In medical federated learning (FL), acquiring labeled data is particularly challenging due to the high cost of expert annotations. Active learning (AL) offers a promising solution by selecting the most informative samples for annotation, thereby improving model performance with minimal labeling effort. However, traditional AL methods, such as entropy-based methods, often suffer from instability in non-IID federated settings and may lead to redundant sample selection. To address these challenges, we propose Prototype-Based Hierarchical Clustering Federated Active Learning (PBHC-FAL), a novel Federated Active Learning method tailored for medical scenarios. Our approach begins with a prototype-based outer clustering, where each client utilizes its locally trained model to compute prototypes of labeled samples and cluster unlabeled samples accordingly. Besides, we introduce a fine-grained inner clustering within these outer clusters to further improve sample diversity. Extensive experiments on medical image classification tasks demonstrate the effectiveness of our proposed method.