Deep learning methods have been demonstrated effective in medical image segmentation tasks. The results are affected by data imbalance problems. The inter-class imbalance is often considered, while the intra-class imbalance is not. The intra-class imbalance usually occurs in medical images due to external influences such as noise interference and changes in camera angle, resulting in insufficient discriminative representations within classes. Deep learning methods are easy to segment regions without complex textures and varied appearances. They are susceptible to the intra-class imbalance problem in medical images. In this paper, we propose a two-stage global-local framework to solve the intra-class imbalance problem and increase segmentation accuracy. The framework consists of (1) an auxiliary task network(ATN), (2) a local patch network(LPN), and (3) a fusion module. The ATN has a shared encoder and two separate decoders that perform global segmentation and key points localization. The key points guide to generating the fuzzy patches for the LPN. The LPN focuses on challenging patches to get a more accurate result. The fusion module generates the final output according to the global and local segmentation results. Furthermore, we have performed experiments on a private iris dataset with 290 images and a public CAMUS dataset with 1800 images. Our method achieves an IoU of 0.9280 on the iris dataset and an IoU of 0.8511 on the CAMUS dataset. The results on both datasets show that our method achieves superior performance over U-Net, CE-Net, and U-Net++.