ESR-Net: An Efficient Image Super-resolution Network for SPECT Reconstruction

SPECT image reconstruction is challenging due to the poor spatial resolution of the SPECT camera. Incorporating machine learning (ML) based regularizers into reconstruction algorithms is receiving more attention recently. However, most previous works assume the true images used for training have the same voxel size as the reconstructed images, which can be suboptimal when training images with finer voxel sizes are available. Directly using ML-based algorithms with fine voxel sizes could be very computationally expensive due to the heavy computations involved in forward and backward projections. This paper proposes a novel, efficient image super-resolution reconstruction network (ESR-Net) that can improve the resolution by training the ML based regularizer using true activity maps having finer voxel sizes, while maintaining the computational efficiency by computing the forward and backward projections in coarser voxel sizes through downsampling and upsampling. Simulation results based on digital XCAT phantoms demonstrated that the proposed ESR-Net significantly outperformed other methods (OSEM and BCD-Net), when evaluated qualitatively by visualizing reconstructed images and line profiles, as well as quantitatively by mean recovery coefficients (MRC) and normalized root mean square error (NRMSE).