The generation of medical reports seeks to automatically create detailed written descriptions that correspond to radiology images, a task that is increasingly significant in the healthcare sector. Despite advancements, conventional approaches frequently depend on supervised learning, which necessitates extensive annotated datasets that are hard to come by because of privacy issues and the expense of professional annotations. Furthermore, previous methods struggled to successfully integrate multi-modal information and capture hierarchical visual-semantic links. We suggest a novel hybrid architecture that combines CNNs, GNNs and transformer-based modules for reliable report generation in order to overcome these constraints. The GNN uses superpixel-based segmentation to represent spatial relationships between regions, whereas the CNN extracts fine-grained visual characteristics. A transformer-based encoder-decoder framework is utilized to fuse these features, allowing the model to provide reports that are coherent and semantically rich. The effectiveness of our architecture in producing precise and contextually aware medical reports is demonstrated by evaluations on benchmark datasets, such as IU-XRay, which show higher performance in BLEU, METEOR, and ROUGE measures.