The precise segmentation of breast tumors is of great significance to the clinical diagnosis of breast cancer. Automatic breast ultrasound (ABUS) is a low-cost medical imaging method. However, one breast scan will produce hundreds of three-dimensional ABUS slices. Manual screening for abnormalities is not only time-consuming, but also heavily relies on the doctor's experience. In this study, a new network with Orthogonal Positioning Attention Deep Supervision Network (OPADSN) was proposed for breast cancer tumor segmentation. OPADSN adopts an orthogonal strategy, sending the input to the horizontal axis and the vertical axis to output features in two independent directions. Orthogonal strategy not only retains more location information, but also helps the model locate and identify the target of interest more accurately. In addition, this article introduces the Original Resolution Subnetwork Network (ORSNet) into the context module for the first time to form the Original Resolution Subnetwork Enhancement Context Attention (ORS-CA). ORSNet can solve the problem of edge blur caused by downsampling, thereby generating high-resolution spatial features. Repeated use of the ORS-CA module can establish the connection between the breast tumor area and the boundary clues, making the tumor contour easier to judge. Due to this iterative interaction mechanism between regions and borders, OPADSN can correct different regions in the prediction results. This article conducted an experiment on a dataset of 783 two-dimensional breast ultrasound images. The experimental results show that the OLADSN method achieves 93.1% segmentation accuracy in Mean Dice, which is a satisfactory result.