Segmentation Of Multiple Sclerosis In The Brain Based On Deep Learning

Multiple sclerosis is a common disease of the central nervous system.MS has demyelinated plaque during acute activity,and its old lesions form calcified plaques.Patients have systemic symptoms.Severe patients will cause blindness,memory loss,and even paralysis.At present,the most effective auxiliary diagnosis method is MRI.Because it will take a long time for experts to describe the lesions,it is necessary to realize the automatic segmentation.To solve this problem,this paper proposes a MS lesion segmentation method based on deep learning.Main tasks as follows:(1)Research on current image processing technology,study a large number of related literatures,analyze the background and current research situation at home and abroad,and summarize and compare the commonly used methods.(2)Introduce the basic principles of CNN,describe the development of neural network,analyze the basic principles and functions of various components of CNN.(3)A CNN-based image classification method is proposed to determine whether an MRI has MS.Firstly,the MRI were preprocessed such as de-brain,offset field correction and registration.Then the AlexNet model was improved to obtain a new network model for classification.In addition,we used ImageNet to pre-train,and migrated the pre-trained parameters to the new network.After fine-tuning,input the pre-processed MRI for training and testing.Finally,the classification is implemented.(4)A method based on FCN for MS lesion segmentation was proposed.Firstly,the FCN network structure and the principle of upsampling are introduced.Then the image is presegmented using the superpixel algorithm.Finally,the segmentation results are compared with the existing methods,and the lesion volume is estimated.