| Brain tumor is a common malignant disease that threatens human health.Timely detection and accurately diagnosis can gain valuable time for treatment.Clearly imaging and accurately segmentation of brain tumors are the premise of accurate diagnosis and treatment.However,the brain tissue structure is relatively complex,the shape and growth position of brain tumors are different,the internal gray level is uneven and the edge is fuzzy,which makes accurate segmentation of brain tumor images a challenging task.At present,clinicians often use computer-aided sketch to segment brain tumor images.Although the accuracy of artificial segmentation is high,it will undoubtedly waste a lot of time in the face of the huge amount of image data to be processed every day.Therefore,the research of brain tumor image automatic segmentation method has been the focus of scholars.In recent years,deep learning method has achieved a record performance in various fields,and has won the champion in the natural pattern recognition competition.Researchers also proposed a large number of models for medical image segmentation by improving the network structure,and achieved high accuracy.In this paper,we mainly did the following research on brain tumor image segmentation.(1)In the process of imaging,magnetic field deviation will occur in MRI equipment,which will affect the imaging quality and cause uneven brightness in the image.To solve this problem,this paper used N4 ITK algorithm to correct the bias field.The brightness of the image after the bias field correction is uniform,which improves the computer operability of the image.In addition,because the images came from different devices,the gray distribution range is also different,so the normalization operation is needed before training.In this paper,the 0-1 normalization method was used to unify the gray distribution range of the image.Then,the improved cross entropy loss function was used to enhance the learning ability of the network.In solving the problem of the lack of data,this paper used the data augmentation to expand the dataset and improve the generalization ability of the model.(2)Dilate convolution can enlarge the receptive field without reducing the image resolution.According to the characteristics of dilate convolution,a Multi-scale Feature Extraction Block based on Dilate Convolution(MD)was designed to extract brain tumor image features.Different from other multi-scale feature extraction methods,MD can obtain a larger receptive field through continuous dilate convolution,and fuse the result of each dilate convolution with the input of the module as the output result,so the output not only contains the multi-scale feature map,but also retains the detailed features of the input image.By adding MD to U-Net network,a multi-scale brain tumor segmentation model combined with dilate convolution was proposed,the of complete tumor segmentation is 0.86,which is higher than that of U-Net by 3%.(3)Adding MD increases the depth of the network,which may bring difficulties in training.Therefore,a multi-scale brain tumor segmentation model combining depth supervision was proposed by combining multi-scale feature extraction module with depth supervision module.Firstly,the model uses small convolution kernel to extract the shallow features of the image,and then uses MD to expand the receptive field quickly,while retaining the shallow features of the image to guide the feature recovery in the up-sampling process,which improves the training speed and segmentation accuracy of the network.The of complete tumor,tumor core and enhancing tumor were 0.9323,0.9430 and 0.9077,respectively.In this paper,all the methods were validated on BraTS2017 dataset,the of complete tumor,tumor core and enhancing tumor were 0.89,0.87 and 0.79,respectively.The experimental results show that our method is superior to most brain tumor image segmentation methods and it is proved that small convolution kernel is more effective to extract detail features in shallow network. |
PDF Full Text Request