Deployment of medical image segmentation models remains a challenge largely because of changes in the distribution of image domains for differences in acquisition protocols, devices, and patient demographics. These factors limit model generalization on new unseen datasets, termed out-of-distribution (OOD) datasets. To address this, we introduce “DomainFlow”, a novel single-source domain generalization (DG) method for coronary vessels segmentation in x-ray angiography. In our method, we hypothesize that training a model to predict connectivity masks instead of conventional binary masks preserves model performance against domain shift, by capturing domain-invariant spatial relationships and structural consistency between vessels, reducing the impact of domain-specific noise. To further encourage DG through model robustness, our model learns a Gaussian posterior at the latent space, which is further refined through a supervised prior, therefore, exposing the model to a broader range of data variations during training. Our extensive evaluation on three x-ray angiography datasets shows that DomainFlow outperforms the state-of-the-art methods with 0.85–3.85% and 2.86–5.09% Dice score improvements in OOD and in-domain datasets, respectively, thus, demonstrating its potential for real-time clinical applications.