EVENT DOSSIER
Visual Language Models (VLMs) face a key gap in standardizing raw heterogeneous medical data in medical AI applications. The research team constructed the Medical Data Standardization Benchmark (MDS-Bench) containing 1,939 pieces of medical data covering diverse clinical practices, radiological modalities, and catalog layout tasks, to evaluate the models’ ability to recognize source formats, convert images, extract text, and organize structured image-text pairs. Experimental results show that even the best-performing VLM (Gemini 3 Flash) achieved only a 48.6% success rate in end-to-end tasks. The study indicates that standardizing raw medical data is a key bottleneck hindering VLMs from being used for diagnosis in real clinical settings.
本文提出视觉语言模型（VLMs）在医疗 AI 应用中面临的关键缺失环节：原始异构医疗数据的标准化。研究构建了医疗数据标准化基准（MDS-Bench），包含人工标注的 1,939 个涵盖多样化临床实践、放射学模态及目录布局的任务，评估模型识别源格式、转换图像、提取文本及组织结构化图文对的能力。实验表明，即使表现最佳的 VLM（Gemini 3 Flash）端到端成功率仅为 48.6%。该研究指出，原始医疗数据标准化是阻碍 VLMs 在真实临床实践中进行诊断的关键瓶颈。