Learning Correspondence:One-shot Brain Anatomical Structure Segmentation Algorithm Based On Deep Learning

The accurate segmentation of medical images depicts the different anatomical structures and abnormal tissues of the human body.With sufficient labeled data,deep convolution neural networks have made a breakthrough in the research of medical image segmentation.However,obtaining a large number of annotations is laborintensive and easy to make mistakes.So,it is urgent to propose a deep-learning-based method,which only needs one or several labeled subjects to effectively solve the problem of data shortage.The brain anatomical structure segmentation with MRI is taken as an experiment in this thesis.The correspondence between one labeled volume and unlabeled volumes is learned through a deep network,and is applied to the segmentation space to get the segmentation of unlabeled volumes.Therefore,the aim of the algorithm proposed in this thesis is to obtain accurate correspondence,which can be used to propagate labels.The research contents and main contributions include the following aspects.Firstly,we propose a one-shot brain anatomical structure segmentation algorithm based on forward correspondence.The original deep-learning-based segmentation algorithm starts from fully convolution networks(FCN),which replaces the fully connection layers with convolution layers and realizes end-to-end dense segmentation.On the premise of uniform input distribution,the conventional segmentation network can segment the target area from the complex environment by learning the features hierarchically.Inspired by the deformable convolution,the one-shot segmentation algorithm does not directly output the probability map,but outputs the voxel-wise correspondence mapping between source data(labeled data x)and target data(unlabeled data y).Correspondence learns potential correlation mapping between images,which can directly apply to the segmentation space.This algorithm can reduce the demand for annotation data.Secondly,we propose a one-shot brain anatomical structure segmentation algorithm based on bidirectional correspondence.The condition of using one-shot algorithm to get accurate segmentation is that the deep network can learn smooth and precise correspondence mapping.The algorithm based on forward correspondence learning learns forward mapping from source domain x to target domain y,which is not unique.In order to make the network learning unique and precise mapping,the idea of CycleGAN is used to produce the bidirectional reversible mapping.Specifically,we add the backward correspondence mapping from the target domain y to the source domain x,which can constrain the learning of forward mapping.Thirdly,we propose a one-shot brain anatomical structure segmentation algorithm based on cycle-consistent correspondence.The one-shot segmentation algorithms based on forward correspondence learning and bidirectional correspondence learning only constrain the network in the image space and the transformation space.Differently,the segmentation algorithm based on the cyclic consistency adds the anatomical structure space to training and explores the consistency information in the three spaces.That consistency information is used as supervised signals to constrain correspondence learning.