Deep Learning Denoising in Histo-Projection TOF Data Non-Rigid Motion Estimation and Correction

Previously, we developed the method for non-rigid motion correction in the quasi-image space of histo-projections. The TOF locality property is used to locally perform motion correction; that is, the approximation of motion as locally rigid on a scale of TOF resolution. The motion vector field (MFV) is estimated from motion-sampled histo-images, which represents the most likely annihilation locations and is a combination of all histo-projection data. The histo-image is relatively noisy, compared with a full statistical reconstruction image from the same data set. Excessive noise leads to a suboptimal estimation of MVF when a standard optical flow method is applied. Nevertheless, histo-images may be denoised (and deblurred) based on target images of histo-projection data reconstruction, resulting in improved quality of histo-image generation.Deep Learning methodology is used in histo-image denoising, where computationally expensive statistical reconstruction from the same data provides training images. U-Net CNN, found to be effective in denoising PET images, is utilized.Biograph Vision Quadra (Siemens Healthineers) patient data was employed in investigations. This PET/CT scanner of long axial field of view (LAFOV) offers a very good time-of-flight (TOF) resolution of about 215 ps, suitable for quasi-image motion correction techniques. Data-driven motion detection is particularly attractive in LAFOV scanners since the entire patient is visible at once.The initial results showcased that DL histo-image denoising and deblurring improve MFV estimation. This procedure leads to less motion-blurry images in the case of respiratory motion correction, performed entirely in the TOF projection space.