Research On Brain Tumor MRI Image Segmentation Algorithm Based On Convolutional Neural Network

According to the clinical medicine,glioma is one of the most common and invasive tumors,which results in a very short life expectancy for patients,so the timely and effective treatment is crucial.Doctors need the accurate segmentation of brain tumor magnetic resonance imaging(MRI)images for early diagnosis,treatment planning,but the amount of 3D image data generated by MRI is very large,the shape and boundary of brain tumors are irregular and fuzzy.These factors hinder doctors from effectively and manually segmenting tumor boundaries,limit the precise quantitative measurement of brain tumors in practice.Therefore,a fully automatic segmentation of brain tumor images by means of automatic segmentation algorithms has become a development trend.But the traditional segmentation algorithms have certain limitations for the irregular shape and blurred boundary of gliomas,so the automation of brain tumor segmentation is still a challenging problem.This paper focuses on the brain tumors image segmentation task,takes multi-modality MRI images as the research object,studies the MRI image segmentation of brain tumors based on the deep learning algorithm.The models designed in this paper were evaluated on the Multimodal Brain Tumor Segmentation Challenge(Brats)2015 training data set.The main contributions of the paper are as follows:(1)Aiming at the limitation of small convolutional kernels' detection range in traditional convolutional neural network(CNN)models,an improved CNN model called Small Kernel Two-path CNN(SK-TPCNN)model is proposed based on the characteristics of brain tumor images.This model parallelizes the small convolutional kernels and the large convolutional kernels in the form of two paths,extracts the local texture features and the wider spatial neighborhood information of the image while enhancing the nonlinear mapping ability of the network.The experiment results show that the SK-TPCNN model realizes the simultaneous extraction and fusion of local and global features,achieves good segmentation results.(2)Aiming at the defects of SK-TPCNN model which is prone to feature redundancy and over-fitting,a joint optimization model of SK-TPCNN and random forest(RF)is proposed.The output features of the SK-TPCNN optimal parameters model are effectively integrated using an RF classifier.The experiment results show that the joint optimization model of SK-TPCNN and RF significantly improves the segmentation result of SK-TPCNN on multiple metrics.(3)In view of the shortcomings of traditional CNN based on the image patch classification for automatic segmentation,such as data redundancy and cumbersome training process,the U-net model is used to solve the segmentation problem firstly.Then,an improved model called Deeper ResU-net is proposed.By extending the U-net model to a deeper layer,and adding the residual unit to ensure effective and stable training of the model.The experiment results show that Deeper ResU-net can effectively eliminate the degradation phenomenon of deep network during training while expanding network depth,and improve segmentation performance.