Retrieval-Augmented Generation (RAG) allows large language models to generate summaries focusing primarily on textual content. However, in specialized academic fields, documents often contain multimodal information such as images and tables, which are crucial for fully understanding and deepening the comprehension of the literature. Meanwhile, summary generation methods that focus solely on text are unable to scale to the volume of text that typical RAG systems can index. To overcome this limitation, we introduced a new summary system, Global-Graph RAG Summarizer (GGRS), specifically designed for multiple PDF-format academic documents in the medical field that include not only text but also images and tables. GGRS combines graph-based RAG with LLMs to effectively process and integrate multimodal data, producing high-quality summaries. Evaluation results show that our GGRS system performs exceptionally well in generating semantically rich summaries, achieving win rates of 89% in accuracy and 94% in comprehensiveness compared to GPT-4.