Research On Medical Image Segmentation Based On Deep Convolution Network

The purpose of medical image segmentation is to extract relevant features from medical images.Medical image features are complex and lack simple linear features.Moreover,the segmentation results are affected by some difficult issues,such as gray unevenness,volume effect,approximation of gray between different soft tissues,artifacts.Therefore,medical image segmentation has always been a challenging research topic.Recently,the commonly used deep convolutional neural network is called U-net,which only use long jump connections to fuse output features to recover lost spatial information in downsampling.This structure will lose some of the shallow space information and also limit the further deepening of the network due to the gradient disappearance problem.To address this issue,a deep fully residual convolutional neural network that combines the U-net with the ResNet for medical image segmentation is proposed.In this paper,the short skip connection(residual learning)in ResNet is added to the structure of U-net,combined with the identity mapping feature in residual learning,which not only solves the gradient problem,but also improves the performance of the network.Although the depth of the network is deepened,the model proposed in this paper has fewer parameters than U-net.we evaluated the performance of the proposed model and other state-of-the-art models on the Electron Microscopy(EM)images dataset and the Computed Tomography(CT)images dataset.The results show that the segmentation accuracy of our method is better than most other methods.Although the above proposed model has a certain improvement in the segmentation precision,it only predicts the class of each pixel independently,i.e.,the interrelationship between pixels is easily ignored so that the continuity of the segmentation is not good enough.In order to further improve the segmentation result and segmentation continuity of the model,this paper introduces the Generative Adversarial Networks(GANs)into the previously proposed model.On the one hand,the characteristics of global prediction of GANs are used to improve the segmentation continuity of the model.On the other hand,the adversarial network is used to improve the robustness of the model under small-scale data and prevent over-fitting.In this work,the model proposed in the third chapter is used as the generation model,and the prediction of the label class is generated based on the input RGB image.The regression loss is calculated pixel by pixel.The discriminative network in the model performs high-order regular statistics on the difference between the prediction image and the label image,and it provides a self-learning global loss statistical method for generating the model.The results show that the segmentation-adversarial model has improved the overall segmentation continuity and segmentation accuracy on the EM datasets and CT dataset.