基于投影域多物理因子噪声插入法仿真的低剂量CT重建基准数据集

计算机断层成像在临床诊断中至关重要，然而X射线辐射剂量与图像质量的矛盾制约了低剂量CT的临床应用。遵循ALARA（辐射剂量尽可能低）原则，降低剂量会导致投影数据信噪比下降，重建图像出现严重噪声与伪影，影响诊断准确性。近年来，基于深度学习的低剂量CT成像方法通过学习低剂量到标准剂量图像的映射关系，展现出显著潜力，但其临床转化受限于算法泛化能力不足，核心瓶颈在于高质量配对数据的匮乏。主要原因包括：配对数据因伦理与辐射限制难以获取；不同设备与协议导致数据异构性强；现有公开数据集规模有限、多样性不足。为应对这一挑战，本研究构建了一个多中心、多部位、多剂量水平的低剂量CT仿真数据集。数据来源于三家厂商设备，涵盖头部、胸部、腹部扫描，并采用多种重建核进行图像重建。通过对330余例临床常规剂量数据进行计算机仿真，生成了包括1/2、1/4、1/6、1/8、1/10等多种剂量水平下的低剂量CT图像。本数据集旨在为低剂量CT成像算法，特别是数据驱动型方法的研究与评估，提供大规模、多样化的基准数据资源，以促进模型泛化能力的提升，推动低剂量CT技术向临床可靠应用转化。

Computed Tomography (CT) is critically important in clinical diagnosis, yet the conflict between X-ray radiation dose and image quality hinders the clinical application of low-dose CT. Following the ALARA (As Low As Reasonably Achievable) principle, reducing radiation dose leads to decreased signal-to-noise ratio (SNR) of projection data, resulting in severe noise and artifacts in reconstructed images and compromising diagnostic accuracy. In recent years, deep learning-based low-dose CT imaging methods have shown remarkable potential by learning the mapping relationship from low-dose to standard-dose images, but their clinical translation is limited by insufficient algorithm generalization capability, with the core bottleneck being the shortage of high-quality paired data. The main reasons are as follows: paired data is difficult to obtain due to ethical and radiation restrictions; strong data heterogeneity caused by different equipment and scanning protocols; limited scale and insufficient diversity of existing public datasets. To address this challenge, this study constructs a multi-center, multi-anatomical-site, multi-dose-level simulated low-dose CT dataset. The data is sourced from scanners of three manufacturers, covering head, chest, and abdomen scans, with images reconstructed using multiple reconstruction kernels. Through computer simulation of more than 330 sets of clinical routine-dose CT data, low-dose CT images under various dose levels including 1/2, 1/4, 1/6, 1/8, and 1/10 of the routine dose are generated. This dataset aims to provide large-scale and diverse benchmark data resources for the research and evaluation of low-dose CT imaging algorithms, especially data-driven methods, to promote the improvement of model generalization capability and facilitate the translation of low-dose CT technology into reliable clinical applications.

搜集汇总

背景与挑战

背景概述

该数据集是一个为低剂量CT重建算法研究提供基准数据的大规模仿真数据集。它通过计算机仿真方法，基于330余例临床常规剂量CT数据，生成了涵盖头部、胸部、腹部等多部位、1/2至1/10等多种剂量水平的低剂量CT图像。其核心特点是多中心、多部位、多剂量水平，旨在解决高质量配对数据匮乏的瓶颈，以提升深度学习模型的泛化能力，推动低剂量CT技术的临床转化。