Abstract
Due to the scarcity of annotated data and strict privacy regulations in cross-center deployment, single-domain generalization (SDG) has become imperative for high-precision medical image analysis under resource constraints. Most existing methods rely heavily on static isolated 2D image textures for feature extractions. However, these methods often fail to break the accuracy ceiling as they ignore longitudinal context and inherent structural correlations, causing early perception bottleneck and weak structural robustness. To address these issues, we propose a novel Prior-Aware Spatio-Temporal Evolutionary Distillation Dual-Stream Network (PASTE-D2Net) for accurate medical single-domain generalization, while providing meaningful interpretation on the results. In particular, a new spatio-temporal evolutionary distillation strategy is designed to capture the longitudinal and cross-sectional patterns of diseases, endowing PASTE-D2Net with dynamic evolutionary inference to effectively enhance the prominence of early lesions. Furthermore, a novel structure-texture dual-stream integration mechanism is introduced to compensate for the impacts of missing modalities, promoting the capacity of PASTE-D2Net in anchoring on anatomical structure to enhance its robustness towards complex clinical scenarios. Extensive experimental results on multiple datasets indicate that PASTE-D2Net outperforms the state-of-the-art methods and provides meaningful interpretations, offering a new data-efficient paradigm for clinical deployment.
Funding Information
  • Hubei Provincial Natural Science Foundation of China (2026AFB556)