Research Of Medical Image Segmentation Model Based On Deep Learning

In the medical field,the use of deep learning as a machine learning and pattern recognition tool has always been the focus of research.The application and research of deep learning technology in the medical field mainly includes: intelligent image recognition,medical image assisted diagnosis,diagnosis and treatment result prediction,drug development,and gene sequencing.Medical imaging is an important adjunct in modern healthcare and computer-aided diagnostic systems.In view of the limitations of current large medical image data and manual segmentation and semi-automatic segmentation methods,it is particularly important to deal with medical image information by means of big data analysis.There is an urgent need for an efficient automatic segmentation method.In recent years,researchers have proposed a variety of segmentation methods,However medical images are easily affected by internal or external factors,which makes medical image segmentation still a challenging task.Therefore,accurate and efficient analysis of medical images is of great practical significance for formulating precise treatment plans,improving the efficiency of medical staff,taking appropriate treatment measures in a timely manner,and balancing the imbalance of medical resources.In this paper,based on the current problem,using the powerful nonlinear representation ability of deep learning,the improved 3D-FCN+CRF model and the MS-CapsNetGAN model are used to process the segmentation tasks of brain tumor images.The main research contents of this thesis are as follows:(1)Overview of image segmentation and research status.First,the seven methods commonly used in image segmentation at this stage are introduced.The principles,steps,and advantages and disadvantages of each method are described.Then,in order to scientifically quantify the performance of image segmentation algorithm,the DSC method for image segmentation result evaluation and the evaluation indexes such as accuracy,accuracy,recall rate,over-segmentation rate and under-segmentation rate are introduced.(2)Brain tumor image segmentation based on 3D-FCN+CRF model.Firstly,the convolutional neural structure and the principle and function of each layer in the structure are introduced.Then,the convolutional neural network can not extract the difference information between the modes.The difference in tumor size of different image layers is significant.Defects such as low segmentation accuracy,this paper proposes to solve the above problems based on the 3D-FCN network model,and uses the conditional random field CRF to process the image boundary problem in the backend,which increases the segmentation accuracy of the image and overcomes the different image layers between the brain tumors.The difference in size and position;Finally,through the BRATS2015 dataset,the Clinical Data dataset and the multimodal magnetic resonance MRI image of 100 patients in a hospital,the Dice coefficient reached 89.23%,96.77% and 91.64%.It is proved that the model can significantly improve the segmentation accuracy,better extract the difference information between the modes,and segment the brain tumor more accurately.(3)Brain tumor image segmentation based on MS-CapsNetGAN model.Firstly,the capsule network and the layered structure and principle and function of generating the anti-network are introduced.Then,the capsule network is used instead of the standard convolutional neural network,applied to the discriminator that generates the anti-network,and the multi-scale capsule coding unit is introduced at the bottom.A multi-scale generated confrontation capsule network MS-CapsNetGAN model was proposed;subsequently,the MS-CapsNetGAN model was used to perform validation experiments on the MNIST and CIFAR-10 datasets,and the MS-CapsNetGAN model was compared with three mainstream classifiers(Fisherfaces,LeNet and ResNet).the model shows that the model has better segmentation effect.Finally,the brain MRI image of 100 patients in a hospital was segmented by MS-CapsNetGAN model,and the Dice coefficient reached 93.61%.The experimental results show that the constructed MS-CapsNetGAN model can accurately segment MRI images and has important application value.