Research On Hippocampal Segmentation Algorithm Based On Semi-supervised And Boundary Enhancement

In recent years,with the aging of the population,the incidence of Alzheimer’s disease has gradually increased.In addition,the process of developing Alzheimer’s disease patients with dementia symptoms is irreversible.There is no medical treatment to cure the disease,so the early diagnosis of Alzheimer’s disease is particularly important.Because of the early atrophy of hippocampus in Alzheimer’s disease,it is of great significance to quickly and accurately obtain the volume change of hippocampus in brain MRI images.Most of the current advanced hippocampus segmentation algorithms are based on deep learning algorithms,but two problems faced by deep learning-based hippocampus segmentation tasks: 1)Scarce MRI samples containing hippocampal body annotations,which are difficult to meet the requirements for the capacity of training samples;2)In the magnetic resonance image,there is an insignificant boundary between the hippocampal structure and its surrounding tissues,and the hippocampus in the human brain is small,making it difficult for the hippocampus segmentation algorithm based on deep learning to accurately segment the boundaries.Aiming at the above problems,taking the hippocampus segmentation in brain MRI images as the research goal,two algorithm of hippocampus segmentation based on deep learning are proposed around the algorithm segmentation accuracy and model robustness.The main work of this thesis is as follows:1)This thesis designs a semi-supervised hippocampus segmentation algorithm based on generative adversarial networks.First,a pair of generators is constructed based on the generation of the adversarial network,and the weight of one generator network is obtained by a moving average of the weight of the other one.Then using the unlabeled data to train paired generator network for unsupervised training,and using the labeled data to train adversarial network for supervised adversarial training,thereby constructing a generative adversarial network based on semi-supervised training.On the EADC-ADNI dataset,the adversarial training strategy,semi-supervised training strategy,the combination of the two,and the fullysupervised training strategy are compared.The results show that in IOU accuracy,the semisupervised strategy and the adversarial training strategy are improved by 1.8% and 0.58%,respectively.And the combination of the two algorithms improves the accuracy of IOU by2.53%.Without considering the complexity of the network structure,the semi-supervised training strategy based on generative adversarial networks designed in this thesis can effectively improve the performance of the hippocampus segmentation algorithm.2)This thesis proposes a hippocampus segmentation algorithm for training networks with attention mechanism using boundary enhanced loss.In order to make full use of the sample spatial information and simplify the network structure,a fully supervised model based on 3D convolution is designed.First,a U-shaped three-dimensional convolutional neural network with an attention mechanism is constructed.The low-level information of the Ushaped segmentation network is weighted and reused to improve the efficiency of feature flow in the neural network.Then,for the problem that the boundary between the hippocampal structure and the surrounding tissue structure is not obvious,a boundary-enhanced loss function is designed.The convolutional layer of the fixed kernel is designed to automatically extract the hippocampal boundary of the training label.The performance of the loss function and network structure proposed in this thesis is analyzed on the EADC-ADNI dataset.Compared with several advanced three-dimensional segmentation algorithms based on deep learning,the experimental results show that the proposed algorithm has the best performance,reaching 89.41% in Dice accuracy.By enhancing the network’s attention to the boundary area during traing,the model obtains a smaller verification variance and is more robust.