the staging of patients with diffuse glioma,and,combined with current molecular pathology for molecular staging of diffuse glioma,can more finely guide the classification and prognosis prediction of diffuse glioma for the purpose of precision and individualized medicine.