Self-normalization for a 1-mm 3 resolution clinical PET system using deep learning.

Normalization in positron emission tomography (PET) corrects for non-uniformity of sensitivity across all system lines of response (LOR). Self-normalization is a framework that aims to estimate normalization components from the emission data without a separate scan of a normalization phantom. In this work, we propose for the first time an image-based end-to-end self-normalization framework using conditional generative adversarial networks (cGAN). We evaluated different approaches by exploring each of the following three methodologies. First, we used images that were either unnormalized or corrected for geometric factors, which encompass all time-invariant factors, as input data types. Second, we set the input tensor shape as either a single axial slice (2-D) or three contiguous axial slices (2.5-D). Third, we chose either Pix2Pix or polarized self-attention (PSA) Pix2Pix, which we developed for this work, as a deep learning network. The targets for all approaches were the axial slices of images normalized using the direct normalization method. We performed Monte Carlo simulations of ten voxelized phantoms with the SimSET simulation tool and produced 26,000 pairs of axial image slices for training and testing. The results showed that 2.5-D PSA Pix2Pix trained with geometric-factors-corrected input images achieved the best performance among all the methods we tested. All approaches improved general image quality figures of merit peak signal to noise ratio (PSNR) and structural similarity index (SSIM) from ~15% to ~55%, and 2.5-D PSA Pix2Pix showed the highest PSNR (28.074) and SSIM (0.921). Lesion detectability, measured with region of interest (ROI) PSNR, SSIM, normalized contrast recovery coefficient (NCRC), and contrast-to-noise ratio (CNR), was generally improved for all approaches, and 2.5-D PSA Pix2Pix trained with geometric-factors-corrected input images achieved the highest ROI PSNR (28.920) and SSIM (0.973).