3D多模态医疗数据集-分割

大规模通用 3D 医疗图像分割数据集 (M3D-Seg)。3D 医学分割是医疗图像分析的主要挑战之一。目前，由于隐私和成本的限制，缺乏大规模的公开利用的 3D 医学图像和标注。 为此，我们收集了 25 个公开的 3D CT 分割数据集，包括CHAOS, HaN-Seg, AMOS22, AbdomenCT-1k, KiTS23, KiPA22, KiTS19, BTCV, Pancreas-CT, 3D-IRCADB, FLARE22, TotalSegmentator, CT-ORG, WORD, VerSe19, VerSe20, SLIVER07, QUBIQ, MSD-Colon, MSD-HepaticVessel, MSD-Liver, MSD-lung, MSD-pancreas, MSD-spleen, LUNA16. 这些数据集被统一编码从 0000-0024，总计 5,772 个 3D 图像和 149,196 个 3D mask 标注。其中，每个 mask 所对应的语义标签可以通过文本表示。每个文件夹内，包含 ct 和 gt 两个子文件夹存储数据和标注，并利用 json 文件作为 split。 dataset_info.txt 描述了每个数据集标签的文本表示。作为通用分割数据集，更多的公开和私有数据集可以按相同格式统一进来，从而构建大规模的 3D 医疗通用分割数据集。

Large-scale General-purpose 3D Medical Image Segmentation Dataset (M3D-Seg). 3D medical image segmentation is one of the core challenges in medical image analysis. Currently, the availability of large-scale publicly accessible 3D medical images and their corresponding annotations is severely restricted due to privacy concerns and high annotation costs. To address this issue, we collected 25 public 3D CT segmentation datasets, including CHAOS, HaN-Seg, AMOS22, AbdomenCT-1k, KiTS23, KiPA22, KiTS19, BTCV, Pancreas-CT, 3D-IRCADB, FLARE22, TotalSegmentator, CT-ORG, WORD, VerSe19, VerSe20, SLIVER07, QUBIQ, MSD-Colon, MSD-HepaticVessel, MSD-Liver, MSD-lung, MSD-pancreas, MSD-spleen, and LUNA16. All these datasets are uniformly encoded with IDs ranging from 0000 to 0024, totaling 5,772 3D images and 149,196 3D mask annotations. The semantic label corresponding to each mask can be represented via text. Each dataset folder contains two subfolders named "ct" and "gt" for storing raw data and annotations respectively, and a JSON file is used for dataset splitting. The "dataset_info.txt" file describes the text representations of the labels for each dataset. As a general-purpose segmentation dataset, more public and private datasets can be incorporated in the same format to build a larger-scale 3D medical general segmentation dataset.

搜集汇总

背景与挑战

背景概述

该数据集是一个大规模通用的3D医学图像分割数据集（M3D-Seg），整合了25个公开的3D CT分割数据集，包含5,772个3D图像和149,196个3D掩码注释。数据集支持传统分割任务和文本提示分割等高级任务，采用统一的格式和组织方式，便于使用和扩展。