PET breathing motion image artifact registration correction method of motion manifold decomposition model based on multi-head attention

The invention relates to a PET breathing motion image artifact registration correction method based on a motion manifold decomposition model, and belongs to the technical field of nuclear medicine imaging. After PET clinical data and simulated PET data after three-dimensional reconstruction are obtained, the obtained image is input into a motion manifold decomposition model based on multi-head attention to obtain a multi-dimensional local registration manifold domain, and the multi-dimensional local registration manifold domain generates a manifold domain of the highest dimension in a competitive weighting mode. And the distortion is applied to a spatial transformation network (STN) to carry out distortion reduction on the PET image, so that the effect of correcting the artifacts of the respiratory movement is achieved. Wherein the multi-head attention-based motion manifold decomposition model is used for reducing the motion gradient change of the PET image pair through a stepped manifold learning mechanism,