Research On Semi-Supervised Medical Image Segmentation Algorithm With A Small Amount Of Labeled Data

With the rapid development of information technology,image segmentation task in medical field has attracted extensive attention and research.Medical image segmentation can predict the location of lesions or organs,assist doctor in diagnosis and analysis of diseases,and reduce the rate of missed diagnosis and misdiagnosis of diseases.However,the cost of labeling medical images is very high.Usually,there is only a small amount of labeled data,which leads to the low segmentation accuracy of supervised learning method in medical image segmentation task with less labeled data.In order to solve the problem of medical image segmentation task caused by a small amount of labeled data,semi-supervised learning is applied to the medical image segmentation task in this thesis.The specific research work is as follows:(1)Aiming at the scarcity of medical image labeled data,a semi-supervised medical image segmentation algorithm based on multi-transformation consistency regularization is proposed.Based on the idea of consistency regularization,complex spatial transformation,non-spatial transformation and disturbance are introduced.The features of unlabeled data are learned by minimizing the transformation consistency loss of multiple transformation branches,and the features of labeled data are learned by minimizing the supervision lossbetween label map and prediction map.Transformation consistency regularization introduces the unlabeled data features into the training of the model,and achieves good segmentation results on ISIC skin lesion dataset and Kvasir-SEG polyp dataset.(2)In order to improve the ability of the model to learn the features of unlabeled data,a dense contrastive learning algorithm is proposed.By introducing a dense projection head with convolutional layer,the global contrastive learning between single feature vector in the current contrastive learning research work is improved by dense contrastive learning between feature maps.Positive samples are composed of feature maps encoded by two transform views of the same image,and negative samples are composed of feature maps encoded by transform view of different images.Each feature vector in the feature map represents the feature of local pixels in the image.The experimental results show that dense contrastive learning can learn more pixel-level feature representations from unlabeled data.(3)In order to solve the problem of high pixel similarity between medical image data,a semi-supervised image segmentation algorithm based on image patch dense contrastive learning is proposed on the basis of dense contrastive learning.The algorithm is divided into pre-training stage and fine-tuning stage.In the pre-training stage,the unlabeled data is divided into multiple image patches,and two branches of image patch contrastive learning and dense contrastive learning are constructed.The encoding network is trained by minimizing the weighted sum of the loss of the two branches.In the fine-tuning stage,the segmentation network loads the encoding network weight parameters generated by pre-training stage,and performs supervised training on the labeled data.The experimental results on Montgomery dataset,JSRT dataset and ISIC skin lesion dataset show that the algorithm can effectively learn pixel features of unlabeled data and improve the accuracy of medical image segmentation.