Research On Domain Adaptation For Multimodal Medical Image Analysis

In recent years,with the rapid development of computer technology,especially the introduction of deep learning,the field of medical image analysis has become more and more automated and intelligent.The applications of these technologies in the medical field include disease diagnosis and classification,lesion segmentation and detection,etc.As we all know,the performance of deep neural network depends to a large extent on the quality of training set data.However,the labeling of medical image data often requires a lot of time and cost of professionals.At the same time,medical image data sometimes involves data privacy,etc.problem,which makes the training of neural networks difficult to carry out.Therefore,it is of great significance to study domain adaptation methods for multimodal medical image analysis.On the basis of domestic and foreign research,this thesis conducts a series of researches on domain adaptation methods suitable for medical image analysis,mainly including the following three points:Domain-adaptive methods for medical image segmentation.This thesis proposes a domain adaptation method based on self-integrating networks.The basic network adopts U-Net,which introduces an attention mechanism to identify image features,and further uses these features to guide the calculation of the consistency loss of the target domain.Guided by the consistency loss,the student network can learn from the output of the teacher network,while the student network is also affected by its source domain image segmentation loss.The parameters of the teacher network are updated by exponential moving average.In this thesis,public fundus image datasets and cell image datasets are used to conduct experiments,and the experimental results confirm the effectiveness of adaptive methods in this field.A passive domain data domain adaptation method for medical image segmentation.This thesis proposes a domain adaptation method,which does not require the data of the source domain image,but uses the characteristics of the discriminator in the generative adversarial network to identify the image features,so that the discriminator learns the characteristics of the target domain,and then the discriminator's The parameters are transferred into the downsampling part of the segmentation network.This thesis uses public medical image datasets to conduct experiments.The experimental results show that in the absence of source domain data,this method can effectively reduce the loss caused by the segmentation model during the migration process,and has a certain degree of advancement.A passive domain data domain adaptation method for medical image classification.The method is based on the generative adversarial network,and uses the residual network as the classifier,uses the random algorithm to obtain random noise,and then inputs the label into the generator according to the category of the target domain image.The input of the discriminator in the generative adversarial network is the generated image and the real image,and the authenticity is judged.The generated target domain-style images and pseudo-labels are then input into the classifier to guide the training of the classifier.This thesis selects three public datasets on the classification of the degree of diabetic retinopathy.The experimental results show that this method solves the problem of low accuracy after model migration caused by passive domain data to a certain extent,and has certain practical value.