Off-label Information Based Medical Image Semantic Segmentation

With the advent of big data era,most industies produce millions of data every day.In the medical field,medical images are generated and accumulated rapidly.Semantic segmentation of medical images can help doctors detect the location and size of anatomical structures and aid in making therapeutic schedule.With the development of deep learning,deep convolutional network based medical image segmentation has become an important research field.Existing works focus on designing better network structures and better training methods to make the network learn features with better representation power.These works only use label information in the training phase.However,in medical image,there are more information other than label that can help the segmentation task.This thesis utilizes unlabeled data,context information,location and shape information to propose three novel works.Collecting and labeling medical image is difficult but there are much unlabeled data.To utilize unlabeled data,we propose a novel semi-supervised segmentation method called difference minimization network(DMNet).The segmentation network of DMNet is composed of a shared encoder branch and two decoder branches.DMNet utilizes unlabeled data by minimizing the difference between the segmentation masks generated by two decoder branches.To improve the segmentation performance of DMNet,we also adopt sharpen operation and adversarial training method.Compared to existing semi-supervised methods which are based on self-training or co-training,DMNet minimizes the difference between two masks generated by two decoder branches.Also,DMNet is an end-to-end method.Experiment result on a CT image dataset of kidney tumor and a MRI image dataset of brain tumor shows that DMNet can surpass baselines and achieve the state-of-the-art result.Pixels around the target tissue are sometimes more informative.So we propose a novel method called local-global context network(LGCNet)to utilize local and global context information.LGCNet is composed of a local context module(LCM)and a global context module(GCM).LCM divides the feature map into local regions and extracts local context information in each local region.GCM extracts some global context vectors which contain global context information,then distributes these global context vectors to each position of feature map.Compared to existing methods which focus mainly on global context information only,LGCNet extracts local and global context information at the same time.Experiment result on a CT image dataset of kidney tumor and a MRI image dataset of brain tumor shows that LGCNet can surpass existing methods and achieve the state-of-the-art result.Three-dimensional tissues often have relatively fixed location and shape.To utilize the location and shape information,we propose a novel graph convolution based threedimension tissue segmentation method called Graph-UNet.Graph-UNet constructs a adjacency matrix by statistical analysis of relationship of voxels with the same label in training set.Graph-UNet embeds the graph convolution into the segmentation network and uses the constructed adjacency matrix as the supervision information of the adjacency matrix generated by the segmentation network.Experiment result on a3 D tooth dataset shows that Graph-UNet can segment tooth well and the performance of Graph-UNet is better than existing methods.