Obtaining labeled data from medical images is very expensive and labor intensive. We consider the problem of unsupervised domain adaptation (UDA) in cross-modality medical image segmentation, aiming to perform segmentation on the unannotated target domain with the help of labeled source domain. We present a framework that combines unsupervised domain adaptation, registration, and pseudo-label learning to effectively and flexibly adapt to multiple target domains. We introduce a novel image translation method based on anatomy space and a novel operation of matching and registration to improve pseudo-labels, which effectively mitigates the large cross-modality domain gap. Experiments demonstrate that our method achieves the average Dice Similarity Coefficient of 0.6053 and Normalized Surface Dice of 0.6462 on 13 abdominal organ segmentation tasks. Moreover, it significantly improves the inference speed, with an average running time of 20.7 s, and uses only an average of 1107332.6 MB of total GPU memory on final testing set.