Large Language Models (LLMs) have recently shown strong performance on Question Answering (QA) tasks. However, in the medical domain, they often hallucinate due to gaps in domain-specific knowledge. Retrieval-Augmented Generation (RAG) mitigates this issue by grounding answers in external information, but most existing approaches rely on single-round or single-retriever paradigms, which struggle with multi-hop questions or when relevant knowledge is distributed across heterogeneous sources. To this end, we propose the Iterative Dynamic Routing Framework (IDRF) for medical QA. IDRF performs iterative reasoning and retrieval, enabling the model to generate multiple retrieval intents, access heterogeneous knowledge sources in parallel, and progressively construct the final answer. The framework consists of three components: a master agent that plans each round, a retriever module with four specialized retrievers, and a summary agent that condenses retrieved results into feedback for the master agent. Experiments on three public medical QA datasets show that IDRF significantly outperforms competitive baselines, and ablation studies confirm the contribution of each component.