MedErr-CT
收藏资源简介:
MedErr-CT数据集是一个用于评估医疗多模态大语言模型(MLLMs)在CT报告中识别和纠正错误能力的视觉问答(VQA)基准。该数据集由韩国蔚山大学医学院生物医学工程系的Sunggu Kyung等人创建,包含六种错误类型,分为视觉中心错误和词汇错误,并分为三个任务难度等级:分类、检测和纠正。数据集采用CT-RATE和RadGenome-Chest CT数据集的区域级报告,利用DSPy框架和LLaMA-3.3-70B模型生成错误报告。MedErr-CT数据集旨在解决临床诊断中CT报告错误率高的问题,通过评估MLLMs在识别和纠正CT报告中的错误的能力,推动更可靠和临床适用的MLLMs的开发,最终帮助减少诊断错误并提高临床实践中的准确性。
The MedErr-CT dataset is a visual question answering (VQA) benchmark designed to assess the capability of medical multimodal large language models (MLLMs) to identify and correct errors within CT reports. It was developed by Sunggu Kyung et al. from the Department of Biomedical Engineering, College of Medicine, University of Ulsan, Republic of Korea. The dataset includes six error types, which are classified into two categories: visual-centric errors and lexical errors, and is stratified into three task difficulty tiers: classification, detection, and correction. The dataset utilizes regional-level reports sourced from the CT-RATE and RadGenome-Chest CT datasets, and generates erroneous reports via the DSPy framework and the LLaMA-3.3-70B model. The core objective of the MedErr-CT dataset is to address the high error rate issue of CT reports in clinical diagnosis. By evaluating the ability of MLLMs to identify and correct errors in CT reports, it promotes the development of more reliable and clinically applicable MLLMs, ultimately helping to reduce diagnostic errors and improve accuracy in clinical practice.




