Accurate and efficient medical image segmentation is the key step for visual analysis and auxiliary diagnosis of diseases. However, medical image segmentation has not reached the practical clinical standard due to the uncertainty and the feature drift that affects the performance of supervised learning. Therefore, the paper considers the fuzzy problem of uncertain feature assignment in semantic segmentation and proposes a novel Uncertainty-guided Feature Learning Network (UFL-Net) for medical image segmentation. The key idea lies in its feature learning module, which can effectively quantify and utilize feature uncertainty. UFL-Net can adaptively guide features according to feature quantization outcomes and adjacent features to mitigate deviations in learning targets attributable to feature uncertainty. We propose an edge-aware prior enhancement module and a feature fusion module to improve the segmentation results further. Experiments on various datasets substantiate that UFL-Net achieves superior segmentation capabilities. Additionally, the uncertainty quantification and the uncertainty-guided feature learning module developed in the paper demonstrate versatility, easily applicable to diverse datasets, or networks.