The advancement of medical artificial intelligence (AI) has set the stage for the realization of conversational diagnosis, where AI systems mimic human doctors by engaging in dialogue with patients to deduce diagnoses. This study introduces an innovative approach using external planners augmented with large language models (LLMs) to develop a medical task-oriented dialogue system. This system comprises a policy module for information gathering, a LLM based module for natural language understanding and generation, addressing the limitations of previous AI systems in these areas. By emulating the two-phase decision-making process of doctors disease screening and differential diagnosis. we designed two distinct planners. The first focuses on collecting patient symptoms to identify potential diseases, while the second delves into specific inquiries to confirm or exclude these diseases. Utilizing reinforcement learning and active learning with LLMs, we trained these planners to navigate medical dialogues effectively. Our evaluation on the MIMIC-IV dataset demonstrated the system's capability to outperform existing models, indicating a significant step towards achieving automated conversational disease diagnostics and enhancing the precision and accessibility of medical diagnoses.