| Primary brain tumors that originate in the brain and almost no exact causative agent has been found clinically.In our country every year tens of thousands of patients with primary brain tumors,the incidence of gliomas account for nearly half of the ratio,the most common primary brain tumor.Compared with computed tomography,magnetic resonance imaging has a clearer contrast of soft tissue and can better display the intrinsic structure of the brain.This is of great clinical significance for the diagnosis and treatment of gliomas.In the current clinical diagnosis and treatment,the acquisition of information such as the size,shape and distribution of glioma,and the relative position of the surrounding organs and tissues is a key prerequisite for the quantitative analysis of the tumor.At the same time,for the surgical resection of gliomas and the development of radiotherapy programs,it also has an important guiding role.Therefore,the study of the precise division of glioma is a very important research direction.In MR images,although multi-modality techniques can be used to image many tissue characteristics of gliomas and improve the accuracy of segmentation,accurate segmentation is still a huge challenge.This is mainly because glioma is different from normal tissue,and the shape,texture and size of different individuals often have not a few differences.Further,the limitations of the imaging technology has brought inhomogeneity and image noise,so that the image is complex and difficult to extract glioma its essential characteristics.Currently more common brain tumor segmentation algorithms are mainly classified as threshold-based segmentation,region-based segmentation,pixel-based segmentation,model-based segmentation,atlas-based segmentation and so on.The threshold-based segmentation algorithm is simple and direct but its accuracy is often low,mostly as the first step of other algorithms.Region-based segmentation can obtain regions with similar characteristics,but the segmentation ability of fuzzy boundaries formed by partial volume effect is poor,and is sensitive to noise and easily over-segmentation.The segmentation method based on pixel classification has different segmentation characteristics according to the complexity of the classifier,but all require a lot of annotation images for training,in which the artificial neural network has more accurate segmentation results but also requires a lot of annotation data and training time.The model-based segmentation algorithm can introduce useful structural prior knowledge,but it is sensitive to the initialisation and location of the contour curve and easily falls into the local minima.At the same time,its generalization and anti-disturbance properties are also poor.Segmentation methods based on atlases can take advantage of the high precision of manually segmented images,but their performance is heavily dependent on the registration accuracy of the pattern.In recent years,due to the continuous improvement of the neural network performance,the increasing number of segmented glioma images and the rapid increase of hardware computing performance,these factors make the obstacles to research on the segmentation algorithm based on artificial neural network greatly reduced so that artificial neural network becomes the current major research focus.Fully convolutional DenseNet is one of the best artificial neural networks at present,which makes the network reuse the image features by the form of densely connected blocks,which effectively improves the efficiency of feature utilization and the optimization of network parameters.Fully convolutional DenseNet has achieved excellent performance in the segmentation of natural images,but the direct application to the multi-modality MR image data of gliomas has not achieved good results.Due to the differences in the amount of data and the characteristics of the data,some hyper-parameter values of network structure and training data need to be re-selected so that the network structure and glioma data have a good match.In this paper,experiments are carried out on the six parameters such as training data dimension,image block size,whether to expand the data,the size of the network structure,the specific form of the loss function and the stride size of the fusion step in the image prediction process.The results show that the algorithm chooses the 3-D volume data with the image block size of 64 pixels as the training data and flips left and right for data augmentation,and selects the smaller fusion step under the condition of the acceptable predicted time and chooses the larger network structure under the condition of the limit training time and uses the loss function of Dice loss value by sample-by-sample can improve the performance of fully convolutional DenseNet in glioma segmentation with limited hardware conditions and time.Since the task objectives of the International Multimodal Brain Tumor Segmentation Challenge used in this paper are different from those of the general natural image segmentation,the original full convolutional DenseNet parameter optimization process can only indirectly improve the segmentation accuracy of the target region.In order to directly improve the segmentation accuracy of glioma target region,this paper proposes a multi-Dice loss function to guide the optimization of network structure for multi-region segmentation.After extracting the common image features of all the tumor tissues,each network branch in the structure performs targeted image feature learning on a type of tissue in the target area,so that the network can learn more distinguishable tissue characteristics and improve Segmentation accuracy of MR images of gliomas.In this paper,we use the challenge dataset BraTS2015 for segmentation task.Experimental results show that the proposed method can indeed improve the segmentation accuracy of the brain glioma target region.The average Dice similarity coefficient of the three target regions of the whole tumor area,the tumor core area and the enhancing tumor area of the test set reached 0.851,0.712 and 0.630,respectively. |
PDF Full Text Request