Gastric cancer has been one of the leading causes of cancer death. To assist doctors on diagnosis and treatment planning of gastric cancer, an accurate and automatic segmentation of gastric tumor method is very necessary for clinical practices. In this paper, we develop an improved U-Net called hybrid blocks network (HBNet) to automatically segment gastric tumor. In contrast to the standard U-Net, our proposed network only has one down-sampling operation, which further improves the performance on segmentation of small target tumor. Meanwhile, we innovatively devise a combination of squeeze-excitation residual (SERes) block and dense atrous global convolution (DAGC) block instead of the original convolution and pooling operations. Both high-level and low-level feature information of the tumor is effectively extracted. We evaluate the performance of HBNet on a self-collected ordinary CT images dataset from three medical centers. Our experiments demonstrate that the proposed network achieves quite favorable segmentation performance compared with the standard U-Net network and other state-of-the-art segmentation neural networks.