The Research Of Segmentation In MRI Using Multi-scale And Fully Convolutional Network

Segmentation of Magnetic Resonance Imaging(MRI)in medical images is one of the important steps in clinical diagnosis.The traditional artificial segmentation results are easy to be influenced by the experience and subjective factor of doctors,and it is a heavy and time-consuming work.Automatic segmentation of computer-aided diagnosis can improve the accuracy and efficiency of segmentation.Most of the traditional automatic segmentation methods are based on the processing technology of digital image,and segmentation is realized by extracting the shallow features of the image,such as the texture and so on.Segmentation based on shallow features is not stable,easy influenced by contraction and noise data,and it is not able to extract deep abstract features of regions of interest,so the generalization ability is not good.In recent years,the development of deep learning has promoted the semantic segmentation of images using convolutional neural networks in the field of computer vision,and laid the foundation for automatic segmentation in MRI images.It can improve the segmentation effect by learning the characteristics from the data with good ability classification and generalization.The main contents of this researches are as follows:(1)A Side U-Net segmentation model is proposed.The network used a mixed dilation convolution and multi-scale up-sampling module to effectively recover the local features which maybe fade in the down-sampling process,and combines the side output module with the ground truth to calculate the error at different scales directly to increase the efficiency of the network updating at different scales.The experimental results show that the stacked network structure proposed in this thesis has higher accuracy when using the same loss function.(2)In order to solve the problem of uneven distribution of positive and negative pixels in MRI,an improved balanced cross-entropy loss function is proposed,which transforms the problem of category imbalance in sample pixel number into the problem of difficulty degree of sample pixel classification.It can improve the accuracy of classification by giving more weight to the sample of difficult classification.(3)Proposed a dense aggregation network based on side output network,which makes the relationship between the modules more closely.The feature map of side output can be used as an input to continue the operation of the next layer in network,and the dense up-sampling convolution is used to learn to recover the up-sampling feature without adding additional operations and parameters.The experimental results show that the precision of proposed deep aggregation network is 89.82%.