Search

CN-121639514-B - Low-dose PET image denoising reconstruction method based on flow model

CN121639514BCN 121639514 BCN121639514 BCN 121639514BCN-121639514-B

Abstract

The invention discloses a low-dose PET image denoising reconstruction method based on a flow model, which comprises the steps of obtaining a three-dimensional low-dose PET image Three-dimensional high-quality PET image corresponding to same Constructing a conditional flow model with a three-dimensional neural network architecture as a core, learning a conditional vector field, defining a denoising process as a normal differential equation with a low-dose PET image y as a condition, and describing data points in the following way The continuous track in the model is trained to obtain a network model, and the trained model is used for a new low-dose PET image y And (5) reasoning to obtain the denoised high-quality PET image. According to the low-dose PET image denoising reconstruction method based on the flow model, the network is trained by random noise and using the normal-dose PET image and the low-dose PET image which are clinically used, and the high-efficiency sampling technology is adopted in the flow model reasoning sampling stage to realize the rapid denoising reconstruction of the low-dose PET image.

Inventors

  • YU FENG
  • SHEN JIALE
  • SU XINHUI
  • SUN XIAONAN
  • XIE XUBIN
  • LUO WEI

Assignees

  • 浙江大学

Dates

Publication Date
20260512
Application Date
20260205

Claims (10)

  1. 1. The low-dose PET image denoising reconstruction method based on the flow model is characterized by comprising the following steps of: Acquisition of three-dimensional low dose PET images Three-dimensional high-quality PET image corresponding to same Constructing a pair training data set; preprocessing a training data set; constructing a conditional flow model taking a three-dimensional neural network architecture as a core, and learning a conditional vector field , wherein, For interpolation of the image, t is a continuous time variable, the denoising process is defined as the ordinary differential equation ODE conditioned on the low dose PET image y, describing the data points At the position of A continuous track within, starting from an a priori noise distribution at t=0, to a clean image distribution at t=1 Ending, the whole process is conditioned on the low dose PET image y; Model prediction with mean square error loss minimization using AdamW optimizers Vector field with ideal linear track Training to obtain a network model according to the difference between the two; For a new low dose PET image y, a trained model is used Reasoning is carried out, an index function controlled by a super parameter k is adopted to define a non-uniform sampling strategy of N non-uniform distribution time steps, and the image state is iteratively updated through a numerical integration method, so that a denoised high-quality PET image is obtained.
  2. 2. The method for denoising reconstruction of low dose PET image based on flow model according to claim 1, And carrying out standardized reconstruction and SUV unit conversion on the training data set, cutting the complete three-dimensional PET image into uniform size and extracting overlapped three-dimensional image blocks.
  3. 3. The method for denoising reconstruction of low dose PET image based on flow model according to claim 2, The standardized reconstruction involves converting the PET image into standard uptake value SUV units and unifying all images to 192 x 288 x 520 voxel sizes by clipping to remove unnecessary background areas.
  4. 4. The method for denoising reconstruction of low dose PET image based on flow model according to claim 3, The method for extracting the overlapped three-dimensional image blocks is to extract the image blocks with the sizes of 96 multiplied by 96 from the unified size image by taking 48 voxels as step sizes as training input so as to improve the utilization efficiency of the memory and keep the space continuity.
  5. 5. The method for denoising reconstruction of low dose PET image based on flow model according to claim 1, The conditional flow model is based on a three-dimensional U-Net architecture, and comprises 206.96M trainable parameters.
  6. 6. The method for denoising reconstruction of low dose PET image based on flow model according to claim 1, During training, random sampling time is used Noise and Data pair Constructing an interpolated image Calculating a loss function And updating model parameters by gradient descent The batch size was set to 4, the initial learning rate was 1×10 -4 , and the decay was to 1×10 -8 using the cosine annealing strategy.
  7. 7. The method for denoising reconstruction of low dose PET image based on flow model according to claim 1, In the non-uniform sampling strategy, time steps , Is determined by the following formula: When k is a negative value, the time steps of the region where t is close to 1 are more dense, so that the reconstruction precision of the fine structure is improved.
  8. 8. The method for denoising reconstruction of low dose PET image based on flow model according to claim 7, And N is set to be 2, and k= -15 is combined, so that two non-uniform time steps are concentrated at the tail end of the track, and high-fidelity reconstruction is realized and the calculated amount is remarkably reduced by matching with Euler method numerical integration.
  9. 9. The method for denoising reconstruction of low dose PET image based on flow model according to claim 1, In the reasoning stage, a low-dose PET image with the size of 192 multiplied by 288 multiplied by 520 is axially divided into 6 192 multiplied by 288 multiplied by 96 image blocks by adopting a sliding window method, 10 voxel overlapping is arranged between adjacent blocks, and random seeds are fixed in the same subject to ensure the consistency of the image block generation.
  10. 10. The method for denoising reconstruction of low dose PET image based on flow model according to claim 9, After each image block independently executes denoising, seamless splicing is realized by carrying out weighted average on the overlapped area, so that a denoising reconstructed PET image with a complete size is obtained, edge artifacts are avoided, and global consistency is improved.

Description

Low-dose PET image denoising reconstruction method based on flow model Technical Field The invention belongs to the field of medical image processing, and particularly relates to a low-dose PET image denoising reconstruction method based on a flow model. Background Positron Emission Tomography (PET) is a critical quantitative functional imaging modality. However, the use of low dose acquisitions in order to reduce patient radiation exposure and scan time clinically can lead to a dramatic decrease in image signal-to-noise ratio (SNR) and a serious degradation in image quality, which directly hampers accurate detection and reliable quantitative analysis of lesions. In order to solve this problem, a denoising reconstruction method based on deep learning has become the mainstream. However, the prior art paradigm is generally faced with a number of key challenges that prevent their wide clinical use, firstly, with drawbacks in reconstruction fidelity. The common approach represented by Convolutional Neural Networks (CNNs), while effective in denoising, tends to produce excessively smooth results that obscure the fine structure of the image and may mask microscopic lesions, directly affecting the accuracy of the diagnosis. While other approaches that attempt to preserve detail often involve the risk of training instability and introducing artifacts. Second, there is a serious bottleneck in computational efficiency. While advanced generation models such as Denoising Diffusion Probability Model (DDPMs) can realize high-fidelity reconstruction of the tip, their reasoning process is extremely slow-usually thousands of expensive iterative samples are required to generate one image. This time cost makes it completely unsuitable for clinical workflows requiring fast feedback. Finally, there is a limitation in model design, even a novel model with higher efficiency in theory, in practical application, a large number of sampling steps are still needed to ensure accuracy due to the dependence on an inefficient numerical sampling strategy, so that the theoretical speed advantage of the model cannot be realized. In summary, the existing method has difficulty in achieving both reconstruction fidelity and computational efficiency. Therefore, there is a strong need in the art for a new PET denoising reconstruction method that must efficiently process complete PET image data and employ an optimized sampling strategy to achieve high fidelity, high robustness reconstruction in a very short inference time. Disclosure of Invention The invention provides a low-dose PET image denoising reconstruction method based on a flow model, which solves the technical problems, and concretely adopts the following technical scheme: a low-dose PET image denoising reconstruction method based on a flow model comprises the following steps: Acquisition of three-dimensional low dose PET images Three-dimensional high-quality PET image corresponding to sameConstructing a pair training data set; preprocessing a training data set; constructing a conditional flow model taking a three-dimensional neural network architecture as a core, and learning a conditional vector field , wherein,For interpolation of the image, t is a continuous time variable, the denoising process is defined as the ordinary differential equation ODE conditioned on the low dose PET image y, describing the data pointsAt the position ofA continuous track within, starting from an a priori noise distribution at t=0, to a clean image distribution at t=1Ending, the whole process is conditioned on the low dose PET image y; Model prediction with mean square error loss minimization using AdamW optimizers Vector field with ideal linear trackTraining to obtain a network model according to the difference between the two; For a new low dose PET image y, a trained model is used Reasoning is carried out, an index function controlled by a super parameter k is adopted to define a non-uniform sampling strategy of N non-uniform distribution time steps, and the image state is iteratively updated through a numerical integration method, so that a denoised high-quality PET image is obtained. Further, the training data set is subjected to standardized reconstruction and SUV unit conversion, the complete three-dimensional PET image is cut into uniform size, and overlapping three-dimensional image blocks are extracted. Further, the standardized reconstruction includes converting the PET image into a standard uptake value SUV unit and removing unnecessary background regions by clipping, so that all images are unified to 192×288×520 voxel sizes. Further, the extracting of the overlapped three-dimensional image blocks is to extract the image blocks with the size of 96 multiplied by 96 from the uniform size image by taking 48 voxels as step sizes as training input, so as to improve the utilization efficiency of the memory and preserve the space continuity. Further, the conditional flow model is based on a three-dimens