Automatic generation of radiology reports from medical imaging has significant clinical utility. This task requires precise pathology identification and coherent and clinically relevant descriptions. This paper introduces Dual-path Enhanced Language for Medicine, DuEL-Med, a novel architecture that integrates a vision encoder-fed large language model (LLM) with an explicit pathology classification informatics path. A lightweight classification head is fine-tuned via LoRA to produce probabilistic pathological predictions to guide and refine the direct LLM-generated text to produce the final report. Experiments on the large-scale MIMIC-CXR and the smaller IU X-Ray datasets demonstrate DuEL-Med’s adaptive benefits. In MIMIC-CXR, it improves the GREEN score from 0.262 to 0.273 by adding correct findings. In IU X-Ray, it demonstrates robustness gain in correcting hallucinated findings from the baseline model, leading to a 15.4% improvement in high-confidence negative cases. This study underscores the value of integrating explicit clinical classification to create more reliable and context-aware automated reporting systems. The training weight and code is available at https://github.com/jink-ucla/DuEL-Med .