Multi-Tracer PET Image Reconstruction And Artificial Intelligence-Assisted Medical Imaging Application

Positron Emission Tomography（PET）is a mature and advanced molecular imaging technology for nuclear medicine,which can carry out qualitative or quantitative studies at the living molecular or cellular level of organisms.The radiotracer drugs for PET are diverse,and different tracers can characterize the uptake differences of different diseased cells.Compared with other imaging techniques,PET imaging has high sensitivity and high specificity,so it occupies an important position in many medical imaging.However,due to factors such as the hardware limitations of PET scanners,the dose safety of the tracer and the absorption ratio of the tracer by tissue cells,the resolution of PET imaging is often low,especially for some tracers with short half-lives.The scan delay is shorter and the imaging noise becomes more obvious.In addition,a single radioactive tracer is commonly used for imaging in clinical practice,but each tracer can only characterize information in one cell,which may lead to false-negative or false-positive diagnosis.On the other hand,dynamic PET parametric imaging with higher quantitative accuracy is difficult to be used in clinical practice due to the longer scanning time and the need for invasive sampling for some blood input functions.Therefore,PET imaging has been an area of great concern.With the promotion of the artificial intelligence wave in recent years,deep learning has been gradually applied in the field of medical imaging.In view of the above imaging problems,this paper proposes a multi-tracer PET image reconstruction algorithm and a medical imaging application technology based on deep learning.The current relatively mature algorithm frameworks are improved and applied to multi-tracer PET and dynamic PET,achieving high-quality PET imaging and effectively preserving specific tumor area information.For the research content,the main research work and innovation points of this paper are as follows:（1）A multi-tracer-based PET image reconstruction algorithm is proposed.Clinically,for some specific diseases,multiple PET tracers are often required for characterization,thereby improving the sensitivity and specificity of identifying diseased tissue.Therefore,this paper proposes for multi-tracer PET to use the PET image of one tracer with higher imaging quality（usually longer half-life）as prior information to guide another tracer with lower imaging quality（usually shorter half-life）PET imaging.The implementation is based on the popular and mature kernel method in recent years for feature guidance,and the kernel matrix of prior information is constructed through k nearest neighbor search to guide image reconstruction.The introduction of high-quality tracer PET can greatly enrich the metabolic information of the tumor region to be reconstructed.Verified by simulation and clinical experiments,the proposed method can effectively reconstruct the rich functional metabolic information of multi-tracer PET while image denoising and edge detail preservation.Especially in clinical trials,the proposed method achieves a tumor signal-to-noise ratio of 19.70 d B,which is 28.3% higher than other compared algorithms（Gaussian and non-local means smoothed maximum likelihood expectation maximization,MR-guided kernelized expectation maximization（KEM）,and multi-guided-S KEM algorithms）.In conclusion,the proposed multi-tracer PET imaging method can maximize the advantages of PET imaging,thereby improving the accuracy of PET imaging in clinical diagnosis.（2）A dynamic PET parametric imaging method based on a generative network model is proposed.Studies have shown that dynamic PET parametric images achieve more accurate quantitative performance than clinical routine static PET,and show higher specificity in characterizing malignant tumors.However,due to the long scanning time of dynamic PET,it is difficult to be universally used in clinical practice.Therefore,this paper proposes a dynamic parametric imaging method based on conventional static PET scans for dynamic PET.The adversarial generation network in deep learning is used to learn the mapping relationship between conventional PET and dynamic parametric images,and the clinical experience of the two imaging methods is considered to select the appropriate loss function,then the static PET images can be used to directly generate dynamic PET parametric images.After model training and clinical evaluation,the results show that the proposed method can learn high-quality dynamic parameter images from conventional PET,especially in the malignant tumors and high metabolic areas of images.The structural similarity can reach 0.86,the peak signal-to-noise ratio achieves 23.64 d B,and the root mean square error is 0.08.In addition,the synthesized parametric image effectively combines molecular imaging and quantitative imaging,which can simultaneously reflect the concentration state and metabolic rate of the tracer in the tissue,providing an additional and more reliable reference for clinical practice and helping to improve the accuracy of malignant tumor diagnosis.The multi-tracer PET image reconstruction algorithm and artificial intelligence-assisted medical imaging application proposed in this paper achieve high-quality imaging through the kernel algorithm and deep learning method,especially improve the ability of PET imaging to characterize diseased tissue,which is helpful to the accuracy of clinical diagnosis,and has a certain promotion significance for the application of multi-tracer PET and dynamic PET in clinical practice.