A Training-free Modular Collaborative Framework: Application to Nasopharyngeal Endoscopy Report Generation

Automatically generating accurate diagnostic reports from nasopharyngoscopy images holds significant clinical value. However, the lack of publicly available training datasets in this domain hinders the fine-tuning of pretrained models. Moreover, general-purpose large models often suffer from hallucinations due to the absence of empirical and structured prompt design. To address these challenges, we propose TfMC, a Training-free Modular Collaborative framework that synergistically integrates Multimodal Vision-Language Models (MVLMs) with Large Language Models (LLMs) for zero-shot nasopharyngoscopy report generation. We propose a diagnostic decision tree-driven prompt generation strategy based on clinical expertise, converting node routing decisions into visual question-answering to dynamically guide prompt generation. Additionally, we design a symptom-based image-node matching and knowledge generation strategy to enhance the evidential and targeted of prompts. Experimental results on a clinically dataset demonstrate that our framework effectively mitigates hallucination in zero-shot settings while ensure consistency between report content and diagnostic result, significantly improving report accuracy and interpretability.