Advancing clinical decision-making in oncology requires structured domain knowledge derived from vast biomedical literature. This study presents a novel framework for dynamically refining specialized knowledge graphs by leveraging distant supervision and iterative adaptation. By integrating domain adaptation techniques with deep learning-based entity recognition and relationship extraction, the proposed system efficiently extracts and organizes oncological knowledge without manual annotation. Experimental results demonstrate its effectiveness in identifying new domain-specific concepts, relationships, and structured insights, and demonstrate scalability for automated biomedical knowledge discovery.