LlamaRG: A Multi-View Large Language Model for Radiology Report Generation

Automated radiological report generation using artificial intelligence can increase clinical throughput by creating an assistive tool for physicians. In this study, we introduce llamaRG, a unified vision-language architecture built upon the LLaMA3.2-11B model that generates detailed and context rich free-form text reports from chest X-rays. Prior work commonly relies on dedicated image encoders paired with language decoders, requiring additional alignment or fusion stages. In contrast, llamaRG is built on a natively multimodal LLaMA3.2-11B backbone, allowing us to directly use the advanced language understanding and reasoning capabilities of a pretrained LLM. This unified design supports end-to-end optimization and efficient crossmodal integration without external encoders. Additionally, the model is designed for multi-view radiology report generation, enabling the model to capture the anatomical information from different views of a chest X-ray. This, in turn, helps the model to behave similarly to how radiologists create diagnostic reports based on multi-view evidence in real-world clinical workflow. To further enhance clinical fidelity and semantic coherence, we incorporate a second stage of training with Self-Critical Sequential Training as a reinforcement learning paradigm. In this phase, we utilise diagnostic and linguistic quality metrics to develop a hybrid reward function that further improves the generated reports. Comprehensive experiments performed on the benchmarking dataset MIMIC-CXR demonstrate the superiority of lla-maRG in terms of both textual fluency and clinical accuracy