Generative feature style augmentation for domain generalization in medical image segmentation

Abstract

Although learning-based models have achieved tremendous success in medical image segmentation for independent and identical distributed data, model performance often deteriorates for out-of-distribution data. Training a model for each domain requires extra time and computing resources and increases the annotation burden of physicians. It is hence more practical to generalize the segmentation model trained using a single source domain. In this work, we model domain-level feature style with a flexible probabilistic block, that is framework-agnostic and can be integrated into an arbitrary segmentation network to enhance the model generality on unseen datasets. Specifically, we employ a variational auto-encoder to learn the feature style representations, enabling the generation of diverse feature styles through sampling from a prior distribution. During inference, we replace the target feature style with that of the source domain using a linear transformation. We compare our method with five state-of-the-art domain generalization (DG) methods using prostate MRI data from six centers and spinal cord MRI data from four sites. Evaluation with Dice similarity coefficient score and 95th percentile Hausdorff distance demonstrates that our method achieves superior improvement in model generalizability over other DG models.