DGSA-Net: A Domain-Generalized Structural Attention Network for Single-Source Cross-Dataset Medical Image Segmentation

Abstract

Medical image segmentation models often perform well on individual datasets but fail to generalize across heterogeneous imaging domains due to domain shifts caused by variations in acquisition protocols, resolution, and appearance. To address this challenge, we propose DGSA-Net, a Domain-Generalized Structural Attention Network that emphasizes structural consistency and boundary information for robust Cross-Domain segmentation. The proposed framework introduces a Structural Attention Refinement Module (SARM) to enhance boundary-aware features and suppress irrelevant background responses. It further incorporates hierarchical cross-scale feature integration and boundary-aware progressive decoding to improve lesion localization and contour delineation. A structure-constrained hybrid loss is used to jointly enforce region, pixel-level, and boundary consistency. The effectiveness of the proposed framework is evaluated on three heterogeneous benchmark datasets, including dermoscopic, CT, and chest X-ray datasets, demonstrating that DGSA-Net consistently outperforms state-of-the-art methods. The proposed framework DGSA-Net achieves Dice scores of 0.918, 0.892, and 0.945 on dermoscopy, CT, and chest X-ray datasets, respectively, under in-domain evaluation. For same-objective cross-dataset evaluation, it achieves Dice scores of 0.874 on ISIC 2018 to ISIC 2017 and 0.926 on Shenzhen to Montgomery. As an auxiliary cross-task transfer experiment, LIDC-IDRI to COVID-19 CT-Seg achieves a Dice score of 0.836. Moreover, the proposed model achieves the lowest boundary error (HD95 of 6.49) and reduces the generalization gap to 4.0%. These results highlight the effectiveness of DGSA-Net in learning structurally consistent and domain-robust representations, providing a reliable solution for Cross-Domain medical image segmentation without requiring target-domain adaptation.