CN-118941718-B - Three-dimensional PET image reconstruction method, system and storage medium based on diffusion multi-scale generation countermeasure network

CN118941718BCN 118941718 BCN118941718 BCN 118941718BCN-118941718-B

Abstract

The invention discloses a three-dimensional PET image reconstruction method, a system and a storage medium based on a diffusion multiscale generation countermeasure network, wherein the method comprises the steps of inputting a PET image to be reconstructed into a trained diffusion multiscale generation countermeasure network DMGAN to generate a near-real full-dose PET image, the diffusion multiscale generation countermeasure network comprises a diffusion generator and a U-Net discriminator, the PET image to be reconstructed, namely a low-dose L-PET image, is input into the diffusion generator after being sliced to synthesize the full-dose image, namely an F-PET image, the full-dose image comprises a series of corresponding target slices, a noisy F-PET image is generated at the same time, the generated image is input into the U-Net discriminator, and details are extracted from global and local views to improve the quality of the generated F-PET image. The method can effectively capture and restore image details, so that the generated image is similar to a real full-dose PET image in vision and structure.

Inventors

ZHANG HONG
YU XIANG
ZHONG YAN
WANG JING
JIN CHENTAO
Hu Daoyan

Assignees

浙江大学

Dates

Publication Date: 20260505
Application Date: 20240819

Claims (7)

1. A three-dimensional PET image reconstruction method based on a diffusion multiscale generation countermeasure network is characterized by comprising the following steps: inputting a PET image to be reconstructed into a trained diffusion multiscale generation countermeasure network DMGAN to generate a near-real full-dose PET image; The diffusion multiscale generation countermeasure network comprises a diffusion generator and a U-Net discriminator, wherein a PET image to be reconstructed, namely a low-dose L-PET image or a slice thereof, is input into the diffusion generator to synthesize a full-dose image, namely an F-PET image, and a noisy F-PET image is generated at the same time, the generated image is input into the U-Net discriminator, and details are extracted from global and local views to improve the quality of the generated F-PET image; the loss function in the diffusion generator is: , wherein x represents an L-PET image, G (x) represents an F-PET image generated by the generator, and y is a corresponding original F-PET image; Charbonnier Loss is introduced to penalize euclidean differences between the generated F-PET image and the original F-PET image: , at the same time, charbonnier Loss penalties are introduced to generate Euclidean differences between the noisy F-PET image and the original F-PET image: , furthermore, taking into account the perceived differences between the generated F-PET image, the generated noisy F-PET image, and the original F-PET image, the feature representations of G (x) and y are extracted using VGG16-Net trained on ImageNet with the perceived loss: , wherein V represents a feature map extracted by VGG 16-Net.
2. The method of three-dimensional PET image reconstruction based on a diffusion multiscale generation countermeasure network according to claim 1, wherein prior to reconstruction, both the training dataset L-PET and F-PET images are preprocessed using statistical parameter mapping for realignment and normalization.
3. The method for reconstructing a three-dimensional PET image based on a diffusion multiscale generating countermeasure network according to claim 1, wherein the U-Net discriminator comprises a downsampling network and an upsampling network connected by a skip-connection and a bottleneck connection bottleneck.
4. A three-dimensional PET image reconstruction method based on a diffusion multiscale generation countermeasure network according to claim 3, wherein the U-Net discriminator performs discrimination on a per pixel z basis, and the encoder loss is: , The loss of the decoder for the average of all pixel values is: , and Respectively representing the discrimination of the discriminator at the pixel (i, j); Is obtained by jumping connections from the middle layer of the encoder network, combining the specific details of the low-level features and global information of the high-level features obtained by upsampling from the upsampling layer; considering the above-mentioned loss function, the goal of the generator is to: , the encouragement generator focuses on the composite image by effectively capturing global structure and local detail, aiming to fool the discriminant more effectively; The base loss of the discriminator is: , the U-Net discriminator returns two values representing the output of the decoder and the encoder respectively, and the intermediate loss is used for describing the loss of the encoder of the U-Net discriminator, which is expressed as: , The total loss of the discriminator is: , where v is a superparameter.
5. A three-dimensional PET image reconstruction system based on a diffuse multiscale generation countermeasure network for implementing the method of any of claims 1-4.
6. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the method according to any of claims 1-4.
7. An electronic device, the device comprising: One or more processors; A memory for storing one or more programs; The one or more programs, when executed by the one or more processors, cause the one or more processors to implement the method of any of claims 1-4.

Description

Three-dimensional PET image reconstruction method, system and storage medium based on diffusion multi-scale generation countermeasure network Technical Field The invention relates to the technical field of medical image engineering, in particular to a three-dimensional PET image reconstruction method, a system and a storage medium based on a diffusion multiscale generation countermeasure network. Background PET, one of the most widely used medical imaging techniques, plays a key role in navigator surgery, medical evaluation, and clinical examination, and unlike other imaging techniques such as magnetic resonance imaging and computed tomography, PET is capable of detecting biochemical and physiological changes. PET is also widely used for prophylactic treatment and early disease recognition, as biochemical and physiological changes usually precede anatomical changes. PET can assess human molecular changes in vivo. Despite the significant advantages of PET, there is an increasing concern about the health risks that radiation exposure may carry during scanning. For example, in clinical practice, the injected dose is often limited by the radiation dose, as more radiation dose is likely to increase the risk of cancer and cause some degree of damage to the body. Therefore, low-dose L-PET capable of image acquisition with minimal radiation exposure is of great interest to researchers. However, L-PET images exhibit higher noise levels, reduced image contrast, and more artifacts than full-dose F-PET images, making accurate diagnosis difficult for a physician. Therefore, it is important to acquire a high quality image from a low dose image to minimize image exposure while maintaining image quality. In order to improve the quality of PET images, many methods have been proposed. One way to achieve high quality PET images is to integrate a priori information into the image reconstruction process. This method allows direct incorporation of imaging physical information. However, it faces challenges associated with intensive computation and must access the physical projection model. Many studies have discussed voxel estimation methods after image reconstruction. These methods include a random forest based regression method, a mapping based sparse representation method, a semi-supervised triple dictionary learning method, and a multi-level canonical correlation analysis framework. While these prior methods have demonstrated promising results, they tend to produce excessively smooth images. In recent years, deep learning methods have been widely studied in the field of medical imaging. The generation of countermeasure and convolutional neural networks has proven successful in low dose CT image denoising. Since deep learning is widely used in various fields, it also has a significant impact in the task of L-PET imaging. Xiang et al propose a complex convolutional neural network model with automatic context learning that predicts high dose PET images using only 1/4 dose of full dose PET images and corresponding MR T1 images. Wang et al developed a comprehensive framework for generating an countermeasure network (GANs) using 3D conditions to generate high quality PET images from corresponding L-PET images. Kaplan and Zhu propose a model that incorporates specific image features into the loss function to denoise 1/10 dose full dose PET image slices and estimate their full dose counterparts. Chen et al propose methods for synthesizing high quality and accurate PET images using PET data alone or in combination with PET and MR information. Ouyang et al propose that generating an antagonizing network can achieve similar performance levels without MR information. Recently, yu et al introduced a simplified L-PET image reconstruction framework that enabled fast generation of F-PET images and improved the overall quality of the final 3D F-PET images using the spatial details of the generated F-PET slices. However, there are still some limitations, such as how to enhance understanding semantic information in L-PET reconstruction from different perspectives, etc. Disclosure of Invention The invention provides a three-dimensional PET image reconstruction method, a system and a storage medium based on a diffusion multiscale generation countermeasure network, and aims to solve the problems. Comprising a diffusion generator and a U-Net discriminator. Experiments prove that the method provided by the invention has advantages compared with other L-PET image reconstruction methods. In order to achieve the above purpose, the technical scheme adopted by the invention is as follows: A three-dimensional PET image reconstruction method based on a diffusion multiscale generation countermeasure network comprises the following steps: inputting a PET image to be reconstructed into a trained diffusion multiscale generation countermeasure network DMGAN to generate a near-real full-dose PET image; the diffusion multiscale generation countermeasure network c