Generative Adversarial Networks For Hippocampus Segmentation

In recent years,with the rapid economic development,the aging of the world’s population has become more and more serious,and the incidence of mental system diseases such as Alzheimer’s disease and major depression has also been increasing year by year.The shape and volume of the hippocampus is inextricably linked to the etiology of Alzheimer’s disease.The hippocampus affects human memory and cognition and other functions.Once a problem occurs,it will have a great impact on people’s daily life.The first step to diagnose Alzheimer’s disease is to segment the shape of the hippocampus from the magnetic resonance image,and then perform the next step of analysis to confirm the diagnosis.At present,the gold standard of hippocampus segmentation is still used by experienced professionals to outline the hippocampus contour of each slice with the help of specialized medical image processing tools.This process is time-consuming,tedious and less repetitive.It usually takes at least two to three hours to segment a pair of hippocampus,and the organizational structure of interest needs to be divided layer by layer,which is highly subjective and often has a certain degree of error.Therefore,in order to overcome the shortcomings of manual segmentation methods,realizing the automatic segmentation of the hippocampus has important practical significance for the diagnosis of neurological diseases.The generative adversarial network was proposed in 2014.At first,the generative adversarial network was only applied to natural images.Due to its powerful performance and flexible framework,it has gradually been applied in the field of medical images in recent years,including medical image segmentation,reconstruction,and denoising.Therefore,this paper studies the hippocampus segmentation method based on the generative adversarial network.The main research contents are as follows:(1)A hippocampus segmentation method is proposed that combines the attention mechanism of residual block and the generative adversarial network.Based on the generation of the adversarial network model,different convolution configurations are proposed to capture the information obtained by the segmentation network.A generative adversarial network model based on Pixel2 Pixel is proposed.The generative model combines the residual network and the codec structure of the attention mechanism to capture more detailed information.The discriminant network uses a convolutional neural network to discriminate between the segmentation results of the generative model and the expert segmentation results.Through the generative model and the discriminant network,the loss is continuously transmitted,so that the generative model reaches the optimal state of segmenting the hippocampus.T1-weighted MRI scans and related hippocampal labels from 130 healthy subjects from the ADNI dataset were used as training and testing data.Experimental results show that the network model can achieve efficient and automatic segmentation of the hippocampus.(2)A hippocampus segmentation method combining convolutional LSTM and generative adversarial network is proposed.Convolutional LSTM is incorporated into the generative model of the generative adversarial network.The spatial features can be extracted through convolution operation,the spatial information between two-dimensional slices can be better processed,and the three-dimensional features can be fully utilized.Because the data of the hippocampus belongs to three-dimensional data,in order to save computer resources and make the model convergence faster,it is necessary to make the model parameters as small as possible,so two-dimensional slice data is used for training.Secondly,the hole convolution is used instead of the general convolution to further improve the segmentation accuracy of the hippocampus under the condition that the parameter quantity is unchanged.After training the model,the 3D reconstruction results are compared with the original hippocampal data.Experimental results show that the model can effectively use the three-dimensional spatial information between data slices of the hippocampus to achieve accurate segmentation of the hippocampus.