Study On The Method Of Predicting Full-Dose PET Images From Low-Dose Based On Deep Learning

PET is a powerful molecular imaging technique in the field of nuclear medicine that presents the metabolic distribution of living tissues and cells in living organisms under the action of radioactive tracers,and is widely used in fields such as oncology and neurology.With the widespread use of PET imaging,the radiation hazard to human body during imaging and the discomfort caused by prolonged acquisition have attracted attention.However,decreasing the activity level of radiotracer injection or shortening the acquisition time can result in degradation of PET image quality.Therefore,how to improve the quality of low-dose PET images to meet the clinical diagnosis needs is the main research problem of this paper.This paper is mainly based on deep learning algorithms to achieve the goal of estimating full-dose PET images from low-dose PET images.In recent years,deep learning algorithms are more and more widely used in medical image processing,especially in denoising,enhancing or predicting full-dose PET images from low-dose PET images.Currently,most of the deep learning methods are used to predict full-dose PET images from low-dose ¹⁸F-FDG PET images and low-dose PET images acquired by conventional PET scanners.To address the above problems,two methods for low-dose estimation of full-dose PET images are proposed in this paper,which work on two main aspects.（1）A deep learning method for estimating full-dose PET images based on low-dose⁶⁸Ga-FAPI PET images is investigated.The commonly used radiotracer in practical clinical applications is ¹⁸F-FDG,whose uptake rate and image contrast are not obvious enough for some specific tumor entities.Therefore,this paper targets a novel radiotracer⁶⁸Ga-FAPI for PET imaging,and generates predicted full-dose PET images from its acquired low-dose ⁶⁸Ga-FAPI PET images.A modified U-Net network was used for feature information extraction of PET images to generate predicted full-dose PET images.Experiments were first conducted on different low-dose PET images,and the results showed that the image quality of predicted full-dose PET（Pre-33%）from 33%low-dose PET images could meet the clinical diagnosis,while the dual-center lesion detection rate of Pre-33%PET images could reach 96%.This means that reducing the standard dose to33%in practical clinical applications can also ensure the accurate localization and diagnosis of lesions.In addition,the method used in this work has certain advantages over other methods（e.g.,Red CNN）in terms of denoising and detail information recovery for low-dose PET images.The deep learning method used to predict full-dose PET images from low-dose PET images based on tracer ⁶⁸Ga-FAPI achieves good denoising effect and better recovery of structural detail information of low-dose PET images,which is a great contribution to the improvement of image quality and clinical diagnosis accuracy of low-dose PET images.（2）An ultra-low-dose PET imaging method based on Cycle GAN network is investigated.Conventional PET scans often require longer PET scanning time in clinical applications,and the image resolution is not high enough and whole-body PET imaging cannot be achieved.Therefore,in this paper,ultra-low-dose PET images acquired by advanced whole-body dynamic scanning u EXPLORER PET/CT are used to estimate full-dose PET images.A Cycle GAN-based network structure is used to learn the mapping relationship between the ultra-low-dose PET images and the full-dose PET images to generate the predicted full-dose PET images.Also,a perceptual loss function is introduced in the network training to guide the generator network to generate PET images with more edge and structural detail information.The effect of different low-dose PET images on the prediction results are explored,and the minimum injection activity required for PET scanning with sufficient detail information retained and able to meet the diagnostic requirements is investigated.The results show that the proposed Cycle GAN network can learn high-quality PET images from ultra-low-dose（5%）PET images,especially for the high metabolic regions of the predicted images,with high structural similarity to the full-dose PET images used as a standard reference,and is expected to be applied in real clinical practice to achieve lower radiation doses while improving the image quality of low-dose PET images and enhancing the high It is expected to be applied in actual clinical practice to reduce the radiation dose while improving the image quality of low-dose PET images and enhancing the accuracy of localization and diagnosis of metabolic regions（tumors）.In this paper,a deep learning method for estimating full-dose PET images from low-dose PET is investigated,which can help to achieve the reduction of radiation hazard while satisfying the diagnostic criteria for PET scans in clinical applications,which is a positive contribution to achieve better clinical services for low-dose PET.