Medical Image Segmentation Using Semi-supervised Generative Adversarial Network

Medical image segmentation is one of the important steps of computer-aided diagnosis.Compared with the traditional direct manual segmentation,the automatic segmentation method based on deep learning can accurately and quickly obtain lesion characteristic information,this method can not only reduce the working pressure of doctors,but also effectively improve the efficiency of clinical diagnosis.However,it is difficult to collect high-quality medical images,the manual annotation process is extremely time-consuming,and this process depend heavily on the subjective perceptions of the experts.The number of standardized data sets is small.The training requirement of deep learning model unable to meet.To address the above-mentioned limitations,this thesis conducted two major projects:1.The shape-aware U-net is proposed.The network includes shape-aware module and shape-aware pyramid pooling module.The shape-aware module can effectively extract the feature from the image and reduce the interference of unrelated regions.The shape-aware pyramid pooling module can deal with the confusion of lesion area boundary better.In the experiment of cervical spondylotic myelopathy segmentation,this model has achieved better performance than the existing segmentation network.2.The Consistent Perception Generative Adversarial Network is proposed.In the network,a similarity connection module is proposed.Non-local operations can be used to capture the dependencies between the lesion regions and improve the segmentation accuracy of the model.A pre-trained assistant network is employed to encourage the discriminator to learn meaningful feature representations.This method can solve the problem of discriminator forgetting in training.Finally,a consistent transformation strategy is adopted.This strategy makes full use of self-supervised information of the input and encourages the segmentation network to predict consistent results for unlabeled data.The re-designed discrimination method performs better in semi-supervised segmentation task.The proposed method was evaluated on the ATLAS data set.Compared with other advanced methods,this model achieves better segmentation performance.In the semi-supervised segmentation experiment,this model uses only two-fifths of the labeled data and achieves similar segmentation performance to U-NET.