Prior‐Knowledge‐Guided Dual‐Stream Fusion Network for Few‐Shot Medical Image Classification

Few‐shot medical image classification focuses on learning effective feature representations from limited labeled data, however, this task remains challenging due to domain shifts and high inter‐class similarity commonly observed in medical imaging. Recent vision–language models, such as contrastive language–image pretraining (CLIP), utilize large‐scale semantic knowledge and have demonstrated significant potential. However, when applied to medical data, CLIP‐based methods tend to emphasize global semantic alignment, thereby failing to adequately capture fine‐grained pathological details that are crucial for reliable diagnosis under limited supervision. To address this limitation, we propose a prior knowledge‐guided dual‐stream fusion network (PKDNet) for cross‐domain few‐shot medical image classification. PKDNet retains CLIP's powerful vision–language semantic priors to provide robust global context, while introducing an auxiliary convolutional branch to explicitly capture local structural and textural information. The two streams are fused through an uncertainty‐aware mechanism that adaptively integrates complementary semantic and structural representations. Extensive experiments on the MedFMC benchmark show that PKDNet consistently outperforms existing approaches, achieving improvements of 10.82%, 55.74%, and 0.5% under the five‐shot setting across three datasets. These results highlight the importance of explicitly combining semantic and structural information when adapting vision–language models to data‐scarce medical imaging scenarios.