DirectPET: Full Size Neural Network PET Reconstruction from Sinogram Data

Neural network image reconstruction directly from measurement data is a relatively new field of research, that until now has been limited to producing small single-slice images (e.g., 1x128x128). This paper proposes a new more efficient network design called DirectPET which is capable of reconstructing a multi-slice Positron Emission Tomography (PET) image volume (i.e., 16x400x400) by addressing the computational challenges through a specially designed Radon inversion layer. We compare the proposed method to the benchmark Ordered Subsets Expectation Maximization (OSEM) algorithm using signal-to-noise ratio, bias, mean absolute error and structural similarity measures. Line profiles and full-width half-maximum measurements are also provided for a sample of lesions. The analysis shows that DirectPET is able to produce images that are quantitatively and qualitatively similar to the OSEM target images used during training. Additionally, we conduct an experiment where DirectPET is trained to map low count raw data to high count target images, which is a common application of image-space PET deep learning methods. The results show that DirectPET is able to maintain the image quality seen when trained to map high count raw data.