Traditional Few-Shot Medical Image Segmentation (FSMIS) methods primarily focus on mining information from support image to guide query image segmentation, while insufficiently exploring the contextual information within the query image itself. Due to significant anatomical variability and differences in imaging conditions, the targets in support and query images may exhibit substantial discrepancies, even within the same class. Consequently, inadequate utilization of query context limits the model’s ability to adapt to intra-class variations, reducing its generalization capability. To address this issue, inspired by Mamba’s capacity for context modeling with linear complexity, we propose ABE-Mamba, a novel query-centric FSMIS framework. ABE-Mamba enhances the model’s adaptability to intra-class variations by effectively mining the query image’s contextual information, and introduces an adversarial training mechanism to improve high-order consistency between the predicted mask and the ground truth. Firstly, a Bidirectional Enhanced Mamba module is introduced into the generator to fully mine the rich semantics in query images by facilitating bidirectional enhancement of both local and global context. Secondly, a cross-SS2D block is designed and incorporated into the discriminator to improve its capability in distinguishing subtle differences between real and generated masks, thereby enhancing the alignment between predicted masks and ground truths. Thirdly, a pyramid structure is integrated into the generator to facilitate multi-scale feature extraction, thereby improving robustness to objects of varying sizes. Extensive experiments on three widely used datasets (Syn-CT, CHAOS-MRI, and Card-MRI) demonstrate that ABE-Mamba achieves competitive performance.