Cross-Domain Image Segmentation Algorithms Based On Cycle-Consistency And Representation Learning

In medical image assisted diagnosis,the accurate segmentation of lesion regions is crucial for physicians to develop appropriate treatment plan.With the advancement of medical imaging technology,the volume of image data has increased significantly,making traditional manual film-reading method insufficient to meet the demand.In recent years,automatic medical image segmentation techniques based on deep convolutional neural networks have achieved remarkable success,even comparable to professional clinicians.Typically,deep neural networks are trained and tested under the assumption that all samples are from the same probability distribution.However,due to the variations in scanning devices,clinical centers,imaging parameters,etc.,the acquired medical images exhibit significant differences in brightness,contrast,noise levels,and other aspects.These datasets from different devices are defined as different domains.It has been observed that the performance of a well-trained model on one domain drops significantly when testing on samples from different domains,i.e.,the domain shift problem.To solve the domain shift problem,researchers have proposed crossdomain techniques,such as domain adaptation and domain generalization.The goal of this technique is to improve the generalization performance of deep learning models on unlabeled target domains when only the labelled source domains are available.With domain adaptation and domain generalization methods,cross-domain techniques can help models overcome the discrepancies in data distribution between different domains,and improve their generalization ability.In this thesis,two types of cross-domain techniques,domain adaptation and domain generalization,are explored to alleviate the performance degradation caused by domain shift problem,based on retinal OCT images from three different devices.The main works and innovations of the thesis are as follows:1)A single-source domain adaptation image segmentation algorithm based on adversarial transfer learning is proposed.First,an image style transfer module is designed based on a cycle generation adversarial network to reduce the difference in appearance between source and target domain samples.Second,a feature transfer module is applied to extract the source and target features in an adversarial learning manner to obtain similar feature representations(i.e.,domain-invariant features)of both domains.Finally,a new automatic pseudo-label selection method based on discrepancy and similarity methods is proposed,which introduces the target domain samples into the source domain for iterative training to facilitate the network to capture the target domain-specific features during the training process.2)A domain generalization image segmentation algorithm based on style and semantic consistency is proposed.First,the source domain samples are augmented using existing data manipulation techniques to generate multiple pseudo-target domains with different styles.Then,a style unification strategy is proposed to transfer all generated pseudo-target domains into the source domain space to learn a mapping model from the target domain to the source domain.This helps prevent the propagation of representation discrepancies to the subsequent segmentation networks,thus improving model generalization and ensuring segmentation accuracy on the source domain.By introducing the style-invariance of the instance normalization,the style unification network is also robust to unseen target domains.Finally,a two-branch segmentation network with semantic consistency feature attention is proposed to capture domain-invariant features from a single source domain while suppressing the expression of domain-variant information.