Prediction Of Molecular Subgroups Of Adult-Type Diffuse Gliomas And Pediatric Low-Grade Gliomas By Radiomics Models

The fist part Radiomics Features from Diffusion Tensor Imaging Predict Molecular Subgroups of Adult-type Diffuse GliomasBackgroundGliomas are one of the most common intracranial tumor lesions,representing approximately 30%of all Central Nervous System(CNS)tumors and 80%of all malignant intracranial brain tumors.Compared with other intracranial space-occupying lesions,such as such as meningioma,acoustic neurinoma and ependymoma,the primary brain glioma is highly invasive and destructive,so that its treatment is the most complex and has a poor prognosis.Glioblastoma(grade WHOⅣ)is the most malignant,with a median survival of only about 15 months.Recent studies on genetic molecular markers in glioma have shown that genetic molecular information in primary gliomas is closely related to the overall survival time and prognostic status of patients.The fourth edition of the World Health Organization(WHO)Central nervous System Tumor Classification(CNS),revised in 2016,incorporated molecular diagnosis into traditional diagnostic criteria,most notably the isocitrate dehydrogenase(IDH)mutation and the deletion status on chromosome 1p/19q.Patients with IDH mutant glioma have relatively slow progression and longer survival;patients with IDH wild-type tumor have rapid progress,easy to relapse after surgical resection and worse prognosis.IDH mutant gliomas with 1p/19q codeletions have a better prognosis than non-codeletion gliomas and are more sensitive to adjuvant chemotherapy.Therefore,the accurate preoperative prediction of glioma isocitrate dehydrogenase mutations and chromosome lp/19q codeletion status can provide important guidance for risk-benefit assessment and individualized treatment decisions.However,the determination of molecular marker information on tumor genetics today requires tissue biopsy or surgical resection to obtain tissue specimens,followed by gene sequencing or immunohistochemical testing of tumor tissue.These invasive means cannot provide preoperative information on tumor molecular markers,so it is particularly important to find a non-invasive,preoperative and highly accurate method to predict the genetic molecular markers of tumors.Magnetic resonance imaging(MRI)plays an important role in the diagnosis of glioma and postoperative recurrence screening.In recent years,many researchers have used nuclear magnetic resonance imaging conventional sequences(T1,T2,TIC,FLAIR,ADC,etc.)to establish radiomics models to deeply explore the classification,genotyping and prognosis of glioma,and have made remarkable achievements.Given the limitations of conventional sequences,diffusion tensor imaging(DTI)that can noninvasively map the white matter nerve fiber pathway in vivo has gradually developed.The goal of this study is to use diffusion tensor imaging to extract imaging omics features to predict molecular subsets of adult-type diffuse glioma,and thus provide guiding information on glioma treatment to improve patient prognosis and improve survival.MethodsDiffusion tensor imaging data and clinical information of 510 patients with primary adult diffuse glioma confirmed by histopathology after surgical resection at the First Affiliated Hospital of Zhengzhou University from 2013 to 2019 were retrospectively analyzed.After enrollment screening,a total of 291 adult patients with diffuse glioma were included in this study.They were divided into three molecular subgroups:IDH mutant 1p/19q common deletion type(IDHmut-Codel,73 Cases),IDH mutation 1p/19q non-common deletion type(IDHmut-NonCodel,70 Cases)and IDH wild type(IDHwt,148 Cases).We randomly divided these cases into training and test sets in a 6:4 ratio,including 169 and 122 patients,respectively.We extracted imaging omics features from the DTI of each patient,then based on the data of the training set,removed redundant features,used the Boruta algorithm to select the features most related to the molecular subgroups classification to build a random forest prediction model,and finally evaluated in the test set.Results:We extracted 4788 imaging omics features from four diffusion maps(mean diffusion coefficient,fractional anisotropy,axradialial diffusion coefficient,anddiffusion coefficient).After removing redundant features and feature screening,we selected 19 imaging omics features and established a predictive model.In the test set,our model predicted IDHmut-Codel with 84.6%accuracy(area under the subject operating characteristic curve[AUC]=0.756),80.3%(AUC=0.879)for IDHmut-NonCodel,and 90.2%for IDHwt(AUC=0.950).Conclusion:The radiomics model based on non-redundant features extracted from DTI can accurately predict molecular subgroups(IDHmut-Codel,IDHmut-NonCodel,and IDHwt)of adult diffuse gliomas.Machine learning algorithms combined with magnetic resonance imaging provide the possibility for the preoperative non-invasive prediction of genetic molecular markers in adult-type diffuse gliomas.The second part Radiomics Features from Multiparametric Magnetic Resonance Imaging Predict Genetic Molecular Markers of Pediatric Low-grade GliomasBackgroundPediatric low-grade gliomas(PLGG)is the most common brain tumor in children,accounting for more than 30%of central nervous system tumors in children.In clinical practice,we found that PLGG generally has a relatively static biological behavior and a good prognosis,with a 5-year overall survival rate of approximately 95%.However,due to the infiltration of surrounding normal brain tissue,about 30%of patients with PLGG may experience multiple progression.Recently,considerable progress has been made in investigating molecular alterations in pediatric low-grade gliomas.The RAS/MAPK pathway is altered in most PLGGs,where BRAF is an important component,and a KIAA1549-BRAF fusion exists in approximately 32%of patients.In addition,many studies have used genetic molecular markers to classify PLGGs into low-risk group(KIAA1549-BRAF fusion or MYB amplification),intermediate-Ⅰ-risk group(BRAFV600E and/or CDKN2A deletion),intermediate-Ⅱ-risk group(no biomarkers)and high-risk group(TERTp or H3F3A mutation or ATRX deletion),with different progress free survival and overall survival.Compared with the intermediate/high-risk group,PLGG had good survival before the follow-up time point in the low-risk group.Non-invasive prediction of the low-and intermediate/high-risk groups of genetic molecular markers or gliomas remains a challenge in the treatment of PLGG.Magnetic resonance imaging(MRI),as a noninvasive method plays an important role in the diagnosis of glioma,but there are still various problems in its application.Recent advances in artificial intelligence(AI)algorithms have greatly promoted the automatic quantification of magnetic resonance imaging.Many studies have used the AI algorithm to investigate the relationship between imaging features and the molecular pathology of glioma,but there is still a lack of research on low-grade gliomas in children.The goal of this study is to extract radiomic characteristics using multiparametric magnetic resonance imaging(including T1 weighted image,Contrast-enhanced T1 weighted image,T2 weighted image,fluid attenuated inversion recovery image and apparent diffusion coefficient image)of 61 PLGG patients to construct a model of predictive molecular subgroups(low-risk and intermediate/high-risk groups)and genetic molecular markers(KIAA1549-BRAF fusion),so as to provide more reference value for PLGG patients to improve patient prognosis and improve survival.MethodsIn this retrospective study,102 primary pediatric low-grade glioma patients(age<18)confirmed from the First Affiliated Hospital of Zhengzhou University from January 2011 to December 2016 were selected,and 61 pediatric low-grade glioma patients were included after enrollment screening.We divided the patients into low-risk and intermediate/high-risk groups,BRAF fusion-positive groups,and BRAF fusion-negative groups.We extracted imaging omics features from multi-parameter MRI(including T1,T1C,T2,FLAIR and ADC),standardized processing with training set data,removed redundant features and Boruta package features,selected the most relevant features with genotyping to establish a prediction model,and finally used the test set data to evaluate the predictive efficiency.ResultsWe extracted 5985 imaging omics features from five multiparameter MR imaging sequences for each patient and constructed predictive models for predicted low-risk and intermediate/high-risk,BRAF fusion positive and BRAF fusion negative groups using training set data.We constructed the low-risk and intermediate/high-risk PLGGs classification models using the five screened imaging radiomics features with an AUC of 0.850,accuracy of 81.0%,sensitivity of 80.0%and specificity of 81.3%in the test set.The KIAA1549-BRAF fusion prediction model constructed using five imaging radiomics features had an AUC of 0.819,75.0%accuracy,sensitivity of 58.3%and specificity of 91.7%.ConclusionThis study is the first to demonstrate that MRI imaging omics is able to predict molecular subsets of pediatric low-grade gliomas with excellent sensitivity and specificity.The radiomics model established by machine learning algorithms provides the possibility to predict preoperative genetic molecular markers in pediatric low-grade gliomas with multiparametric magnetic resonance imaging.