Medical Image Segmentation Method Based On Slice Context Information

Medical image segmentation is an important basis for auxiliary medical diagnosis.Medical images are different from natural images.There is no strong correlation between natural images.In contrast,due to its imaging technology of medical image,there is a strong correlation between various slices of medical images,so that its features exist not only in the slice,but also between adjacent slices.When segmenting medical images,it is necessary to extract not only the two-dimensional features in the slice,but also the sequence features between adjacent slices,so as to segment the organs in each slice accurately.Therefore,how to extract intra-slice two-dimensional features and inter-slice sequence features and fuse them together is the work of this paper.2D convolutional neural network has made remarkable achievements in natural images,and the two-dimensional feature extraction method of medical images is transferred from it.Although 2D network has strong two-dimensional feature extraction ability,it can not extract features in other dimensions in space,which are part of human organs and have an important impact on the segmentation results.3D convolution neural network mainly processes 3D image data.Referring to 2D convolution network method,a variety of spatial features including two-dimensional features in slices and sequence features between slices are extracted isotropically in all directions of 3D space.However,medical images are anisotropic data stacked by a series of slices,that is,there is a large gap between the pixel spacing in slices and between slices.The direct application of 3D convolution network can not adapt to the characteristics.On the other hand,the amount of 3D network training parameters is huge,the space occupation is high,and it is easy to over fit.Although the existing 3D medical image segmentation methods based on multi view fusion can take advantage of the advantages of 2D Network and the amount of parameters is not too large,the feature extraction process on each view(coronal plane,sagittal plane and transverse plane)is exactly the same,and they also do not pay attention to the anisotropic characteristics of medical data,and there are some irrelevant features in the image,It does not filter the irrelevant features in the shallow features,which affects the segmentation effect.In addition,the influence of respiratory movement on abdominal organs leads to organ edge blur in slices and different organ contour sizes between different slices.2D method can not deal with these two problems,resulting in these problems still exist in the output probability map,but the existing fusion methods do not deal with this problem alone.To solve the above problems,this paper proposes a fusion method based on single view,which only uses the horizontal plane slice of human organs.This method includes two parts:intra-slice two-dimensional feature extraction network H-Felay Net and inter-slice sequential feature extraction network Mac Net.Because the skip connection between U-Net layers simply integrates the shallow features and introduces more irrelevant features into the high layer of the network,this paper designs H-Delay Net network,which uses the channel attention mechanism to assign different weights to different feature channels to suppress irrelevant feature channels,and uses the spatial attention mechanism to focus the network on important features,weaken the relationship between the target organ and irrelevant features to reduce their interference to the training process.On the other hand,this paper designs a Mac Net network.Before using Conv LSTM to extract the inter-slice sequence features,the fuzzy edges are mapped to the high-dimensional space by using the multi-fuzzy-degree module to find a clear boundary in the abstract features.On the other hand,the atrous spatial pyramid pooling module is used to apply the different field convolutions to different slices to extract the corresponding features in different density regions,so Conv LSTM can more accurately learn the information of organ changes between adjacent slices.Finally,by the maximizing fusion method,the segmentation results of the two networks are complementary,and the advantages of the two networks are combined,so that the fused segmentation results surpass any previous network.This method is evaluated on the PROMISE12 dataset and the LITS dataset.The main evaluation index is Dice.Our method exceeds the classical segmentation method on PROMISE12 and obtains better segmentation results.H-Delay Net shows higher accuracy and better universality in LITS.