Automated chest X-Ray report generation using vision-language models: a Llama 3.2 based approch

Abstract

Automating radiology report generation from chest X-ray images offers the potential to ease the workload of radiologists while also improving diagnostic accuracy and efficiency. Recent advancements in vision-language models have shown strong promise in aligning visual data with natural language descriptions. In this study, we introduce a framework for automated chest X-ray report generation that leverages a LLaMA 3.2 vision-language architecture. Our approach combines image classification features with multimodal language modeling using a teacher-student learning paradigm. This integration allows the model to generate radiology reports that are both clinically accurate and contextually meaningful. By guiding the student model through teacher supervision, the framework enhances the coherence and relevance of the generated text in clinical settings. We evaluate the proposed method using a widely adopted chest X-ray dataset. The experimental results show that our model achieves strong performance, recording a BLEU score of 0.503 and a METEOR score of 0.657. These results surpass those of baseline methods in terms of both accuracy and fluency, demonstrating the effectiveness of the approach in producing high-quality medical reports. The findings suggest that large language models, when carefully adapted for multimodal medical data, can generate radiology reports that closely resemble those written by experts. This highlights the practical potential of our framework to support real-world clinical decision-making and improve healthcare delivery.