Study Of Brain MRI Tumor Segmentation Based On Fully Convolutional Networks

Glioma is the highest prevalence and mortality of brain tumors.Once the patient is diagnosed,the patient can survive for no more than two years.The tumor is found earlier,the patient can get better treatment.Magnetic resonance imaging(MRI)is a very important method for the glioma diagnosis.However,it is difficult to separate tumor areas from MRI images of patients5 brains.At present,the tumor is mainly separated by doctors by hand.Manual segmentation not only needs much professional knowledge,but also takes much time.Therefore,a fully automatic segmentation method of glioma is very practical.But glioma can infiltrate into the normal tissue area and make the tumor boundary to be blurred.In addition,according to the image-forming principle of magnetic resonance imaging,even the same patient,images obtained under different instruments have different grayscale ranges.Because of the segmentation difficulties,traditional image segmentation algorithms can not achieve good segmentation results.In recent years,the deep convolutional neural network has great advantages in computer vision compared with other traditional machine learning methods.With the development of deep learning algorithm and the realization of stable and high performance,from the perspective of effective deep learning,deep learning can better solve the problem of brain glioma segmentation.In recent years,deep convolutional neural network has achieved great success in the field of computer vision.Based on the present situation,this paper applies the deep convolutional neural network to the brain MRI tumor segmentation.The main work is as follows:(1)Due to the features of glioma and MRI,the traditional image segmentation methods are difficult to separate the tumor area.In order to solve this problem,the paper applies the fully convolutional neural network to the brain MRI tumor segmentation.The task of the paper is to separate the tumor areas in MRI images of the brain.But the tumor area is small,and the internal structure of the tumor is more complex.If the training process directly uses the original MRI images of the brain,which can separate the outline of the tumor.However,it is not very effective for the internal structure of the tumor.The paper divides the brain tumor segmentation task into two steps.First,a rough segmented network is trained to detect the tumor areas in MRI images of the brain.A fine segmented network is trained by extracted tumor area to use to finely separate the tumor area.(2)Due to the interference of the offset field in the image-forming process of MRI images,the brain tissue can not be displayed in the image by its own gray level value.Therefore,the paper proposes an offset field correction method.In addition,different imaging instruments may result in different gray level ranges for the same object.After correcting the original MRI images of the brain,the gray level regularization of the brain MRI images is performed to make the gray level values of all the brain MRI images are concentrated in the same range.(3)For the traditional fully convolutional neural network,the size of the feature maps will be reduced to half of its original size after each pool.After multiple pooling layers,the input image will lose a lot of details.In this paper,dilated convolution structure is applied to the fully convolutional neural network to make the size of the feature map same as the original size and the obtained feature map more densely.At the end of the network structure,the multi-scale convolution kernel is used to further improve the performance of the network.Finally,the conditional random fields of the fully connection is used to optimize the output of the fully convolutional neural network.All the image data used in this paper derive from the BRATS 2015 database.After a lot of experimental analysis and comparison,it can be found intuitively that the proposed algorithm has better brain glioma segmentation performance compared with other similar methods and can provide reference for the diagnosis of auxiliary clinical medicine.