Research On Medical Image Segmentation Algorithm Based On Consistency Label-transferring Learning

Like image classification,image segmentation has become one of the most important fields in medical imaging researches.The aim of image segmentation is extracting and segmenting image regions or features with specific meanings,which is the foundation of medical diagnosis and treatment,imageology and pathology research.Therefore,accurately and steadily segmenting human organs or lesion areas from medical images plays an important role in the diagnosis and treatment of diseases and many other clinical medical problems.However,due to the particularity and complexity of clinical medical imaging,it is necessary to solve the problems of inter-individual differences and irregularities in the process of image segmentation.Therefore,proposing an effective medical image segmentation algorithm has great significance to improve the effect of medical image segmentation.Automatic lesion segmentation of medical image is one of the key technologies of computer-aided diagnosis(CAD),which has been widely used in clinical medicine.It has rapidly replaced the inefficient traditional manual segmentation of medical image.In recent years,due to the emergence of deep learning technologies such as convolutional neural networks,the performance of automatic medical image segmentation on specific issues such as image segmentation of cardiac magnetic resonance and lung nodule detection has reached or exceeded the level of professional physicians.However,most of these methods are usually trained by supervised learning,and a large amount of high-quality labeled training data is required during the training stage.The data annotation of medical images is a very time-consuming process.For example,accurate annotations of radiological images must be manually annotated by experienced radiologists and then carefully checked by other experts.In recent years,related researchers have proposed many medical image segmentation algorithms to reduce the dependence on labeled data in the process of medical image segmentation.The main methods can be divided into the following three types: 1)Self-supervised learning algorithm: Through the structure or characteristics of a large amount of unlabeled data,automatically constructing labels to train the network model,and then transferring knowledge on this model;2)Weakly-supervised learning algorithm: Without pixel-level segmentation tags,only using hand-drawn image contours,bounding boxes,or image-level category tags for medical image segmentation;3)Semi-supervised learning algorithm: Using a small amount of labeled data and a large amount of unlabeled data for model training,and obtaining the results of image segmentation with high-performance.In most medical image segmentation problems,several data contain segmentation label,so the semi-supervised learning algorithm that makes full use of the existing segmentation and labeled data has the most extensive application.Therefore,the research of this article mainly focuses on the segmentation of medical images by semi-supervised learning.The challenge is how to take advantage of a large amount of unlabeled data to improve segmentation performance and obtain an accurate lesion mask(Mask).In this article,we propose a semi-supervised medical image segmentation algorithm based on Consistency label-transferring learning.For a small amount of medical image data with segmentation labels,the cross-entropy loss between the predicted lesion mask by the prediction model with the attention mechanism and the real segmentation label is optimized.For the medical image data without segmentation labels,it is based on the image segmentation technology of Cycle GAN to realize the transformation of data between different classes,so as to maximize the identification of the location of the lesion area that needs to be changed during the class transformation process.In the training process of Cycle GAN,only medical imaging data with class annotations(disease/non-disease)is used.The twoway class transformation model is established between disease data and non-disease data,and the attention mechanism is used to detect the two the most discriminative semantic feature between each class,so as to identify the diseased area.Finally,we can also use newly generated disease images and corresponding lesion masks as new training data with segmentation labels to further improve the performance of semisupervised segmentation.We use the proposed model to perform several experiments on the Bra TS and ISIC datasets.The experimental results show the effectiveness of the model on the medical image segmentation problem.On the data with only a small number of segmentation labels,it greatly exceeds the method by supervised learning.