CNN-based Multimodal MRI Brain Tumor Image Segmentation Modeling

MRI brain tumor image segmentation refers to the separation of the entire tumor area,tumor core area and tumor enhancement area from normal brain tissue.The main challenge of traditional segmentation algorithms is the large gray-scale similarity between brain tissues in MRI images and the differences between different cases.Multi-modality MRI brain tumor image segmentation can make full use of the feature information of different modalities in MRI images and improve the effectiveness of segmentation.It has been a research focus of brain tumor image processing in recent years.The thesis uses CNN,the mainstream image processing technology in deep learning,as a research tool to improve the segmentation efficiency and effect,and constructs a multi-modal MRI brain tumor image segmentation model based on 2D and 3D,respectively.The 2D model is suitable for fast segmentation and pre-segmentation,and the 3D model is suitable for precise segmentation.The main research work of this thesis is also divided into two parts:(1)CNN-based 2D multimodal MRI brain tumor image segmentation modeling and experimental analysis.Due to the relatively small amount of data,the 2D model has the advantages of fast segmentation speed and easy adjustment.To this end,the thesis builds a 2D network model U-Net-Mix suitable for brain tumor segmentation tasks based on the basic framework of fully convolutional neural networks.The main design of the model has four points.One is to design the convolution block and the deconvolution to deepen the network depth to improve the segmentation accuracy;the second is to add a dilated convolution to expand the receptive field without changing the existing feature extraction method.In order to make up for the lack of overall information loss of 2D images;the third is to use a variety of image data expansion methods to increase data diversity to alleviate the problem of low model generalization caused by too little 2D image data;the fourth is to use intensity standardization and normalization algorithms to balance the gray value difference between different images.The U-Net-Mix model has been tested on the industryrecognized BraTS 2017 dataset.The results show that the Dice segmentation coefficient can reach 0.90 on the entire tumor area.(2)CNN-based 3D multimodal MRI brain tumor image segmentation modeling and experimental analysis.MRI brain tumor images are essentially 3D data,and analysis using 3D models can undoubtedly improve segmentation accuracy.Therefore,based on the 2D model,the thesis constructs a 3D fully convolutional neural network segmentation model U-Net-Deep.The main design of the model also has four points.One is to add the self-designed feature extraction module and upsampling module to continue to deepen the network to extract more features;the second is to introduce Dropout technology to alleviate the overfitting phenomenon caused by the high complexity of the deep network;the third is to design a superimposed residual block to alleviate the vanishing gradient problem in deep networks;the fourth is to use a mixed loss function instead of the traditional Dice loss function to fully consider the contribution of weights in a relatively small area to alleviate the problem of data imbalance.The U-Net-Deep model has been tested on the industry-recognized BraTS 2018 dataset.The results show that the Dice segmentation coefficient reaches 0.91,0.82 and 0.66 on the full tumor area,the tumor core area,and the tumor enhancement area,respectively.Slicing and segmenting 2D MRI brain tumor data,the model has the advantage of rapid segmentation while ensuring a certain accuracy;segmenting 3D MRI brain tumor data,although the accuracy has been improved to a certain extent,the training time of the 3D model relative to the 2D model has increased many times,and the cost is too high.In order to improve the accuracy of brain tumor image segmentation while ensuring segmentation efficiency,2D MRI brain tumor data and 3D MRI brain tumor data can be combined in the future for exploration.