Existing few-shot segmentation algorithms have made notable advancements in the field of medical image segmentation. However, most general methods use a single global prototype extracted from the support image to represent all information, which can lead to ambiguities of the local parts and not fully utilize the information of the query image. In this paper, we propose two novel modules, i.e., Multi-Support Prototype Generator (MSPG) and Cross Prototype Generator (CPG), for adaptively extracting multiple prototypes and leveraging query information. Specifically, MSPG is a training-free approach that extracts more representative prototypes for organs with different scales. CPG fuses support and query information to generate cross prototypes that guide the prediction of query masks. By integrating MSPG and CPG, we propose the Cross Prototype Network (CPNet), which leverages the information of support and query images to mitigate distribution shifts between query images and the support set. In particular, CPNet achieves state-of-the-art performance with higher accuracy and lower deviation on two widely used datasets, abdominal MR (ABD) and cardiac MR (CMR), under two different settings.