Multimodal MRI Radiomics Intelligent Diagnosis Of Pediatric Brain Tumors And Its Application

Central nervous system tumors are the second highest incidence tumors in children after leukemia,and they are also the important cause of death in children.Posterior fossa tumors account for about 55%-70%of brain tumors in children.Among them,ependymoma(EP),medulloblastoma(MB),and pilocytic astrocytoma(PA)are the most common pediatric posterior fossa tumors.As these three kinds of tumors are different in malignant,treatment,and prognosis.Accurate and non-invasive identification of them before operation is helpful to the formulation of personalized surgical strategies and timely communication with patients and their families.Magnetic resonance imaging(MRI)has the advantages of high resolution,multi-sequence,and non-radiation,so it is often used in clinical evaluation of brain tumors.Radiomics focuses on the potential gray level variation of images which are invisible by human eyes,extracts the effective texture information from images,and quantifies them.Machine learning technology can be used to establish an accurate and reliable diagnostic model and save the cost of human resources.The combination of multimodal MRI,radiomics,and machine learning could predict tumor type,grade,survival period,curative effect and so on,assist doctors in clinical decision-making.Due to the particularity of the pediatric population,it is difficult to collect cases on a large scale.At present,similar studies are few in this research field.Therefore,it is essential to establish a research system for pediatric brain tumors.Based on radiomics and machine learning techniques,this paper aimed at the value of multiregional and multimodal MRI in preoperative diagnosis of pediatric posterior fossa tumors.We have carried out the following research on methods and applications:1.Established a relatively complete radiomics data processing scheme.For the lack of a unified and effective radiomics processing scheme,and limited MRI sequences and tumor regions in the current pediatric brain tumors research.In this paper,bias field was corrected,image intensity across scanners was normalized,multimodal MRI was registered,three-dimensional volumes of interest were semi-automatically segmented,and the images were further normalized in order to enhance the inter-class difference.Finally,11958 radiomics features were extracted from multiregional and multimodal MRI.These studies laid a foundation for the establishment and application of subsequent radiomics models.2.Conducted a radiomics model applicable for differentiating pediatric posterior fossa tumors,and compared its performance with that of radiologists.Existing diagnostic models are usually effective in solving one problem area,but are difficult to be applied to other domains.According to the characteristics of pediatric posterior fossa tumors,it is urgent to establish a model that is most likely to be suitable for the data in the research field.In this paper,56 combinatorial models were constructed by using 7 feature selection methods and 8 classifiers.According to the comparison of diagnostic performance,the models suitable for EP,MB,and PA were explored.The results showed that the multiregional and multiparametric MRI radiomics model constructed by fast correlation-based feature selection(FCBF)and support vector machine(SVM)achieved an optimal overall performance(area under the curve[AUC]=0.934),which was outperformed those of radiologists(AUC:0.681-0.904),and better than most related literature,illustrating that the model proposed has clinical value in differentiation of pediatric posterior fossa tumors.3.Investigated the contribution of different MRI sequences and tumor sub-regions to preoperative diagnosis in pediatric posterior fossa tumors,discussed the added value of contrast-enhanced T1 weighted(T1C)imaging.Exploring the diagnostic contribution of specific MRI sequences plays an important role in guiding clinicians to reasonably adopt different examination methods under existing conditions,avoiding unnecessary injuries,and saving medical costs.TIC imaging requires injection of gadolinium-based contrast agents,these drugs potentially affect the growth and development of children.In this paper,radiomics models were applied to explore the contribution of different MRI sequences to discriminate pediatric posterior fossa tumors.The results showed that T1C was not of irreplaceable diagnostic value.The combination of plain scan sequence and diffusion-weighted imaging(DWI)were the preferred acquisition sequences for preoperative diagnosis of pediatric brain tumors.In addition,when edema features were added to contrast-enhancing tumor volume,the performance did not significantly improve.4.Discussed cross-scanners generalization of radiomics features,identified and analyzed image-based biomarkers affecting the classification of pediatric posterior fossa tumors.Transferability of radiomics features across different scanners determines whether this technique can be popularized in clinic.In this paper,the data on six scanners were divided into four groups,and the relevant features extracted from any three groups were tested on the remaining one,so as to investigate the transferability of features across scanners.In addition,on the framework of image-based diagnosis,image biomarkers provide comprehensive information to characterize tumors.In this study,we identified image biomarkers that distinguish EPs,MBs,and PAs,attempted to explain their potential pathological significance.In conclusion,this paper aimed at the method improvement and application investigation towards the existing limitations in the preoperative automatic diagnosis of pediatric posterior fossa tumors.The results of this study have practical clinical value for assisting clinicians in diagnosis decision-making,rationally adopting examination strategies,and avoiding harm to children.