Predicting Cardiovascular and Cerebrovascular Events Based on Instantaneous High-Order Singular Entropy and Deep Belief Network.
Shao, Shiliang; Wang, Ting; Mumtaz, Asad; Song, Chunhe; Yao, Chen
Most medical image segmentation models often assume identical data distributions in the source and target domains, limiting their deployment in different clinical centers, and the performance usually drops dramatically when applied to an unseen target domain. This paper tackles the challenging problem of generalized medical image segmentation on unseen domain data. We discover that the performance degradation is mainly due to the inter-domain distribution shift and the intra-domain appearance variation, and the Extrinsic Attention (EA) and Intrinsic Attention (IA) modules are designed to address these two issues, respectively. Specifically, the EA overcomes the shift between different domains by enriching image features with knowledge from multi-source domains. It dynamically maintains a domain knowledge bank, and then performs feature enhancement based on feature similarity. The IA alleviates the problem of large variation in the target domain by exploring the relationship between pixels and regions within a single image. It conducts pixel-region relation modeling from the coarse segmentation map to facilitate pixel context learning and finally obtain a more accurate segmentation map. To validate the model effectiveness, extensive experiments have been conducted on various medical image segmentation benchmark datasets. The results demonstrate the superiority and feasibility of our method compared to previous domain generalization techniques. Moreover, our model can be extended to various medical image segmentation tasks, providing potential benefits for a wide range of medical image segmentation applications.
This paper develops a novel domain generalization model for medical image segmentation by addressing both the inter-domain distribution shift and the intra-domain appearance variation, and demonstrates its effectiveness through experiments on public datasets.
To achieve generalization to unseen domains, a domain-aware dual attention network is developed for medical image segmentation. The dual attention modules are designed to account for the inter-domain distribution shift and intra-domain appearance variation.