Low-dose Direct PET Image Reconstruction Using Channel Attention for Deep Neural Network

Abstract

Positron emission tomography (PET) is a medical imaging approach widely used in various clinical applications. There is significant value in low-dose PET image reconstruction, because radiation risk is reduced when patients are injected with lower dose of radiotracer. However, this results in a high level of noise in emission data, which degrades the quality of activity distribution images. In this paper, we propose a deep neural network for low-dose PET reconstruction. Using time-of-flight (TOF) sinograms as inputs, it generates high-quality quantitative PET images directly. Specifically, we utilize an encoder-decoder to transfer projections in sinogram domain to activity maps in image domain. Then the outputs of previous stage are restored using a deep neural network with channel attention modules. Residual connections allow abundant low-level features to be bypassed, while channel attention blocks (CABs) capture high-level features by extracting channel statistics. We inject supervision to both the initial output after domain transformation and the final output. The loss function is comprised of the mean square error (MSE) of two outputs and their edge losses. The qualitative and quantitative results demonstrate that the proposed approach is capable of preserving fine details. This method shows promise in improving PET image quality with low-dose emission data.