研究目的
To explore the use of deep learning techniques, specifically convolutional neural networks, for improving the accuracy and efficiency of scatter estimation in positron emission tomography (PET) imaging.
研究成果
The study demonstrates the potential of deep learning techniques for accurate and efficient scatter estimation in PET imaging. The initial results from both convolutional neural networks are promising, though further improvements and more extensive training data are needed to fully realize their potential.
研究不足
The second network's performance showed a non-negligible difference from Monte Carlo simulated data, indicating the need for more training data and potential improvements in the network design.
1：Experimental Design and Method Selection:
The study involves the development of two convolutional neural networks (CNNs) for scatter estimation in PET. The first CNN estimates multiple scatter profiles from single scatter profiles, and the second CNN predicts total scatter profiles directly from emission and attenuation sinograms.
2：Sample Selection and Data Sources:
Training and validation data were generated using SimSET Monte-Carlo simulations from various phantoms, including cylinders of different sizes, an elliptical phantom, and the Zubal phantom.
3：List of Experimental Equipment and Materials:
GE Discovery D610 scanner geometry was simulated with BGO detector specifications. The neural networks were implemented using Caffe.
4：Experimental Procedures and Operational Workflow:
The first network was trained using one-dimensional radial profiles extracted from single and multiple scatter profiles. The second network was trained with emission and attenuation sinograms as inputs to predict total scatter profiles.
5：Data Analysis Methods:
The performance of the networks was evaluated by comparing the predicted scatter profiles with Monte Carlo simulated scatter signals, considered as the ground truth.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
