(1) collecting a computed tomography (CT) data of a scanning object, then injecting a tracer into the scanning object, and using a PET system to scan and collect the sinogram data of the PET;

(2) converting the CT image into an attenuation image at 511 KeV energy;

(3) reconstructing a non-attenuation-corrected PET image based on the sinogram data, and then using the PET image to calculate a non-attenuation-corrected standardized uptake value (SUV) image;

(4) based on the sinogram data and the CT attenuation image, reconstructing an attenuation-corrected PET image, and using the PET image to calculate an attenuation-corrected SUV image;

(5) establishing a flow model for attenuation correction, taking the non-attenuation-corrected SUV image as input, and using the attenuation-corrected SUV image as a ground truth label to train the flow model; and

(6) performing attenuation correction on the non-attenuation-corrected PET image reconstructed from the sinogram data by using the well-trained flow model;

wherein, the flow model is composed of a cascade of multiple reversible modules, and the reversible module consists of a reversible 1×1 convolutional layer and connected with a enhanced affine coupling layer; and

wherein, the 1×1 convolutional layer is used to realize the permutation operation, that is, to disrupt the channel order of the input image, and using the matrix W to multiply the input image and then the output is provided to the affine coupling layer, the expression of the matrix W is as follows:

wherein, P is a permutation matrix, L is a lower triangular matrix with diagonal elements of 1, U is an upper triangular matrix with diagonal elements of 0, and v is a set of numerically learnable vectors.