Segmentation of liver tumors plays an important role in the subsequent treatment of liver cancer. At present, the mainstream method is the fully supervised method based on deep learning, which requires medical experts to manually label a large number of pixel level labels for training, resulting in high time and labor cost. In this article, we focus on using bounding boxes as weak label to complete the segmentation task. It can be roughly divided into two steps. The first step is to use region mining technology to obtain pixel level labels from the bounding box. The second step uses pixel level labels to train the semantic segmentation network to obtain segmentation results. In the whole task, the quality of pixel level labels obtained from bounding boxes plays an important role in the performance of segmentation results. Therefore, our goal is to generate high-quality pixel level labels. Aiming at the problem that the current region mining method based on classification network is inaccurate and incomplete in object location, we use the Adversarial Complementary Learning module to make the network pay attention to more complete objects. We conduct analysis to validate the proposed method and show that our approach performs is comparable to that of the fully supervised method.