Unifying Spatial and Semantic Augmentation Strategies for Cross-Domain Semi-Supervised Medical Image Segmentation

Medical image datasets often suffer from inadequate labeling, primarily due to ethical constraints, expertise challenges, and high annotation costs. In addition, domain shifts, resulting from variations in imaging protocols, equipment settings, and patient demographics, pose further challenges. It is imperative to address the dual challenges of limited annotated data and prevalent domain shifts in the field of medical imaging, as these obstacles significantly impede the effectiveness of deep learning-based segmentation methods in clinical applications. Most existing approaches often address these issues independently, adopting either semi-supervised strategies for label efficiency or domain generalization techniques for distribution shifts, thereby neglecting the inherent interdependence between data scarcity and domain variation. Our research emphasizes the importance of a unified framework that synergistically combines semi-supervised learning and domain generalization techniques. In this paper, we introduce a novel framework called the Spatio-Semantic Augmentation Network (SSAN), designed to operate concurrently at both the image space and semantic feature levels. SSAN integrates two key components: Spatial Dual Transformation (SDT) and Semantic Cascade Perturbation (SCP). SDT, which includes Global Hallucination and Center Mix, generates diverse images and facilitates consistent learning through cross-supervisory interactions. SCP leverages statistical methods to dynamically synthesize diverse feature styles at the semantic level, thereby enhancing the model's exposure to varied data representations. Extensive experimental evaluations demonstrate the efficacy of the SSAN framework in enhancing cross-domain segmentation performance under limited labeled data, improving model generalization to new medical imaging domains. Comparative analysis with state-of-the-art methods reveals the superior performance of SSAN.