HEAL-MedVQA
收藏资源简介:
HEAL-MedVQA是一个全面的基准数据集,用于评估医学大型多模态模型(LMMs)的定位能力和幻觉鲁棒性。该数据集包含67,000个问答对,以及医生标注的解剖分割掩码,旨在解决医学图像解释中LMMs存在的幻觉问题。数据集创建过程中,我们采用了两个创新评估协议来评估视觉和文本的快捷学习,并且数据来源于两个大型公共数据集MIMIC-CXR和VinDr-CXR。HEAL-MedVQA旨在解决医学图像解释中LMMs存在的幻觉问题,提高医学视觉问答的鲁棒性。该数据集在医学图像解释、医学视觉问答等领域具有广泛的应用前景。
HEAL-MedVQA is a comprehensive benchmark dataset for evaluating the localization capabilities and hallucination robustness of medical large multimodal models (LMMs). It contains 67,000 question-answer pairs, together with anatomically segmented masks annotated by physicians, aiming to address the hallucination issues of LMMs in medical image interpretation and improve the robustness of medical visual question answering. During the dataset creation process, we adopted two innovative evaluation protocols to assess visual and textual shortcut learning, and the data is sourced from two large public datasets: MIMIC-CXR and VinDr-CXR. This dataset has broad application prospects in fields including medical image interpretation and medical visual question answering.




