Attention mechanism-based low-dose dual-tracer PET reconstruction method

A method, comprising:

• (1) performing dynamic PET scanning on biological tissues injected with a standard-dose tracer I to obtain a PET Standard-dose Single-tracer Sinogram (SSS1) corresponding to the tracer I;
• and performing dynamic PET scanning on biological tissues injected only with a standard-dose tracer II to obtain a PET Standard-dose Single-tracer Sinogram (SSS2) corresponding to the tracer II at an interval of 5 half-lives of the tracer I, wherein the standard-dose generally refers to a dose that is able to reconstruct an activity map using a traditional reconstruction algorithm;
• (2) calculating, by a PET reconstruction algorithm, PET Standard-dose Single-tracer Activity Maps 1 and 2 (SSA1 and SSA2) corresponding to the SSS1 and the SSS2, respectively;
• (3) normalizing and superimposing the SSS1 and the SSS2 to form a PET Standard-dose Dual-tracer Sinogram (SDS);
• down-sampling the SSS1 and the SSS2 respectively to obtain PET Low-dose Single-tracer Sinograms 1 and 2 (LSS1 and LSS2), and superimposing the LSS1 and the LSS2 to form a PET Low-dose Dual-tracer Sinogram (LDS);
• and normalizing and superimposing the SSA1 and the SSA2 to form a PET Standard-dose Dual-tracer Activity Map (SDA);
• (4) repeating (1) to (3) several times to obtain a plurality of samples, and dividing these samples into a train set, a validation set and a test set, where each set of samples comprises the SSS1, the SSS2, the SDS, the LDS, the SSA1, the SSA2 and the SDA;
• (5) building an attention mechanism-based network model comprising a sinogram part and a reconstruction part, inputting the LDS in the samples of the train set into the sinogram part, and serving the SSS1 and the SSS2 in the corresponding set of samples as labels of the sinogram part, so as to pre-train the sinogram part;
• (6) cascading the pre-trained sinogram part with the reconstruction part, inputting the LDS in the samples of the train set into the sinogram part, serving the SSS1 and the SSS2 in the corresponding set of samples as the labels of the sinogram part and the SSA1 and the SSA2 as labels of the reconstruction part, so as to train the entire network model;
• and
• (7) inputting the LDS in the samples of the test set into the trained network model to achieve low-dose dual-tracer PET image reconstruction and obtain corresponding reconstructed SSS1, SSS2, SSA1 and SSA2.