Multi-modality Medical Image Segmentation with Unsupervised Domain Adaptation

Abstract

Advances in medical imaging have greatly aided in providing accurate and fast medical diagnosis, followed by recent deep learning developments enabling the efficient and cost-effective analysis of medical images. Among different image processing tasks, medical segmentation is one of the most crucial aspects because it provides the class, location, size, and shape of the subject of interest, which is invaluable and essential for diagnostics. Nevertheless, acquiring annotations for training data usually requires expensive manpower and specialised expertise, making supervised training difficult. To overcome these problems, unsupervised domain adaptation (UDA) has been adopted to bridge knowledge between different domains. Despite the appearance dissimilarities of different modalities such as MRI and CT, researchers have concluded that structural features of the same anatomy are universal across modalities, which unfolded the study of multi-modality image segmentation with UDA methods. The traditional UDA research tackled the domain shift problem by minimising the distance of the source and target distributions in latent spaces with the help of advanced mathematics. However, with the recent development of the generative adversarial network (GAN), the adversarial UDA methods have shown outstanding performance by producing synthetic images to mitigate the domain gap in training a segmentation network for the target domain. Most existing studies focus on modifying the network architecture, but few investigate the generative adversarial training strategy. Inspired by the recent success of state-of-the-art data augmentation techniques in classification tasks, we designed a novel mix-up strategy to assist GAN training for the better synthesis of structural details, consequently leading to better segmentation results. In this thesis, we propose SynthMix, an add-on module with a natural yet effective training policy that can promote synthetic quality without altering the network architecture. SynthMix is a mix-up synthesis scheme designed for integration with the adversarial logic of GAN networks. Traditional GAN approaches judge an image as a whole which could be easily dominated by discriminative features, resulting in little improvement of delicate structures in synthesis. In contrast, SynthMix uses the data augmentation technique to reinforce detail transformation at local regions. Specifically, it coherently mixes up aligned images of real and synthetic samples at local regions to stimulate the generation of fine-grained features examined by an associated inspector for domain-specific details. We evaluated our method on two segmentation benchmarks among three publicly available datasets. Our method showed a significant performance gain compared with existing state-of-the-art approaches.