Automated radiology report generation has the potential to assist radiologists and improve diagnostic accuracy. While prior works incorporate medical knowledge graphs to mitigate data bias, their limited scope and labor-intensive construction hinder scalability across domains. To overcome these limitations, we propose a novel framework that automatically constructs and adapts generalizable clinical knowledge from diverse datasets for radiology report generation. Unlike static knowledge graphs, our method simulates human cognitive processes by extracting disease-related keywords from medical reports, forming a unified and transferable knowledge representation without requiring extra medical annotations. To seamlessly integrate knowledge with visual features, we propose a multi-grained alignment strategy that enhances contextual understanding, enabling the model to leverage knowledge to progressively focus on diverse abnormalities. This strategy consists of two key components: (1) a coarse-grained branch that employs standard self-attention to establish global vision-knowledge associations, and (2) a Knowledge-Visual Bidirectional Alignment (KVBA) module, which refines fine-grained vision-knowledge representations through bidirectional cross-attention. Evaluated on public IU-Xray and MIMIC-CXR benchmarks, our method successfully enhances the accuracy of disease-related keywords mentioned in medical reports and achieves competitive results in clinical efficacy metrics in report generation.