Extracting and Encoding: Leveraging Large Language Models and Medical Knowledge to Enhance Radiological Text Representation