FlowNet-PET Training Set

The training set used for FlowNet-PET found in https://github.com/teaghan/FlowNet_PET A dataset of 300 phantoms generated using the 4D extended cardiac-torso (XCAT) anthropomorphic digital phantom (2x4x4) mm3 voxels. The dataset is split into training (270) and validation (30) sets. For each phantom, ten frames spaced across a single breath cycle were created, each frame having (108x152x152) pixels resembling the activity distribution throughout the phantom. To emulate a simple representation of PET data acquisition, these distributions were used to sample a random number (between 1x106 and 9x106) of counts per frame. The parameters for each XCAT phantom were varied randomly. This included the gender, axial section included, transaxial shifts, scaling factors in each direction, size and location of the lung lesion, extent of the diaphragm motion, extent of the chest expansion, and the activity of each organ.本数据集为GitHub仓库https://github.com/teaghan/FlowNet_PET中公开的FlowNet-PET模型所用训练集，包含300个体模样本，均基于体素（voxel）尺寸为(2×4×4) mm³的4D扩展心脏躯干（XCAT）拟人化数字体模生成。数据集按270个训练样本、30个验证样本的比例划分。针对每个体模，生成了分布于单次呼吸周期内的10帧时序图像，每帧包含(108×152×152)个像素，用于表征体模内的放射性活度分布。为模拟PET数据采集的简化表征，基于上述活度分布对每帧图像进行采样，得到1×10⁶至9×10⁶范围内的随机计数。每个XCAT体模的参数均经随机调整，涵盖受试者性别、所包含的轴向断层、横断面偏移、各方向缩放因子、肺部病灶的尺寸与位置、膈肌运动幅度、胸廓扩张幅度，以及各器官的放射性活度。

搜集汇总

背景与挑战

背景概述

FlowNet-PET Training Set是一个用于PET成像运动校正的机器学习训练数据集，包含300个基于4D XCAT数字体模生成的幻影，每个幻影有10个帧模拟PET数据采集，体素尺寸为(2x4x4) mm³，帧尺寸为(108x152x152)像素，并随机变化多种生理参数以增强多样性。数据集分为270个训练样本和30个验证样本，总大小为3.3 GB，适用于FlowNet-PET项目，旨在支持PET图像中的运动伪影校正研究。