Artificial Intelligent Based Diagnosis For Pediatric Posterior Fossa Tumor Using Magnetic Resonance Imaging

Children’s brain tumors account for about 25% of children’s cancers,which is the second most common cancer after children’s leukemia,and posterior fossa brain tumors account for60% of children’s brain tumors.Among the posterior fossa brain tumors,the three most common tumors are medulloblastoma,astrocytoma and ependymoma.Due to the different features and malignancy of these three tumors,the therapeutic schedule and prognosis results are also different.Therefore,in order to determine the therapeutic schedule for children more quickly and accurately,the significance of accurate early diagnosis is particularly obvious.Magnetic resonance imaging(MRI)has been widely used in children brain tumor because of its advantages such as no radiation,high resolution,high tissue contrast and multi-modality imaging.Also,with the development of computer technology,decision support based on MRI becomes possible.Driven by big data,the decision support system can not only reduce the work of doctor effectively and improve the accuracy of diagnosis,but also extract more information that cannot be discovered by doctors’ subjective clinical experience.It has important significance in clinical work and scientific research.This thesis used multi-modal MRI data to intelligently diagnose three common posterior cranial fossa brain tumors in children.Deep learning technology and radiomics were mainly used to achieve automatic segmentation and classification of brain tumor.The survival analysis of children with medulloblastoma was carried out using characteristics of the model.The work of this thesis mainly includes the following four aspects:(1)In terms of clinical multi-modal MRI data,a complete set of data preprocessing workflow is proposed.The preprocessed data has better consistency in image size,field of view,contrast and intensity uniformity.(2)The Accurate segmentation of brain tumor tissue is achieved based on multi-modal MRI using deep learning technology.With the improvement of the U-Net network structure:batch standardization layers and residual modules are added,and the accuracy of the final segmentation reached 83.88%.(3)Three types of children’s posterior fossa brain tumors were automatically classified using deep learning method and radiomics methods,respectively.In terms of deep learning approaching,the method of transfer learning is further used,and brain tumor image slices are used to fine-tune the pre-trained deep learning model,and a final classification accuracy rate of 83.33% is obtained.In radiomics approaching,feature extraction is performed on the tumor area,including histogram features,shape features and texture features.traditional machine learning SVM classifier and random forest classifier are used respectively to achieve the classification of brain tumors.Among them,the random forest classifier has better performance,with an overall accuracy rate of 91.67%,and 100% accuracy and recall rate on astrocytoma.(4)Combined with deep learning and radiomics,survival analysis was performed on children with medulloblastoma.A large number of deep features which were extracted by the deep learning model combining with the radiomics features were selected by a variety of feature selection methods.A Cox regression model was fitted by the selected features and finally the risk index,which was calculated by Cox regression model,successfully divided the children into The two groups of high and low risk,and the prediction of survival time is achieved.