Unsupervised Deep Learning Based Denoising For PET Static Image And Parametric Image

Positron emission tomography(PET)is a powerful imaging method in nuclear medicine.PET can detect the metabolic activity of molecules in living tissues,which is widely used in oncology,cardiology,and neurology.At present,due to the limitation of physical degradation factors and the restriction of the injection dose in real medical practice,PET image generally has the problem of low signal-to-noise ratio.Therefore,the main purpose of this thesis is to improve the quality of the PET image.In this thesis,we used the unsupervised deep learning technology for PET static image and parametric image denoising.In recent years,deep learning has been widely used in the field of medical imaging due to its superior performance.Common deep learning denoising methods need the use of low-quality/high-quality image pairs to train the network.But in real clinical practice,considering the limitation of dose,it is difficult to obtain high-quality PET images that can be used for training.Therefore,this thesis proposed an unsupervised deep learning denoising framework——Conditional Deep Image Prior(CDIP).Firstly,we apply the CDIP framework to a single patient's static PET image denoising prob-lem.We used the patient's noise PET image itself as the training label of the neural network and added the computed tomography(CT)or magnetic resonance(MR)image of the same patient as the anatomical prior to further improve the image quality.The denoising problem was transferred to the neural network training.The network output is the denoised image.Only using the noisy static PET image and the registered CT/MR image can achieve excellent denoising effects.In this method,we do not need pretraining or high-quality images as training labels.Then,we embedded the CDIP framework into a single patient's PET parametric image esti-mation.PET parameter image can represent the tracer binding rate calculated from the dynamic data of PET according to the kinetic model.Compared with the static PET image,it has higher contrast and specificity.We combined the unsupervised deep learning method with the Logan reference tissue model and used the ADMM algorithm to solve the optimization problem in three steps.Then we obtained the denoised parametric image.This is the first application of unsuper-vised deep learning to Logan parameter image estimation.In addition,considering the specificity of each individual patient,we tried to introduce pop-ulation data into the CDIP framework for static PET denoising.This method was combined by population training and individual fine-tuning.In the population training step,a group of patients'PET/anatomical image pairs were used to pre-train the network.In the individual fine-tuning step,the pre-trained network parameters are used as the initial parameters of the new network and some initial parameters are fixed to make better use of population information.Then the PET/anatomical image pair of a single patient is used to fine-tune the new network.The two networks have the same structure,with anatomical images as input and noise images as pretraining and fine-tuning labels.This method does not need low-quality/high-quality training pairs and can be implemented on the most commonly used PET/CT or PET/MR datasets.The innovation of this thesis is that this is the first work to utilize the unsupervised deep learning method for PET image processing.Our method can solve the problem that the existing deep learning method needs high-quality images' guidance,which is hard to get in real clinical practice.At the same time,our method can introduce anatomical prior(CT or MR)into the process of denoising,and further improve the quality of PET image.All of the above three denoising methods cover the common denoising problems in PET.