Recent advances in multimodal large language models have unlocked new opportunities for developing next-generation auxiliary diagnostic systems capable of jointly interpreting medical images and clinical text. However, general-purpose models such as GPT-4o exhibit limited performance in specialized medical domains like dental panoramic X-ray diagnosis, primarily due to insufficient encoding of domain-specific knowledge and the instability of full-parameter fine-tuning on small-scale datasets. To overcome these limitations, this study introduces a specialized multimodal framework for oral disease analysis built upon the LLaVA-Med architecture. The proposed framework features a collaborative dual-model design that simultaneously handles diagnostic inference and image-quality assessment, along with an integrated multi-round reasoning mechanism that improves robustness through iterative prediction and self-correction. Experimental results show that our approach significantly outperforms conventional full-parameter fine-tuning and mainstream multimodal baselines, offering an efficient and reproducible paradigm for adapting large models to medical domains.