Generation of PET Attenuation Map for Whole-Body Time-of-Flight 18F-FDG PET/MRI Using a Deep Neural Network Trained with Simultaneously Reconstructed Activity and Attenuation Maps
Generation of PET Attenuation Map for Whole-Body Time-of-Flight 18F-FDG PET/MRI Using a Deep Neural Network Trained with Simultaneously Reconstructed Activity and Attenuation Maps
复制标题
DOI：
10.2967/jnumed.118.219493
复制
发表时间：
2019-08-01
影响因子：
9.3
通讯作者：
Lee, Jae Sung
中科院分区：
文献类型：
--
作者：
Hwang, Donghwi;Kang, Seung Kwan;Lee, Jae Sung
我们提出了一种新的基于深度学习的方法，以提供比基于Dixon的4段方法更准确的全身PET/MRI衰减校正。我们使用使用活动和衰减的最大似然重建（MLAA）算法估计的活动和衰减图作为卷积神经网络（CNN）的输入来学习CT衍生的衰减图。研究方法：100名癌症患者（38名男性和62名女性;年龄，57.3 ± 14.1岁）的全身F-18-FDG PET/CT扫描数据被回顾性地用于训练和测试CNN。训练修改的U-网络以从MLAA生成的活动分布（MLAda-MLAA）和mu-图（mu-MLAA）预测CT衍生的mu-图（mu-CT）。我们使用了来自60名患者数据的130万个补丁来训练CNN，另外20名患者的数据被用作验证集以防止过度拟合，另外20名患者的数据被用作CNN性能分析的测试集。将使用所提出的方法（mu-CNN）、mu-MLAA和4段法（mu-segment）生成的衰减图与mu-CT（地面实况）进行比较。我们还比较了使用有序子集期望最大化与mu图重建的活动图像之间的体素相关性，以及通过在活动图像上绘制感兴趣区域获得的原发性和转移性骨病变的SUV。结果：CNN生成的噪声衰减图更少，并且比MLAA实现了更好的骨识别。mu-CNN和mu-CT之间骨骼区域的平均Dice相似系数为0.77，显著高于mu-MLAA和mu-CT之间的相似系数（0.36）。此外，CNN结果显示了与基于CT的结果的最佳逐像素相关性，并且与基于CT的衰减校正相比，显著减少了活动图的差异。结论：拟议的深度神经网络为511 keV光子产生了比目前用于全身PET/MRI研究的4段方法更可靠的衰减图。
We propose a new deep learning-based approach to provide more accurate whole-body PET/MRI attenuation correction than is possible with the Dixon-based 4-segment method. We use activity and attenuation maps estimated using the maximum-likelihood reconstruction of activity and attenuation (MLAA) algorithm as inputs to a convolutional neural network (CNN) to learn a CT-derived attenuation map. Methods: The whole-body F-18-FDG PET/CT scan data of 100 cancer patients (38 men and 62 women; age, 57.3 +/- 14.1 y) were retrospectively used for training and testing the CNN. A modified U-net was trained to predict a CT-derived mu-map (mu-CT) from the MLAA-generated activity distribution (lambda-MLAA) and mu-map (mu-MLAA). We used 1.3 million patches derived from 60 patients' data for training the CNN, data of 20 others were used as a validation set to prevent overfitting, and the data of the other 20 were used as a test set for the CNN performance analysis. The attenuation maps generated using the proposed method (mu-CNN), mu-MLAA, and 4-segment method (mu-segment) were compared with the mu-CT, a ground truth. We also compared the voxelwise correlation between the activity images reconstructed using ordered-subset expectation maximization with the mu-maps, and the SUVs of primary and metastatic bone lesions obtained by drawing regions of interest on the activity images. Results: The CNN generates less noisy attenuation maps and achieves better bone identification than MLAA. The average Dice similarity coefficient for bone regions between mu-CNN and mu-CT was 0.77, which was significantly higher than that between mu-MLAA and mu-CT (0.36). Also, the CNN result showed the best pixel-by-pixel correlation with the CT-based results and remarkably reduced differences in activity maps in comparison to CT-based attenuation correction. Conclusion: The proposed deep neural network produced a more reliable attenuation map for 511-keV photons than the 4-segment method currently used in whole-body PET/MRI studies.