Deep Image Reconstruction for Reducing Limited-Angle Artifacts in a Dual-Panel TOF PET

Abstract

Dual-panel PET scanners have many advantages in dedicated breast imaging and on-board imaging applications since the compact scanners can be combined with other imaging and treatment modalities. The major challenges of dual-panel PET imaging are the incomplete sampling and data truncation, which can cause severe limited-angle artifacts. In this incomplete sampling case, time-of-flight (TOF) provides new information and thus reduces the artifacts, however, the problem is still quite challenging even with 200 to 300 ps timing resolution. In this work, we explore deep learning based image reconstruction for limited-angle artifacts reduction for dual-panel TOF PET imaging. The deep image reconstruction consists of two components, namely, TOF ordered subsets expectation maximization (OSEM) reconstruction, and a deep neural network for limited-angle artifacts reduction. We adopt and optimize a U-net based architecture for limited-angle artifacts reduction (LaU-net) to predict expected images from limited-angle TOF reconstructions. We perform numerical simulations with a generic 2D dual-panel TOF PET system with timing resolution of 300 ps and angular coverage of 90°. We generate 640 2D training datasets by performing TOF ordered subsets expectation maximization (OSEM) reconstructions from randomly generated phantom images. Then 3 additional folds of datasets were obtained using data augmentation by flipping horizontal and vertical dimensions for each dataset. We used Kullback-Leibler divergence as loss function for nonnegative images, and the Adam optimizer for training. We show from both random phantoms and a high resolution hot-rod phantom that the deep reconstruction can substantially reduce limited-angle artifacts and improve quantitative accuracy of reconstructed images.