Research On Thalamic Segmentation Based On Convolutional Neural Network

The thalamus is the transit station of the human brain,receiving all sensory signals except olfaction and transmitting it to the cerebral cortex.The thalamus can be divided into multiple nuclei,and each nuclei have its specific function.The nucleus is connected to a specific cortical area or becomes a relay between cortical connections.Many neurological diseases are closely related to the damage of the thalamic nuclei,such as Alzheimer’s disease,Parkinson’s disease,schizophrenia,epilepsy and many other diseases.Deep brain stimulation surgery can effectively treat these diseases by implanting a pacemaker on a specific nucleus of the thalamus.Therefore,accurate thalamic segmentation has important value and significance for brain cognitive research,mechanism research and diagnosis and treatment of neurological diseases.Magnetic Resonance Imaging(MRI)has many advantages such as high resolution,good contrast,and no radiation.It is widely used in brain cognitive research and diagnosis and treatment of neurological diseases.Magnetic resonance imaging provides a good basis for accurate segmentation of the thalamus.Manually segmenting an MRI image is a very time consuming and cumbersome task.Automatic and accurate thalamic segmentation is of great value for subsequent diagnosis and treatment.As a new technology developed in recent years,convolutional neural networks are increasingly used in medical image segmentation.However,existing methods for convolutional neural networks fail to adequately consider imaging noise and low contrast between the thalamus and surrounding tissues.In response to these challenges,this thesis proposes a new type of convolutional neural network,the Residual Dense U Network(RDU-Net).RDU-Net introduces dense connections in the classic DenseNet,improves the feature extraction capability of each layer of the network,reduces the image noise,and makes the extracted features better distinguish the thalamus and surrounding tissues.At the same time,the ResNet residual learning strategy is introduced to solve The problem of gradient dispersion caused by network deepening has greatly reduced the training difficulty of deeper networks,and the segmentation effect has been significantly improved.Finally,the bottleneck design in InceptionNet is introduced,and the network parameter size is compressed to ensure the feature extraction of the network.Under the premise of capability,it solves the problem that the network requires too much computing resources.In order to evaluate the performance of the RDU-Net network model proposed in this thesis,we selects the real universal IBSR and HCP data sets for experiments.The segmentation performance was evaluated using Dice Similarity Coefficient(DSC),Intersection over Union(IOU),Absolute Volume Difference(AVD),and Hausdorff Distance(HD).Experiments show that RDU-Net can effectively segment the thalamus in brain MRI images,and each evaluation index has significant advantages over the classical network model.It verifies the effectiveness of RDU-Net thalamic segmentation and has great clinical application prospects.