Abdominal multi-organ segmentation in multiple MR sequences plays a crucial role in assisting clinical decision making. However, the annotation of multi-sequence MR images is time-consuming and labor-intensive, and the limited availability of labels constrains the advancement of multi-sequence MR abdominal segmentation. In this study, we propose a three-stage method for multi-sequence MR abdominal organ segmentation, incorporating unsupervised domain adaptation, semi-supervised learning, registration, and anatomical structure constraints. Specifically, in the first stage, pseudo-labels for T1-weighted (T1W) sequences are generated using labeled CT samples and unlabeled T1W sequence samples through unsupervised domain adaptation and semi-supervised learning. The second stage involves using a T1W-to-Multi-sequence label transfer module to share the pseudo-labels from T1W sequence with other sequences that have minimal positional differences, followed by training a multi-sequence segmentation network via semi-supervised learning. In the third stage, iterative training is conducted using the pseudo-labels generated by the segmentation model, with an Anatomy-aware module employed to enhance the accuracy of the pseudo labels. Our method achieved an average score of 81.60% and 89.83% for the organ DSC and NSD on the validation set and the average running time and area under GPU memory-time curve were 11.80 s and 26888 MB, respectively. Our code is available at https://github.com/Ho-Garfield/FLARE2024_he .