Weakly-supervised medical image segmentation with only image-level annotation is particularly challenging to infer precise pixel-wise predictions. Existing works are usually highly restricted by the assumption that the medical images for training and testing are under the same distribution. However, a robust weakly-supervised segmentation model needs to show accurate inference on medical images from unseen distributions. Different feature distributions can lead to a dramatic shift in the feature activation and class activation map (CAM), which in turn leads to the degradation of pseudo labels. In this paper, we aim to learn generalizable weakly-supervised medical image segmentation by focusing on enhancing the domain invariance for pseudo labels. A novel domain-invariant CAM learning scheme (D-CAM) is proposed, in which the content and style are decoupled during training. By inferring domain-invariant pseudo labels, the supervision of a segmentation model is more generalizable to different target domains. Extensive experiments under multiple generalized medical image segmentation settings show the state-of-the-art performance of our D-CAM. Source code is available at https://github.com/JingjunYi/D-CAM .