Abstract: We propose an unsupervised unimodal registration method based on Wasserstein Gan. Different from the existing registration methods based on deep learning, the proposed method can finish training without ground truth or preset similarity metrics. The network is composed of a generation network and a discrimination network. The generation network extracts the potential deformation fields between fixed images (positive images) and moving images, and predicts the registered images (negative images) by interpolation; the discrimination network then judges the similarity between the positive images and negative images that are input alternately, and feeds back the judgment result as a loss function to drive the network parameter update. Finally, through adversarial training, the registration image generated by the generation network deceives the discrimination network and the network converges. In the experiment, we randomly selected 30 cases of LPBA40 brain dataset, 25 cases of EMPIRE10 lung dataset and 15 cases of ACDC heart dataset as the training datasets, with 10 cases of LPBA40 brain dataset, 5 cases of EMPIRE10 lung dataset and 5 cases of ACDC heart dataset as the test datasets. The results of registration were compared with those obtained using Affine algorithm, Demons algorithm, SyN algorithm and VoxelMorph algorithm. The DICE coefficient (DSC) and the normalized correlation coefficient (NCC) evaluation index of the proposed algorithm were the highest, indicating a better registration accuracy of our method than Affine algorithm, Demons algorithm, SyN algorithm and the current unsupervised SOTA algorithm VoxelMorph.

Key words: unsupervised; Wasserstein Gan; unimodal registration; adversarial training