Aegis: A domain generalization framework for medical image segmentation by mitigating feature misalignment

Domain shift caused by variations in data acquisition significantly impedes the deployment of medical image segmentation models in clinical settings. Domain generalization aims to mitigate performance degradation induced by domain shift by training a model using source domain data and generalize well to unseen target domain. In this work, we have an interesting observation: domain shift results in significantly different activation patterns across domains even they have semantically identical input. This cross-domain “feature misalignment” phenomenon motivates us to develop a hypothesis: mitigating cross-domain feature misalignment may enhance domain generalization. To this end, we propose a framework called Aegis, which employs style augmentation to generate augmented image features that simulate domain shift. Subsequently, we introduce a dual attention-guided feature calibration (DAFC) module to facilitate feature interaction between source and augmented images, thereby establishing an implicit alignment constraint within the shared feature space. Furthermore, we propose an uncertainty-guided feature alignment (UFA) loss, which quantifies segmentation discrepancies caused by domain shift and incorporates an uncertainty-weighting mechanism to enhance the alignment of hard-to-classify pixel regions. These components work in synergy to effectively mitigate cross-domain feature misalignment, promote robust feature alignment, and ultimately improve cross-domain generalization. Extensive experiments conducted on three widely used benchmarks demonstrate that the proposed framework significantly outperforms existing methods in domain generalization. Code is available at https://github.com/Zerua-bit/Aegis.