摘要目的:探讨不同深度学习策略对低剂量脑 18F-FDG PET图像的降噪效果。 方法:本研究为回顾性方法学研究，分析2023年5月至2024年1月间中山大学附属第五医院的50例患者(男35例、女15例，年龄20~87岁)的脑部PET/CT影像资料。患者先行CT扫描，按体质量注射3.7MBq/kg 18F-FDG后扫描2min，获得全剂量脑PET正弦图。将全剂量PET列表模式数据降采样至全剂量计数水平的1/2、1/4和1/20，生成低剂量PET正弦图。采用三维(3D)有序子集最大期望值迭代法(迭代数2，子集数20)对全剂量和低剂量正弦图进行重建，并进行随机、衰减和散射校正。采用4种深度学习的去噪方法：(1)基于低剂量PET的单模态3D生成对抗网络(GAN)-1；(2)引入注意力机制的低剂量PET单模态注意力GAN(AttGAN)-1；(3)低剂量PET与CT双模态输入的AttGAN-2；(4)分频低剂量PET和CT同时作为输入的分频双模态AttGAN(AttGAN-FS-2；分频率处理时，对PET重建图像先行傅里叶变换，分离高低频图像，再进行逆傅里叶变换等处理，学习得到最终去噪后图像)。将50例数据通过无放回的简单随机抽样方法分为训练集(70%)、验证集(10%)和测试集(20%)，行5折交叉验证。使用归一化均方误差(NMSE)、结构相似性(SSIM)、峰值信噪比(PSNR)、对比度噪声比(CNR)和特选脑区的SUV mean和SUV max误差对不同模型的降噪图像进行评估，采用Wilcoxon符号秩检验分析不同方法间上述指标的差异。 结果:在1/2、1/4和1/20剂量条件下，AttGAN-FS-2都展现出最好的性能，其NMSE、SSIM、PSNR和CNR均与低剂量PET和GAN-1降噪后的图像差异有统计学意义( Z值：2.92~6.15，均 P<0.005)。如在1/20剂量下，GAN-1、AttGAN-1、AttGAN-2、AttGAN-FS-2模型的NMSE和SSIM(中位数)分别为0.08和0.87、0.08和0.88、0.07和0.89、0.06和0.91( Z值：3.24~5.77，均 P<0.005)。 结论:结合多种深度学习策略的AttGAN-FS-2模型能对低剂量PET图像展现出更好的降噪效果。

abstractsObjective:To investigate the denoising performance of different deep learning (DL) strategies on low-dose brain 18F-FDG PET images. Methods:This retrospective methodological study was conducted on brain PET/CT images of 50 patients (35 males, 15 females, age 20-87 years) who received 3.7MBq/kg 18F-FDG at the Fifth Affiliated Hospital of Sun Yat-sen University between May 2023 and January 2024. Full-dose PET data were acquired with 2min scan. CT scans were acquired before PET scanning. Low-dose PET sinograms were generated by down-sampling the full-dose list mode data to 1/2, 1/4, and 1/20 of full-dose count level. Both full-dose and low-dose sinograms were reconstructed with random, CT-based attenuation and scatter corrections using the three-dimensional (3D) ordered-subsets expectation maximization (OSEM) algorithm (2 iterations, 20 subsets). A total of 4 DL denoising methods were established: (1) 3D conditional generative adversarial networks (GAN) using only low-dose PET as input (GAN-1); (2) 3D attention-based GAN (AttGAN) with low-dose PET input (AttGAN-1); (3) 3D AttGAN with low-dose PET and CT inputs (AttGAN-2); (4) 3D AttGAN with frequency-separation using low-dose PET and CT inputs (AttGAN-FS-2). For AttGAN-FS-2, during the frequency division process, high- and low-frequency components were extracted from the PET reconstructed images via Fourier transform, then inversed Fourier transform, denoised separately, and finally combined to produce the final denoised images. The dataset was separated into training (70%), validation (10%) and testing (20%) sets using simple random sampling without replacement with a fixed random seed. A 5-fold cross-validation scheme was then applied to test all 50 patients. Performance was evaluated against full-dose PET using normalized mean square error (NMSE), structural similarity (SSIM), peak signal-to-noise ratio (PSNR), contrast-to-noise ratio (CNR), SUV mean and SUV max bias of selected brain ROIs. Wilcoxon signed rank test was used to analyze the differences between the denoising methods. Results:AttGAN-FS-2 showed the best performance among all dose levels, with statistical difference as compared by low-dose PET and GAN-1 denoised images for NMSE, SSIM, PSNR, and CNR ( Z values: 2.92-6.15, all P<0.005). NMSE, SSIM quantitative evaluation results (median) of each model at 1/20 dose were: GAN-1: 0.08, 0.87, AttGAN-1: 0.08, 0.88, AttGAN-2: 0.07, 0.89, AttGAN-FS-2: 0.06, 0.91, respectively ( Z values: 3.24-5.77, all P<0.005). Conclusion:The DL-based method combined with multiple strategies AttGAN-FS-2 shows improved denoising performance for low-dose brain PET images.