Unsupervised Domain Adaptation For Medical Image Segmentation

Medical image segmentation is a key task in the application of computer-aided diagnosis and treatment.With the rapid development of deep learning technology,medical image segmentation method based on deep learning has achieved remarkable results.However,in the medical image segmentation task,due to the differences of data acquisition equipment and standards,images from different centers and devices often have serious domain drift problems,which leads to the performance degradation of segmentation model on different domain data.In order to solve this problem,unsupervised domain adaptive research in medical imaging is gradually emerging.Although recent unsupervised domain adaptation work has achieved some results in the field of medical image analysis,there are still great challenges.For example,image translation technology(image style transfer)is often used in unsupervised domain adaptation to reduce the difference between two domains.However,image translation may change the semantic information of the original image,resulting in the deterioration of network domain adaptive performance.In order to deal with this challenge,the existing technologies either separate the image translation model from the domain adaptive segmentation model,so that the image translation module does not affect the training of the domain adaptive segmentation model,or correct it through a costly two-way learning process.In this study,we propose a new unsupervised domain adaptive method,which can effectively correct image translation errors and is simpler than other correction methods.Firstly,a simple and effective baseline model combining image alignment and feature alignment is proposed.Based on the baseline model,we propose an effective self improving domain adaptation(SIDA)method.This method introduces two self-monitoring tasks(image translation degree prediction task and contrast learning task)to ensure the semantic consistency between the translated image and its feature representation,and improve the effectiveness and robustness of image translation module for segmentation network.In this paper,the unsupervised domain adaptive pancreas segmentation(CT-MRI)task is fully studied and analyzed.The results show that the proposed SIDA is effective and superior to other advanced algorithms.