The increase of fake news on digital platforms poses significant challenges to information integrity and public discourse. Existing detection methods focusing on text-only or simple multimodal fusion fail to capture complex inter-modal dependencies and contextual relationships. In this paper, we propose Bi-modal Representation Attention with Dynamic heterogeneous graphs (BRAD) for multimodal fake news detection that dynamically constructs and processes heterogeneous graphs from news content. Our approach leverages both textual and visual information to create better multimodal representations, where each node in the graph can adaptively select the most informative neighborhood type for embedding updates through a learnable decision mechanism. The BRAD architecture consists of specialized networks for neighborhood selection and representation learning, enabling fine-grained control over information propagation in the graph structure. We evaluate our method on the Fakeddit dataset, a large-scale multimodal benchmark with diverse fake news categories. Experimental results demonstrate that our approach achieves superior performance compared to existing multimodal baselines with acc., precision, recall, and f1-score values of 91.28%, 87.32%, 91.23%, and 89.23%, respectively.