Study On The Segmentation Method Of Multiple Sclerosis Lesions In Brain MRI Images

Multiple sclerosis(MS)is a common central nervous system disease.The main symptoms are numbness and dizziness in the limbs,which may cause stroke,visual dysfunction,etc.,usually in the white matter of the brain.Brain Magnetic Resonance Imaging(MRI)technology has the advantages of high resolution and no radiation to humans,and has become the main means of clinical diagnosis of MS.Segmentation of MS lesions from MRI images can establish a basis for subsequent lesion reconstruction,volume estimation,and disease assessment.However,due to the 3D multi-fault nature of MRI images,imaging experts are time consuming and laborious in manually segmenting multiple sclerosis lesions and subjectively uncertain.Therefore,studying an algorithm for automatic segmentation of MS lesions in the brain is of great significance for computer-aided diagnosis of MS.This paper proposes two different MS segmentation methods for 2D and 3D images.Aiming at the 2DMRI tomographic image,a method based on the maximum inter-class variance is proposed.Using the fast optimization feature of the differential search algorithm,the optimal threshold is obtained to segment the brain tissue,and then the MS lesion tissue is further segmented on the white matter segmentation result.The method was carried out in the Brainweb simulation MRI image dataset,and good experimental results were obtained.The DSC average value can reach 92.41%,and it has good stability.In addition,for real 3D MRI images,this paper proposes an algorithm based on 3D Convolutional Neural Network(CNN).The algorithm is based on a concatenation of two convolutional neural networks of the same structure.The first stage CNN segmentation of possible candidate lesion voxels,while the second stage CNN is used to reduce the number of voxels misclassified in the first stage.MS lesions occupy a small volume in the brain,and the health voxels are much larger than the lesion voxels,which results in an imbalance in training data.In order to solve the data imbalance problem,we selected the same number of healthy voxels in the first stage of network training for training.In addition,because of the small amount of MRI datain the brain when training the network,we use the data expansion method to solve the over-fitting problem.The algorithm was tested on the MSC lesions challenge dataset of MICCAI2008 and MICCAI 2016,and achieved good results.When the lesion volume is more than 5ml,the DSC average value can reach 70%,with high sensitivity and reasonable false positive rate,and the VD difference value is the lowest.The comprehensive performance index is better than the EM,RF and 2D CNN methods proposed by predecessors.