Aiming to address the issue of inadequate and singular feature extraction in instance segmentation and drawing inspiration from the design concept of RefineNet and relative multi-scale attention, we propose a novel mask branch module known as MSAR (Multi-Scale Attention Refinenet). Initially, features extracted by the backbone network or FPN at different scales are utilized as input. Subsequently, high-level and low-level feature maps are generated through relative multi-scale attention. These generated feature maps are then fused using the RefineNet module to produce the final mask. Testing on the COCO dataset reveals that when MSAR is incorporated into BlendMask, the mask AP can reach 35.55%, indicating that MSAR can enhance the accuracy of masks generated by instance segmentation networks.