Neural Architecture Search for Unsupervised PET Image Denoising

Unsupervised learning methods effectively reduce the noise level of positron emission tomography (PET) images with limited training data. Recent research indicates that the performance of these methods is greatly influenced by the network architecture. However, there has been a lack of investigation into the optimal network architecture for unsupervised PET imaging in previous studies. To address this gap, we developed a neural architecture search method to search for a better network architecture for unsupervised PET image denoising tasks. Our approach searches the network architecture in two separate spaces: the network-level search space and the cell-level search space. Continuous relaxation techniques are utilized to reduce time consumption during the search process. In our proposed framework, high-count PET images were used to search the network architecture, while low-count PET images were used to optimize operation parameters. After identifying the optimal network architecture, we evaluated its performance on phantom data and patient data with a variety of tracers. Our experimental results demonstrated that the searched network outperformed other methods.