摘要
医疗视频诊断需从动态组织反应中推断临床决策。现有方法依赖端到端学习，存在重外观轻病理、缺乏临床先验及无反事实对比等局限。本文提出 MedVCR，一种模拟临床诊断思维的反事实推理框架。该框架包含基于扩散模型的反事实生成器、编码临床规则的反事实表示学习模块，以及结合视频级评估与帧级分析的双重诊断策略。在完全监督和弱监督设置下，MedVCR 较基线提升 2.6%-10.2%，验证了其有效性。
AI 推荐理由
论文核心提出反事实推理框架，模拟临床诊断思维，直接针对推理机制进行创新。
研究机构
Center for Data Science in Clinical Medicine, Peking University Third Hospital
The State Key Lab of Brain-Machine Intelligence, Zhejiang University
Department of Gynecology and Obstetrics, 7th Medical Center of Chinese PLA General Hospital
School of Computer Science, Peking University
School of Psychological and Cognitive Sciences, Peking University
State Key Lab of General AI, Peking University
NaT’Eng Research Center of Visual Technology
Beijing Key Laboratory of Behavior and Mental Health, Peking University
Embodied Intelligence Lab, PKU-Wuhan Institute for Artificial Intelligence
论文信息