Research On MRI Segmentation Methods For Infant Brain

The accurate segmentation of brain into cerebrospinal fluid,gray matter,and white matter helps doctors assess the development of the nervous system and diagnose brain diseases.In recent years,deep learning methods have made great progress in the field of medical image analysis,but there are still certain limitations in infant segmentation tasks: 1)Magnetic Resonance Images(MRIs)will inevitably have partial volume(PV)effects and intensity inhomogeneity in the imaging process,resulting in deviations between image pixel values and true values.Reduce the robustness of data-driven deep learning methods.2)For the infant brain at iso-intense phase(approximately 6 to 9months of age),due to continuous myelin formation and maturation,gray matter and white matter show similar intensity distribution in the image,which results in extremely low gray levels between the two tissues.At the same time,there are a large number of edges in the infant brain image,which further increases the difficulty of segmentation.In view of the above chanllenges,this article uses deep learning methods to study infant brain segmentation algorithms.The specific research work is as follows:1.This thesis proposes a rotation-driven convolution mechanism,which can rotate the feature map when performing 2D convolution,use the same convolution kernel to extract the features of different axes,and introduce multi-axis information into the network.Then,the feature maps of different axial planes are rotated to the same axial plane,and the maximum feature response value is taken at the same position as the final convolution result.This process is called Max-View-Maps,which can effectively reduce the impact of PV effect.Based on the new convolution mechanism,this thesis constructs Rubik-Net and conducts experiments on i Seg2017,i Seg2019 and IBSR data sets.Experiments show that the proposed convolution mechanism can effectively extract the spatial information of the image,which can help the network to better complete the segmentation task.2.In this thesis,the active contour model energy term is introduced into the cross-entropy loss function to assist the network in learning the edges in the image and help the network better locate the edge information.Then,this thesis proposes a cascaded neural network segmentation model.In one-stage segmentation,the cerebrospinal fluid is separately segmented by the network.A symbolic distance function is constructed according to the edges of the cerebrospinal fluid and gray matter to restrict the existence range of gray matter.In the two-stage segmentation,the symbol distance function is added to the network input as a priori information.The proposed loss function is used for training,and the gray matter and white matter are separated by the model.The results on the i Seg2017 dataset show that this method can effectively improve the accuracy of segmentation.