Single-source Domain Generalization in Deep Learning Segmentation via Lipschitz Regularization

Mazlum Ferhat Arslan, Weihong Guo, Shuo Li

Abstract: Deep learning methods have proven useful in medical image segmentation when deployed on independent and identically distributed (iid) data.However, their effectiveness in generalizing to previously unseen domains, where data may deviate from the iid assumption, remains an open problem.In this paper, we consider the single-source domain generalization scenario where models are trained on data from a single domain and are expected to be robust under domain shifts.Our approach focuses on leveraging the spectral properties of images to enhance generalization performance. Specifically, we argue that the high frequency regime contains domain-specific information in the form of device-specific noise and exemplify this case via data from multiple domains. Overcoming this challenge is non-trivial since crucial segmentation information such as edges is also encoded in this regime. We propose a simple regularization method, Lipschitz regularization via frequency spectrum (LRFS), that limits the sensitivity of a model’s latent representations to the high frequency components in the source domain while encouraging the sensitivity to middle frequency components. This regularization approach frames the problem as approximating and controlling the Lipschitz constant for high frequency components. LRFS can be seamlessly integrated into existing approaches. Our experimental results indicate that LRFS can significantly improve the generalization performance of a variety of models.