Fine-tuning Vision Language Models with Graph-based Knowledge for Explainable Medical Image Analysis