Dynamic PET Image Filtering Method Based On Deep Learning And MRI Structure Information

Positron emission tomography(PET)is a functional imaging that can observe metabolic activity at the molecular level in tissues.Dynamic PET imaging can not only observe the metabolic status over time,but also combine with the dynamic model to make quantitative analysis.Although the physical properties of PET continue to develop,dynamic PET scanning has high image noise due to the low count rate in short acquisition time each frame.Clinically,patients are often injected with high doses of tracers to improve image quality,which undoubtedly increases the risk of radiation exposure to patients and the costs of tracer drug.This thesis focuses on how to improve the quality of low count dynamic PET image.Gaussian filtering has been used in clinical practice due to its simplicity and the fast speed,but important structures in images are usually smoothed out.Bilateral filtering,as an extension of Gaussian filtering,takes into account the similarity of image space and neighborhood pixel values,which can preserve the edge while denoising the noisy image,but there are gradient flip artifacts in the edge.Guided filter can keep the edge and avoid artifacts while denoising image.However,this method may lead to blurring of the resulting image due to the coefficient averaging operation,and may introduce the wrong structure to the denoised image when the structure of the guide image and the noise image do not match.With the development of deep learning technology,deep neural network has been widely used in PET image denoising.Compared with traditional filtering methods,convolution neural network(CNN)has the better denoising effect and the faster speed.However,there are three challenges in the research process:(ⅰ)it is difficult to obtain high quality image labels of network;(ⅱ)A large amount of data is demanded for training to prevent overfitting;(ⅲ)Mean square error loss function often leads to blurred image results.In this thesis,a dynamic PET image filtering based on deep learning and MRI structure information was proposed using high-resolution magnetic resonance(MR)images as priori.Based on the linear representation model of spatial variation,the linear coefficients of guided filter were learned at the pixel level,and high-quality PET images are used as labels to avoid the problems of denoised image blurring and the introduction of wrong structures in guided filter.In addition,in order to solve the above challenges,the following three aspects were improved in this thesis:(ⅰ)Total frames of dynamic PET data were summed into one frame as a high-quality network labels.At the same time,the image labels were down-sampled to obtain the network training inputs;(ⅱ)Pre-training the network with a large number of simulation data,and finetuning the last two layers of the convolutional network with limited clinical data;(ⅲ)Based on the L1 norm,the network loss function introduced two edge preserving regularization terms to ensure that the network preserves the image details while denoising,and is not disturbed by the MR image structure.In this thesis,ten groups of dynamic PET/MR data were simulated based on BrainWeb in the simulation study,and eleven groups of dynamic PET/MR data from ADNI database were used for clinical study.The proposed method was compared with Gaussian filter,guided filter,CNN method with mean square error as loss function and its network structure is consistent with the proposed method.Simulation results show that the proposed method is superior to other methods in terms of root mean square error,structure similarity,ensemble variance and ensemble mean square bias.Due to the lack of clinical ground-truth image,we evaluated and compared the result images using local mean and standard deviation,and the results showed that the proposed method performed better denoising performance in a variety of local regions of interest.