Research On Automatic Detection Method Of Brain Tumor Region Based On Multi-sequence MR Image Fusion

Brain tumors are a serious threat to human health.Magnetic Resonance(MR)imaging technology has the characteristics of non-invasive and non-invasive,and has become an important means of brain lesion identification and analysis.Extracting and segmenting the brain MR image of the lesion area,that is,the brain tumor area,is a key technique.Due to the problems of noise,offset field and volume effect in brain MR images,traditional machine learning algorithms or traditional convolutional neural network models is insufficient of extracting and segmenting brain tumor regions.Multi-sequence MR image is composed of a variety of MR image types,the fusions of the brain tumor region more abundant features,thereby improve the segmentation accuracy of brain tumors.Therefore,this paper makes a research of the automatic segmentation method of brain tumor region from the perspective fusions of multi-sequence MR image.The main work is as follow:(1)A multi-sequence MR image brain tumor region segmentation method based on Neighborhood Filter Kernel Low Rank Representation(NFK-LRR)is proposed.By studying the spatial and gray-level features of multi-sequence MR images,the multi-sequence MR images are first fused,and then the neighborhood filter kernel is applied to obtain the characteristics of the high-order structure of the fused data,and the fused data is mapped to the linearly separable high-dimensional space.The neighborhood filter kernel can effectively combine spatial and gray-level features,and combine the related priors to perform homogeneous region segmentation on the image.Based on the discriminate dictionary,low rank and sparse representation modeling is used to construct the brain tumor with feature-driven combination.The low rank representation segmentation method is used to segment the brain tumor region according to the residual representation.Finally,the segmentation results of brain tumor region are obtained by exploiting the prior knowledge of clinical pathological to fine-tune the segmentation,and high segmentation accuracy is taken.(2)In order to further improve the segmentation effect,A multi-sequence MR image brain tumor region segmentation method based on Multichannel Full Convolutional Neural Network(MFCNN)and Markov Random Field(MRF)are proposed.Because the traditionalconvolutional neural network is limited by the size of its convolution kernel,it is difficult to obtain the overall information of the MR brain tumor image,and the convolution and pooling process may cause partial information loss on the shallow layer of the network,resulting in MR.The segmentation accuracy of brain tumor images is degraded.Based on the global channel model with network and the local channel model of shallow network information,the MFCNN model is established based on multi-sequence MR image fusion for preliminary segmentation.Due to the correlation between adjacent pixels of the MR image,adjacent image blocks,and image tags,the MRF is embedded in the MFCNN to segment the boundary and the relationship between the image blocks.Finally,the segmentation results of the final brain tumor region were obtained by fine-tuning the segmentation prior knowledge of clinical pathology.The experiment results show our proposed achieved superior brain tumor segmentation performances in the brain tumor segmentation challenge dataset.(3)A multi-sequence MR brain tumor image processing system is designed and implemented.The processing system includes a registration module,a fusion module and a segmentation module.The segmentation module consists of two parts: segmentation module(1)Multi-sequence MR image brain tumor segmentation based on NFK-LRR,suitable for training dataset is insufficient and there is insufficient training time;segmentation module(2)based on MFCNN and MRF Sequence MR image brain tumor segmentation can achieve better segmentation accuracy when the training dataset is sufficient and there is sufficient training time.