MRI Hippocampus Segmentation Based On Deep Learning

The hippocampus is located between the cerebral thalamus and the medial temporal lobe,and is mainly responsible for the storage of long-term memory.Abnormal volume and function of the hippocampus are closely related to many mental diseases.Therefore,accurate segmentation of the hippocampus can assist physicians in the diagnosis and treatment of related mental diseases,and has great medical value.Magnetic Resonance Imaging(MRI),as an advanced medical image acquisition technology,the acquired image not only has a high soft tissue resolution,but also provides rich contrast and high-resolution 3D brain tissue information.Therefore,studying the volume and shape of the hippocampus in brain MRI images and achieving accurate segmentation of the three-dimensional hippocampus have gradually become an important task in medical image research.The segmentation methods for brain MRI hippocampus are mainly divided into the following categories: manual segmentation,semi-automatic segmentation and automatic segmentation.Usually manual segmentation methods are cumbersome,time-consuming,and error-prone.Semi-automatic segmentation usually relies on other auxiliary technologies such as classifiers and optimizers,and the segmentation accuracy is often not very high.In order to achieve three-dimensional accurate segmentation of the hippocampus,this paper mainly studies the segmentation of the hippocampus through deep learning methods,the main contents are as follows:(1)Hippocampus segmentation algorithm based on convolutional neural network and multi-view ensemble(CNN + BDC-LSTM).First,the convolutional neural network is used to segment the 2D slice sequence in the three views of the coronal,sagittal and axial views,and then the segmentation results in the three views are integrated to obtain the final result.The convolutional neural network consists of the encoder,convolutional long and short memory network(Bi-Directional CLSTM,BDC-LSTM)and the decoder.In order to obtain multi-scale information and expand the receptive field of the convolution layer,the encoder uses asymmetric convolution layers and dialted convolutions of different sizes.In addition,BDC-LSTM is used between encoder and decoder to fully mine relevant information between slice sequences in single view,thereby improving segmentation accuracy.(2)Based on Dilated-3DUnet segmentation hippocampus algorithm.The model is based on a three-dimensional fully convolutional neural network and an end-to-end convolutional neural network is designed.The number of channels of the convolutional layer in this network is distributed in a "pyramid" way,which effectively reduces the scale of the parameters.In addition,the use of 3D dilated convolution as a cascaded convolution operation not only effectively combines the deep and shallow features of brain MRI images,but also expands the convolution without changing the number of parameters.The receptive field of the layer has acquired multi-scale information,which can better capture the shallow features of the brain MRI image,thereby improving the segmentation accuracy.(3)Based on 3D-DilAttenUNet segmentation algorithm of brain hippocampus.The structure of the network is based on 3DUnet,and the attention gate(Attention Gate,AG)inAttention U-Net is added.Because of this,a large number of additional model parameters is not required,the characteristic response of the unrelated background area is suppressed and the model's sensitivity and prediction accuracy are improved.In order to make the model easier to train and quickly converge,the Resiual Unit module is added to encoder of3D-DilAttenUNet.By using the convolution of holes with different expansion coefficients,multi-scale information is obtained.In addition,in order to avoid the problem of "gradient dispersion",deep supervision was added.Experiments on two public databases have found that the 3D-DilAttenUNet model is more effective when processing MRI images,surpassing other algorithms in accuracy,and has a greater advantage.