A medical unsupervised domain adaptation framework based on Fourier transform image translation and multi-model ensemble self-training strategy.

Abstract

Well-established segmentation models will suffer performance degradation when deployed on data with heterogeneous features, especially in the field of medical image analysis. Although researchers have proposed many approaches to address this problem in recent years, most of them are feature-adaptation-based adversarial networks, the problems such as training instability often arise in adversarial training. To ameliorate this challenge and improve the robustness of processing data with different distributions, we propose a novel unsupervised domain adaptation framework for cross-domain medical image segmentation. In our proposed approach, Fourier transform guided images translation and multi-model ensemble self-training are integrated into a unified framework. First, after Fourier transform, the amplitude spectrum of source image is replaced with that of target image, and reconstructed by the inverse Fourier transform. Second, we augment target dataset with the synthetic cross-domain images, performing supervised learning using the original source set labels while implementing regularization by entropy minimization on predictions of unlabeled target data. We employ several segmentation networks with different hyperparameters simultaneously, pseudo-labels are generated by averaging their outputs and comparing to confidence threshold, and gradually optimize the quality of pseudo-labels through multiple rounds self-training. We employed our framework to two liver CT datasets for bidirectional adaptation experiments. In both experiments, compared to the segmentation network without domain alignment, dice similarity coefficient (DSC) increased by nearly 34% and average symmetric surface distance (ASSD) decreased by about 10. The DSC values were also improved by 10.8% and 6.7%, respectively, compared to the existing model. We propose a Fourier transform-based UDA framework, the experimental results and comparisons demonstrate that the proposed method can effectively diminish the performance degradation caused by domain shift and performs best on the cross-domain segmentation tasks. Our proposed multi-model ensemble training strategy can also improve the robustness of the segmentation system.