Application Of Ivim Based Multi-modal Magnetic Resonance Imaging Technology For Pediatric Tumor Diagnosis

Brain tumors are the most common solid tumor among children,which has a high fatality rate and account for 15.5% of brain tumors in the whole age group.Early detection and determination of the children’s brain tumors’ growth location,shape,size,and grade can greatly improve the survival rate,as well as the prognosis of children.Magnetic resonance imaging technology(MRI)uses non-invasive and non-radiation methods to perform high-resolution imaging of children’s brain tumor tissues,and it can reflect the structure and perfusion of children’s brain tumors.Therefore,MRI has become an important imaging basis for the diagnosis of children’s brain tumors.With the continuous development of magnetic resonance technology,functional sequences that can reflect the level of tissue diffusion and perfusion appear in clinical magnetic resonance scans,such as Intravoxel Incoherent Motion(IVIM),Diffusion Weighted Imaging(DWI),and Dynamic Susceptibility Contrast(DSC)sequences.In particular,IVIM sequence can distinguish tissue diffusion and perfusion effects,thereby making up for the shortcomings of the traditional diffuse perfusion sequence.However,there are currently few studies on IVIM in China,so it is rarely used in clinical practice.In recent years,many studies have used the combined analysis of multi-modal MR images,which greatly exert the clinical value of magnetic resonance imaging in the diagnosis of brain tumors.In the past,doctors needed to manually analysis MR images layer by layer,so the diagnosis process was extremely time-consuming and prone to diagnosis deviations caused by human factors.However,with the widespread application of machine learning and deep learning in the field of medical image analysis,Computer-Aided Diagnosis(CAD)methods have made great breakthroughs in the diagnosis of brain tumors,which can solve the problems of manual analysis.This study collected 60 cases of children’s brain tumors with multimodal magnetic resonance data,including IVIM,T1,DWI,and DSC sequences.A set of data preprocessing schemes were first designed,and then the preprocessed multi-modal MR images were used to train brain tumor segmentation models and benign and malignant classification models using machine learning and deep learning methods.Finally,the visualization of the convolutional neural network classification model was carried out.The application value of IVIM sequence in the segmentation of children’s brain tumors and the classification of benign and malignant was explored.The results showed that IVIM performed better in the study of benign and malignant classification than brain tumor segmentation,which had important clinical value.Secondly,this study also used transfer learning method to train the benign and malignant brain tumors classifier by exploring feature extraction transfer learning and fine-tuning transfer learning respectively.The VGG19 and Alex Net pre-training models were selected as the feature extractors,and the depth features of the convolutional layer and the fully connected layer were extracted,then the classifier training was performed.At the same time,this experiment compared the effects of 11 modal combinations,3 input clipping methods,and 4 fine-tuning strategies on the performance of the classification model.The classification model trained by the sequence combination containing IVIM had high classification accuracy.This experiment also proposed two custom shallow convolutional neural networks,Patch Net and Slice Net,to classify benign and malignant brain tumors.Among them,Patch Net achieved better classification performance than Slice Net.Finally,this study conducted a deep neural network visualization study on the classification model obtained from the fine-tuned transfer learning.The network was visualized by the convolution kernel,the forward feature map,the backward feature map and the Class Activation Mapping.Those visualization showed the working mechanism of the network,thereby increasing the credibility of the classification model.This study based on the IVIM multi-modal magnetic resonance imaging,explored the brain tumors segmentation and classification of benign and malignant.The application value and prospect of IVIM sequence in the diagnosis of brain tumors were confirmed.At the same time,a multi-modal MR solution in the intelligent assisted diagnosis of children’s brain tumors was proposed and good results had been obtained.