Deep Learning Accelerates Accurate Scatter Correction with Histo-image in TOF PET/CT System

Benefited from the hardware upgrade, the time-of-flight (TOF) resolution of PET/CT scanners has been significantly improved, which can accurately locate annihilation events and obtain high quality histo-image. On the other hand, scatter correction methods usually rely on iterative calculations to obtain accurate emission distribution, which is time-consuming. In this work, a deep learning assisted scatter correction algorithm was proposed. This algorithm generates the emission distribution from the TOF histo-image with deep neural network, and directly estimates the scatter distribution through the Monte Carlo simulation without further iterations of image reconstruction and scatter estimation. There were totally 205 patient data in the training, validation and test datasets, which were all acquired by the uMI Panorama PET/CT with sub 200 ps TOF resolution. When comparing the proposed algorithm to the ground-truth scatter correction method by taking the clinical reconstructed image as the emission map, the NRMSE of the scatter sinograms was less than 8%, and the organ-based SUV difference of the reconstructed images was less than 2%, except for the stomach (less than 8%). The computation time of scatter estimation was reduced by 78.5 % for non-TOF and 65.3% for TOF protocol, comparing to the clinically applied iterative Monte Carlo scatter simulation. These results prove that our method can significantly accelerate the scatter correction with equivalent accuracy.