Medical question answering (QA) is a type of natural language processing (NLP) task to support a series of medical applications. Pre-training language models (PLMs) are widely used in medical QA tasks by fine-tuning the models with related datasets. To further incorporate medical knowledge, knowledge-enhanced PLMs and knowledge fusion methods have been proposed to integrate medical knowledge bases (KBs) into models. However, medical KBs are structured knowledge that lacks semantic coherence. Existing methods that insert knowledge without conversion may introduce semantically unclear statements and deviate the input text from its original meaning, leading to the knowledge noise (KN) issue. Large language models (LLMs) trained with massive corpora contain vast semantic knowledge. In a range of NLP tasks, particularly text generation tasks, LLMs have demonstrated outstanding capabilities. Combining the medical KBs and LLMs can provide a more diverse and comprehensive body of knowledge. In this paper, we propose a generative knowledge fusion method, GKF-mQA, to combine medical KBs and LLMs to generate coherent medical knowledge. The generated knowledge is selected according to the relevance scores with input text to reduce the KN. The optimal knowledge is then inserted into the input text during the fine-tuning stage, thereby enhancing the performance of medical QA models. Experimental results demonstrate that our method achieves improvements on three medical QA datasets in two languages.