Chest ImaGenome

该数据集名为Chest ImaGenome，在现有研究基础上，通过基于规则的文本分析和基于图谱的边界框提取技术，对报告文本和正面图像中的解剖结构和相关病理进行了结构化处理。该数据集包含18个解剖区域的标注和9个病理标签，病理标签的平均精确度和召回率分别为0.9819和0.9875。其规模涵盖了来自MIMIC-CXR数据集的217,417份报告文本和正面图像，其中500份报告和1000张正面胸部X光片得到了双重标注和边界框。该数据集的任务是进行多标签胸部X光分类，并将发现定位到相应的解剖区域。

This dataset is named Chest ImaGenome. Building upon existing research, it performs structured annotation of anatomical structures and associated pathologies in both the report texts and frontal-view images using rule-based text analysis and graph-based bounding box extraction techniques. This dataset provides annotations for 18 anatomical regions and 9 pathology labels, with the average precision and recall of the pathology labels reaching 0.9819 and 0.9875 respectively. In terms of scale, it encompasses 217,417 pairs of report texts and frontal-view images sourced from the MIMIC-CXR dataset, among which 500 reports and 1,000 frontal chest X-ray images have undergone dual annotation with bounding box annotations. The task supported by this dataset is multi-label chest X-ray classification, as well as localizing identified findings to their corresponding anatomical regions.

搜集汇总

背景与挑战

背景概述

Chest ImaGenome数据集是基于MIMIC-CXR数据集的结构化版本，包含217,417份报告文本和正面图像的标注，覆盖18个解剖区域和9个病理标签，精确度和召回率均较高，适用于多标签胸部X光分类和定位任务。