A whole-body FDG-PET/CT dataset with manually annotated tumor lesions (FDG-PET-CT-Lesions)

Purpose: To provide an annotated data set of oncologic PET/CT studies for the development and training of machine learning methods and to help address the limited availability of publicly available high-quality training data for PET/CT image analysis projects. This data can also be used for machine learning challenges.
Data: The anonymized publication of data was approved by the local ethics committee and data protection officer. 501 consecutive whole body FDG-PET/CT data sets of patients with malignant lymphoma, melanoma and non small cell lung cancer (NSCLC) as well as 513 data sets without PET-positive malignant lesions (negative controls) examined between 2014 and 2018 at the University Hospital Tübingen were included. All examinations were acquired on a single, state-of-the-art PET/CT scanner (Siemens Biograph mCT). The imaging protocol consists of a diagnostic CT scan (mainly from skull base to mid-thigh level) with intravenous contrast enhancement in most cases, except for patients with contraindications. The following CT parameters were used: reference dose of 200 mAs, tube voltage of 120 kV, iterative reconstruction with a slice thickness of 2 - 3 mm. In addition, a whole-body FDG-PET scan was acquired 60 minutes after I.V. injection of 300-350 MBq 18F-FDG. PET data were reconstructed using an ordered-subset expectation maximization (OSEM) algorithm with 21 subsets and 2 iterations and a gaussian kernel of 2 mm and a matrix size of 400 x 400.

All data sets were analyzed in a clinical setting by a radiologist and nuclear medicine physician in consensus identifying primary tumors and metastases in each data set. All FDG-avid lesions identified as malignant based on patient history and prior examinations were manually segmented on PET images in a slice-per-slice manner by a single reader using dedicated software (NORA imaging platform, University of Freiburg, Germany).

We provide the anonymized original DICOM files of all studies as well as the DICOM segmentation masks. Primary diagnosis, age and sex are provided as non-imaging information. In addition, we provide a link to code that can produce a preprocessed version of the data with resampled and aligned PET, CT, and masks as a NIfTI file ready to use in machine learning projects.

数据集用途：本数据集为肿瘤学正电子发射断层显像/X线计算机体层成像（PET/CT）标注数据集，旨在支撑机器学习方法的开发与训练，同时弥补当前PET/CT影像分析项目公开高质量训练数据匮乏的现状。该数据集亦可用于机器学习挑战赛。
数据说明：本数据集的匿名化发布已通过当地伦理委员会与数据保护专员的审批。纳入2014年至2018年间于蒂宾根大学医院完成检查的501例恶性淋巴瘤、黑色素瘤及非小细胞肺癌（NSCLC）患者全身氟代脱氧葡萄糖PET/CT（FDG-PET/CT）数据集，以及513例PET阳性恶性病灶阴性的对照数据集。所有检查均采用同一台顶尖PET/CT扫描仪西门子（Siemens）Biograph mCT完成。
本次检查的影像方案包含诊断性CT扫描：扫描范围多为颅底至大腿中段，多数患者接受静脉增强扫描，存在造影禁忌证者除外。CT扫描参数如下：参考剂量200 mAs，管电压120 kV，迭代重建算法，层厚2~3 mm。此外，所有受试者于静脉注射300~350 MBq的18F-氟代脱氧葡萄糖（18F-FDG）后60分钟，完成全身FDG-PET扫描。PET影像采用有序子集期望最大化（OSEM）算法重建：子集数21，迭代次数2次，高斯核宽度2 mm，矩阵尺寸400×400。
所有数据集均由放射科医师与核医学医师在临床场景下联合阅片，共同标注每例数据中的原发肿瘤与转移灶。基于患者病史与既往检查结果，所有被判定为恶性的FDG摄取阳性病灶均由一名阅片者使用专用软件（德国弗莱堡大学NORA影像平台）在PET影像上逐层手动分割。
本数据集提供所有检查的匿名化原始DICOM文件，以及DICOM格式的分割掩码。非影像元数据包含患者的原发诊断、年龄与性别。此外，我们还提供配套代码链接，可将数据预处理为重采样并配准后的PET、CT及掩码文件，格式为可直接用于机器学习项目的NIfTI文件。

搜集汇总

背景与挑战

背景概述

该数据集是一个全身FDG-PET/CT数据集，包含501例恶性肿瘤患者和513例阴性对照的PET/CT扫描数据，所有病灶由专家手动标注，旨在为机器学习提供高质量训练数据。数据集提供匿名DICOM文件和分割掩码，并附带预处理代码，方便直接用于图像分析项目。