Multi-party Dialogue Reading Comprehension (MDRC) is a task focused on understanding dialogues involving multiple participants and answering related questions. There are two main challenges for research in this field: 1) Complex discourse structure stems due to the frequent switching of speakers and changes in topics. 2)The distinct expressive intentions and speaking styles of individual speakers introduce intricate co-referential data into the dialogue. Therefore, we present the Dual-level Graph Reasoning with Key Block Decoupling (DlGR-KB) method to filter out noisy, irrelevant contexts and maintain comprehensive and accurate dialogue information flow. We first decouple the key block most relevant to the question based on a topic-aware dialogue segmentation method. To maintain the relevance of the information flow, we then employ a dual-level graph reasoning approach that integrates local and global relevance for precise and contextual awareness. At the local level, we construct a Local Interlocutor-Perceived Question Heterogeneous Graph (LIQHG) to directly connect questions with key block contents, aiding precise answer localization. Furthermore, to address the intricacies of co-reference information, we employ a global coreference-aware module (GCM) to refine the semantic logic at the representation level. Experiments conducted on the benchmark datasets demonstrate that our approach obtains consistent and significant improvements, and outperforms the performance of state-of-the-art methods.