MRI Brain Tumor Segmentation Based On Deep Learning

With the rapid development of society,people’s living standards have gradually improved.But at the same time,people’s irregular lifestyles also make more diseases occur,and tumors have even become a common disease.Brain tumors are one of the common malignant tumors with a high fatality rate.Currently,they can only be treated by surgery or radiotherapy.Therefore,early diagnosis of brain tumors is essential to improve the patient’s condition.Magnetic resonance imaging(Magnetic Resonance Imaging,MRI)can help doctors observe the inside of the patient’s brain and quickly determine the area of the lesion.However,brain tumors have different histological subregions and complex structures.There are also different differences in the images scanned by MRI.It is difficult to distinguish the marginal connection between the tumor and the normal tissue.Usually,imaging experts manually mark and segment the tumor area,but it is extremely time-consuming and there may be errors.Therefore,the technology that automatically segment MRI images has become a challenging task in the computer field in recent years.In order to achieve precise segmentation of brain tumors,this paper uses deep learning algorithms to study this problem.The specific work is as follows:(1)A two-dimensional segmentation algorithm based on U-Net is proposed.First,the densely connected blocks in DenseNet are improved and integrated into U-Net,and the densely connected blocks are used to include all previous outputs in each input,which strengthens the information transfer and feature reuse between convolutional blocks,and the feature extraction capability of the U-Net encoder is improved.And the dilated convolution is incorporated into it,so as to expand the receptive field of the convolution kernel,and improve the connection between layers without reducing the resolution.Subsequently,the Conditional Random Field Recurrent Neural Network(CRF-RNN)is added to finely segment the image to form an endto-end training model.By training and verifying the two-dimensional slice data of BraTs2018,this method realizes the segmentation of brain tumors.(2)An improved 3D U-Net network brain tumor segmentation algorithm is proposed.Firstly,the MRI image is preprocessed to perform N4 ITK bias field correction.By correcting the bias field,the intensity unevenness of the scan is solved.Incorporate the Atrous space pyramid pooling structure(ASPP)into the model,and use multiple atrous convolutions with different expansion rates to obtain multi-scale context information to improve feature extraction capabilities.The attention gate(AG)is added to enhance the acquisition of salient features useful for specific areas and suppress irrelevant information.A segmentation layer is added to the network to create multiple segmentation maps with different resolutions.Instance normalization is used instead of traditional batch normalization to compensate for the randomness caused by the use of small batch training due to memory limitations.In addition,the Exponential Logarithmic loss function replaces the traditional linear Dice loss function,and improves the model performance through parameter adjustment control in the loss function.Through the verification of the BraTs2018 data set,this method effectively improves the segmentation accuracy.(3)A brain tumor segmentation algorithm based on 3D SE-RESUNet is proposed.The model uses encoding and decoding modules to achieve the same input and output size.The Squeeze and Excitation structure is added to the designed residual module to automatically obtain the importance of each channel information from the relationship between the characteristic channels,thereby increasing the degree of attention to useful information and suppressing unimportant information.Deepen the network through the residual module,and cascade the encoding and decoding parts,so that the model can obtain more features.And Generalized Dice Loss is used to deal with the imbalance of data categories.Through the training of the BraTs2018 data set and the test of the BraTs2019 data set,this method effectively segmented the disease area.