MR Brain Tumor Image Segmentation Algorithm Based On Fully Convolutional Neural Network

Malignant brain tumor(also known as brain cancer)is one of the most horrible types of cancer in the world.It often causes cognitive impairment and poor quality of life.Due to the superiority and non-invasiveness of magnetic resonance imaging(MRI)technology,it can provide greatly help for the diagnosis and treatment of brain tumors.By tracking the tumor changes of patients,it can provide different treatment plans for different patients.So the brain tumor segmentation based on MR image is becoming a hot issue in research at home and abroad.However,because brain tumor can occur anywhere in the brain and vary in size and shape.These make brain tumor segmentation based on expert labor time-consuming and laborious.And the subjectivity of segmentation result is also large.It is difficult to take into account the accuracy and automation of segmentation based on the traditional machine learning image segmentation algorithms.So it can be seen that brain tumor segmentation is still a challenging task.With the developing of artificial intelligence and deep learning,the method based on convolutional neural network is applied in various scenarios.So this dissertation will be based on multi-modal MR brain tumor images and deep learning methods,do the following researches for brain tumor image segmentation:1.The segmentation accuracy of the existing machine learning algorithms for brain tumor image segmentation is not high.A brain tumor image segmentation method based on improved fully convolutional neural network(FCNN)is proposed.Firstly,the MR brain tumor images of FLAIR T2 and T1 C are gray normalized.And then gray image fusion method is used to get more brain tumor feature information.Furthermore,brain tumor image is segmented cursorily with the improved FCNN which fusing the tertiary pooled feature information.In order to speed up the convergence degree and improve the accuracy of the training model,batch normalization(BN)layer is added after each convolutional layer.Finally,the conditional random fields(CRF)is integrated to the FCNN to fine the segmentation result.Compared with the traditional MRI brain tumor segmentation methods of convolutional neural network(CNN),the experimental results show that the segmentation accuracy and stability has been greatly improved.Average Dice can be up to 91.29%.And the real-time performance of the proposed method is good.2.In order to make the segmentation result provide substantive guidance to the expert physician.And the two-dimensional segmentation result is also not reconstructed in three-dimensional space with the algorithm.An automatic brain tumor image segmentation algorithm based on 3D U-Net and residual technique is proposed.Firstly,FLAIR T2 and T1 C images are de-biased the field effects to improve the image contrast of brain tumor.Then,the fusion network of 3D U-Net and residual modules is established.Dice loss is used as the network optimization function and the adaptive learning rate Adam is used as the network optimization algorithm.In training stage,the training set is used to automatically learn the parameters of the model.And the validation set is used to monitor the over-fitting degree of the model.Finally,the trained model is used to segment and evaluate the brain tumor image on the test set.The experimental results show that the proposed algorithm can achieve high-precision segmentation of threedimensional brain tumor image.The average Dice can be up to 92.52%.Surely,the multi-classification task segmentation of the tumor core part is also successfully achieved.In this dissertation,different brain tumor image segmentation algorithm frameworks are designed from two different perspectives.Both of them have achieved good segmentation results.And it can provide guidance for clinical segmentation.