| Infantile period is the fastest growing period of human brain.During this period,brain regions change rapidly,and it is easy to get all kinds of brain diseases,but in this period infants and young children are unable to express their condition,so a good medical diagnosis method is very important.Infant brain MR images have low resolution,the inversion of gray matter and white matter,large noise and the partial volume effect during development.At present,there is still no perfect segmentation algorithm for infant brain tissue.In recent years,with the development of technology,the application of deep learning in image segmentation is becoming more and more popular.This paper improves the existing deep learning semantic segmentation network to make it more suitable for the segmentation of infant brain tissue images.The main contents of this thesis are as follows:(1)Aiming at the low resolution and high complexity of infant brain MR image,this paper uses deep convolution neural network to extract complex abstract features adaptively,and improved feature extraction algorithm for infant brain.The residual module with bottleneck layer added to the traditional VGG16 structure.On the one hand,this method can avoid the degradation effect caused by the excessive number of layers,the gradient disappearance and gradient explosion when the network layers are designed very deep.On the other hand,the algorithm use bottleneck layer in the traditional residual block,which effectively reduces the training parameters of the network.The residual module is introduced into the semantic segmentation network to improve the ability of extracting features from infant brain tissue.(2)The feature fusion part in deep semantic segmentation network is studied.To avoid the problem of considering only one dimension feature in network feature fusion,a multi-scale and multi-level context information extraction module is proposed.The deep feature map mainly captures the semantic features of classified objects.The shallow feature map captures the details of objects.Combining the two features can improve the accuracy of feature expression.Therefore,this paper designs a context information extraction module,which uses multiple hollow convolution kernels with different expansion rates to form a pyramid structure in parallel to obtain multi-scale features of feature maps in different receptive fields,enhance the recognition ability of the network to target pixels,and improve the accuracy of semantic segmentation.(3)Because of the sampling operations in deep convolution neural network,the information of target location and structure are lost,which resulting in reducing the classification accuracy of edge pixel of brain tissue.Aiming at the problem of poor target edge processing,a boundary points reclassification algorithm is proposed in this paper.Based on deep learning semantics segmentation network,we design a algorithm to solve the pixels with low prediction credibility.By using the method of region division and local fixed dictionary learning,we reclassify the low prediction credibility pixels and further improves the semantics segmentation results to achieve accurate segmentation of infant brain tissue. |
