AISA-DG: Automatic Implicit Style-Augmented Domain Generalization on Optic Disc/Cup Segmentation

Convolutional neural networks have found widespread applications in the field of medical image segmentation, yielding remarkable results. However, in clinical practice, medical images frequently display variations in appearance due to various factors, such as variations in scanner vendors and image quality. Consequently, a noteworthy challenge arises from the existence of domain gaps between training (source domain) and test data (target domain). In response to this challenge, we introduce an innovative style-based domain generalization approach called Automatic Implicit Style-Augmented Domain Generalization (AISA-DG). Specifically, we introduce a novel style-augment module designed to effectively perform implicit style enhancement on source domain data, thereby expanding the style diversity within the source domains and enhancing the adaptability to unknown target domains. Comprehensive quantitative and qualitative experiments on four fundus image datasets demonstrate the effectiveness of our proposed AISA as a state-of-the-art domain generalization approach. The source code can be accessed at https://github.com/LyWang12/AISA-DG.