This study aims to propose an integrated framework, the query-adaptive graph iteration network (QAGIN), to systematically address three core challenges in medical image few-shot segmentation (FSS), including the high intra-class variability of anatomical structures, significant domain shifts across imaging devices and protocols, and inherent ambiguity in boundaries between target foregrounds and complex backgrounds. To this end, the QAGIN architecture is meticulously designed to integrate a series of advanced, mutually synergistic mechanisms. Firstly, the architecture’s query-adaptive prototype generation module utilizes FiLM layers to dynamically calibrate support features and generate prototypes adapted to the feature distribution of query images to overcome domain shifts. On the basis of obtaining high-quality adaptive prototypes, the prototype-guided dynamic graph reasoning mechanism explicitly models the internal structural relationships of query images via the hypernetwork-modulated graph convolutional network (GCN), thereby enhancing the model’s structure-aware capabilities in significant anatomical variation conditions. Subsequently, by taking the structure-enhanced feature representations as the input, the iterative foreground-background optimization module employs a ConvGRU decoder to construct a self-correcting loop and perform multi-round refinement of segmentation results, achieving progressive convergence and refinement of ambiguous foreground-background boundaries. The effectiveness of this framework is validated on the challenging CHAOS-T2 abdominal organ dataset and the MS-CMRSeg cardiac magnetic resonance dataset. Experimental results demonstrate that QAGIN provides a solution with significant advantages in robustness, accuracy, and generalization capabilities for data-scarce clinical application scenarios.