Medical Abnormalities Unveiling (MAU)
收藏资源简介:
MAU数据集是由澳门大学智能物联网与系统协同创新中心创建的医学异常揭示数据集,旨在增强医学大视觉语言模型(Med-LVLMs)在医学图像中的异常检测能力。该数据集包含医学图像、用户查询以及基于异常区域的诊断响应,数据通过GPT-4V模型生成。数据集的设计过程包括利用提示方法生成诊断,并通过两阶段训练方法(异常感知指令调优和异常感知奖励)来训练模型。MAU数据集的应用领域主要集中在医学图像分析,旨在提高模型对医学图像中异常区域的理解和定位能力,从而提升医学诊断的准确性和可靠性。
The MAU dataset is a medical anomaly disclosure dataset developed by the Collaborative Innovation Center for Intelligent Internet of Things and Systems of the University of Macau, which is designed to enhance the anomaly detection performance of medical large vision-language models (Med-LVLMs) on medical images. This dataset comprises medical images, user queries, and diagnostic responses grounded in anomalous regions, with all data generated via the GPT-4V model. The dataset design process involves generating diagnostic content using prompting methods, and training the target model through a two-stage training framework: anomaly-aware instruction tuning and anomaly-aware reward modeling. The application scope of the MAU dataset is primarily concentrated in medical image analysis, with the goal of improving models' ability to understand and localize anomalous regions in medical images, thereby boosting the accuracy and reliability of medical diagnostics.




