Abstract
242
Objectives: Deep neural networks such as convolutional neural networks (CNN) have attracted growing interests in medical imaging such as positron emission tomography (PET) due to its high performances in computer vision tasks. In general, deep neural networks are needed to prepare the large size of training image datasets, however, it is not easy for clinical uses because of difficulty in preparing the large size of high-quality patient datasets. If some cases of diseases are not included in the training datasets, the trained network cannot exactly denoise the PET images of the patients. In addition, training datasets for novel PET ligands are impossible to prepare. Some recent report shows that CNN can get denoised image without training datasets. In this study, we proposed dynamic PET image denoising using CNN without training datasets. The advantage of this study is that training datasets are not necessary, as the original PET data of its own is used to reduce the statistical noise.
Methods: In the proposed method, it is not necessary to prepare additional image datasets. Static images of [18F]fluoro-2-deoxy-D-glucos was acquired the entire data from the start to end of the data acquisition, which is employed as the network input, and dynamic PET images are treated as the training label, and denoised dynamic PET images are represented by the network output with moderate iterations. 3D encoder-decoder architecture was used as the network structure and implemented using TensorFlow and Keras libraries. Mean square error was the optimized loss function and the network was optimized by stochastic gradient descent. In the evaluation, computer simulation data and real monkey brain data were used. We measured the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) index. This study was approved by the ethical committee of the Central Research Laboratory, Hamamatsu Photonics K.K. (HPK-2017-02).
Results: The proposed denoising method can improve the denoising performance compared with the other non-deep learning algorithms. In the computer simulation, the PSNR and SSIM of the proposed method is better than the Gaussian post filtering and guided image filtering in all time frames. In addition, the regional time activity curves of the real monkey brain data treated by the proposed method are smoother than the Gaussian post filtering and guided image filtering.
Conclusions: The simulation results of the present study suggest that the proposed method improves the quantitative accuracy in the kinetic analysis. In addition, the results of the real monkey brain data show the same trend as simulation results. These results indicate that the proposed method is a practical algorithm for dynamic PET data.