Classification And Segmentation Of Pediatric Demyelinating Diseases Based On MRI

Central Nervous System Demyelinating Diseases are a group of brain and spinal cord myelin destruction or demyelination as the main feature of the disease,including acute disseminated encephalomyelitis(ADEM),Neuromyelitis Optica Spectrum Disorders(NMOSD)and Multiple Sclerosis(MS).Clinical practice has shown that various diseases of demyelination are easy to be misdiagnosed through diagnosis of pathology,clinical symptoms and imaging,resulting in errors in follow-up treatment and affecting the prognosis.Therefore,accurate diagnosis is vital to the treatment of demyelinating diseases in children.Professional doctors usually need to combine clinical symptoms and magnetic resonance imaging(MRI)diagnostic technology to diagnose patients and formulate treatment plans.The lesions of demyelinating diseases in the white matter region have similar lesion appearance,location distribution and signal characteristics on MRI,so it is difficult for clinicians to diagnose.This paper conducts research on the segmentation and classification of children’s demyelinating diseases for the above problems,the main innovations are as follows:1.A model based on convolutional neural networks for the classification and lesion segmentation of demyelinating diseases in children.The model inherits the encoding-decoding framework of the U-Net network,the basic architecture of medical image segmentation,and further down-samples the middle layer to continue extracting the depth features for health,ADEM and NMOSD classification.Segmentation-based method can provide pixel-level lesion information and tissue structure information,such as size and location,which will be more conducive to image classification.Classification-based methods can identify regions of interest within the image,such as the white matter region of the brain,which will promote the accuracy of segmentation.Feature fusion of the two methods can not only promote the segmentation task but also promote the classification task.Compared with the existing models,this approach can not only achieve the joint detection of segmentation and segmentation,but also achieve the accuracy of 97.96% and 71.1% on the classification and segmentation,respectively.2.A classification model for demyelinating diseases based on multi-scale MRI segmentation feature fusion is proposed.This model improves the accuracy of ADEM and NMOSD classification by fusing features in the segmentation.The features are obtained using global max pooling and global average pooling on the output segmented image,which are then combined with the features extracted in the downsampling.To address the imbalance of positive and negative samples in the network model training,a dynamic weight loss function is designed to adaptively control the contribution from the segmentation and classification branches.Compared with existing models,this approach achieves a classification accuracy rate of 99.19% for ADEM and NMOSD.3.A detection system for children’s demyelinating diseases is developed,which includes a GUI interface designed using PyQt.The convolution neural network-based model proposed in this paper for children’s demyelination disease classification and lesion region segmentation is applied to the detection system.The system assists doctors in further analysis of segmentation and classification of demyelinating diseases in children.