Search

CN-122023549-A - CT imaging reconstruction method and system based on residual poisson flow generation model

CN122023549ACN 122023549 ACN122023549 ACN 122023549ACN-122023549-A

Abstract

A CT imaging reconstruction method and system based on residual poisson flow generation model, the method improves the existing poisson flow generation model, conditional input is introduced, the poisson flow generation model is conditioned by a component, noise disturbance is added to a target full-angle label image x gt to form a sample x t , the disturbed image and a preliminary reconstruction image x sp are taken as paired data, the conditional poisson flow generation model is input together for training, the model is optimized through weighted mean square error, mapping from x sp to x gt is learned, after the conditional poisson flow generation model is obtained, the model is subjected to gradual iterative sampling by a numerical solution method, each iteration comprises physical consistency correction and residual linear fusion, and finally an output image realizes good balance between texture reduction and physical consistency, so that reconstruction quality under a sparse view angle is remarkably improved.

Inventors

  • SHI YU
  • Fang Changcheng
  • FAN SHUYI

Assignees

  • 苏州工学院

Dates

Publication Date
20260512
Application Date
20250930

Claims (10)

  1. 1. The CT imaging reconstruction method based on the residual poisson flow generation model is characterized by comprising the following steps of: Preliminary image reconstruction step S110: Obtaining a preliminary reconstructed image x sp by using a traditional computer tomography reconstruction method; conditional poisson flow generation modeling step S120: Taking the preliminary reconstructed image x sp obtained in the step S110 as a conditional input, representing as conditional posterior distribution sampling, enabling the generated model to be corrected and refined on the basis of the existing poisson flow generated model structure reference, correcting an evolution equation into a conditional ordinary differential equation, and modeling to obtain a conditional poisson flow generated model; model training step S130: Adding noise disturbance to a target full-angle label image x gt to form a sample x t , taking the disturbed image and a preliminary reconstructed image x sp as pairing data, inputting the paired data into the conditional poisson flow generation model for training, outputting a mapping from x sp to x gt by the model, outputting the model as a direction vector pointing to a real image, and optimizing through weighted mean square error; CT image reconstruction step S140: Taking the preliminary sparse angle reconstruction image x sp as a conditional input, calculating a prediction result by a trained conditional poisson flow generation model For the prediction result Physical consistency correction is carried out to obtain a physical consistency result And to compare the prediction result And physical coherence results Weighted summation is carried out to obtain a fusion result With the fusion result And (3) as a new sampling point, reinjecting the normal differential equation updating process of the conditional poisson flow generation model, entering the next iteration, and obtaining a sparse view CT reconstruction result after the iteration is finished.
  2. 2. The CT imaging reconstruction method as recited in claim 1, wherein: step S110 is: Projection data at a sparse angle is expressed as formula (1): y sp =m (=) ax+n formula (1) Where M (Λ) is a sampling operator defined by an angle sub-sampling mask Λ, Λ is a full angle system projection operator, n is logarithmic domain noise, x is an imaged object, y sp is sparse view projection, solving by using regularized least squares as equation (2) to obtain a preliminary reconstructed image x sp , Wherein, the Is a priori constraint of artificial construction, ω being its weight.
  3. 3. The CT imaging reconstruction method as recited in claim 2, wherein: step S120 is: Taking the preliminary reconstructed image x sp as a conditional input, the mathematical modeling of the conditional posterior distribution sampling can be expressed as equation (3): Wherein, p (x) represents the prior distribution of the full-angle CT image and reflects the statistical rule of the real image, p (x sp -x) represents the process of degenerating the full-angle image into a sparse view angle and reflects the constraint of the measurement condition, x sp is the low-quality reconstructed image obtained in the step S110; And (3) generating a final reconstruction result obtained under the model for the condition. Obtaining a conditional poisson flow model as described in equation (4): The generation process takes x sp as a conditional image in each step, hijacking the sampling process, wherein x is a data point to be generated or reconstructed, r is an expansion radius variable, θ is a model parameter to be trained, formula (4) is a normal differential equation, and the sampling process is solved by a numerical solution after the conditional poisson flow model is modeled.
  4. 4. A CT imaging reconstruction method as claimed in claim 3, wherein: In the step S130 of the above-mentioned process, The training process is described by the formula (5): wherein x ti is a disturbance sample, x sp is a preliminary reconstructed image, x gt is a full-angle label image, sigma is a noise scale, and model parameters theta obtained through training can be used in a conditional poisson flow generation process to improve the quality of a final reconstructed image.
  5. 5. The CT imaging reconstruction method as recited in claim 4, wherein: in the step S130 of the process of the present invention, The model parameters theta obtained through training are that an Adam optimizer is adopted to update the parameters theta of the conditional poisson flow generation model, and the weighted mean square error is utilized to optimize until convergence.
  6. 6. The CT imaging reconstruction method as recited in claim 5, wherein: In the step S140 of the process of the present invention, The pair of prediction results Physical consistency correction is carried out to obtain a physical consistency result The physical consistency result is obtained by adopting the formula (6) The method comprises the steps of A is a full-angle projection operator, M (Λ) is a sparse angle sampling operator, y sp is sparse view projection data, E is an allowable error threshold value, and x TV is a total variation regularization term; as an estimated initial value, equation (6) is solved by ASD-POCS iteration.
  7. 7. The CT imaging reconstruction method as recited in claim 6, wherein: In the step S140 of the process of the present invention, Said predicting said outcome And physical coherence results Weighted summation is carried out to obtain a fusion result Specifically, the fusion result is obtained by weighting and summing according to the formula (7) Wherein, the The result of the prediction generated is represented by, And (3) representing a physical consistency result, wherein alpha E [0,1] is a weight coefficient.
  8. 8. CT imaging reconstruction device based on residual poisson flow generation model, characterized by comprising the following modules: the preliminary image reconstruction module 210: Obtaining a preliminary reconstructed image x sp by using a traditional computer tomography reconstruction method; The conditional poisson flow generation modeling module 220: taking the obtained preliminary reconstructed image x sp as a conditional input, representing as conditional posterior distribution sampling, enabling the generated model to be corrected and refined on the basis of the existing poisson flow generated model structure reference, correcting an evolution equation into a conditional ordinary differential equation, and modeling to obtain a conditional poisson flow generated model; model training module 230: Adding noise disturbance to a target full-angle label image x gt to form a sample x t , taking the disturbed image and a preliminary reconstructed image x sp as pairing data, inputting the paired data into the conditional poisson flow generation model for training, outputting a mapping from x sp to x gt by the model, outputting the model as a direction vector pointing to a real image, and optimizing through weighted mean square error; CT image reconstruction module 240: Taking the preliminary sparse angle reconstruction image x sp as a conditional input, calculating a prediction result by a trained conditional poisson flow generation model For the prediction result Physical consistency correction is carried out to obtain a physical consistency result And to compare the prediction result And physical coherence results Weighted summation is carried out to obtain a fusion result With the fusion result And (3) as a new sampling point, reinjecting the normal differential equation updating process of the conditional poisson flow generation model, entering the next iteration, and obtaining a sparse view CT reconstruction result after the iteration is finished.
  9. 9. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that: The processor, when executing the computer program, implements the steps of the CT imaging reconstruction method based on a residual poisson flow generation model of any of claims 1-7.
  10. 10. A computer-readable storage medium having stored thereon a computer program, characterized by: the computer program, when executed by a processor, implements the steps of the CT imaging reconstruction method based on a residual poisson flow generation model of any of claims 1-7.

Description

CT imaging reconstruction method and system based on residual poisson flow generation model Technical Field The invention relates to the technical field of medical image and image reconstruction, in particular to a CT imaging method and a CT imaging system based on residual poisson flow generation model in the sparse view angle in the field of X-ray computed tomography. Background X-ray Computed Tomography (CT) is an important tool for medical image diagnosis. However, conventional CT scans require a large number of projection angle acquisitions, which can lead to higher radiation doses and increased health risks for the patient. To reduce the radiation dose and shorten the scan time, sparse angle sampling (spark-view CT) is often used, i.e., to reduce the number of projection angles. However, due to incomplete projection data, serious streak artifacts and structural distortion often occur in the directly reconstructed image, which affect the accuracy of diagnosis. In recent years, a method based on deep learning and generative modeling is introduced into sparse CT reconstruction, so that the image quality can be improved to a certain extent. The first such iterative reconstruction method is to obtain a reconstruction result by combining statistical modeling of a projection domain with prior constraint of an image domain. Typical methods include regularization methods based on Total Variation (TV), non-local total variation (NLTV), low rank constraints, and the like. The method can improve the image quality under sparse sampling to a certain extent, but has high calculation complexity and slow convergence speed, and more importantly, the prior constraint depends on an artificial structure and has stronger subjectivity. Meanwhile, regularization parameters are difficult to adjust in a self-adaptive mode, the images are excessively smooth due to the fact that the regularization parameters are too strong, and artifacts cannot be restrained due to the fact that the regularization parameters are too weak. The second type is sparse angle CT iterative reconstruction (DiffusionMBIR) based on diffusion model, which uses the full angle sampling CT image as training data in network training stage, and constructs training sample by gradual noise adding process, and trains the ability of denoising network learning to recover clear image under different noise levels. The training process can enable the model to master the statistical distribution characteristics of the real CT image, and clear images are gradually generated through hundreds of steps of iterative reverse sampling from Gaussian noise in the reasoning stage. In each iteration, projection domain data consistency constraint or image domain regularization operation can be added, so that the generated result is ensured to be consistent with sparse projection measurement to a certain extent. But the network training needs to be started from random noise, and hundreds of steps are iterated gradually to generate images. Although the method can obtain a result with higher quality, the reasoning process is tedious, the calculation cost is high, and the requirement of clinical application on instantaneity is difficult to meet. In addition, when physical consistency constraints (such as projection domain data consistency, non-negativity or total variation constraints) are introduced into the diffusion model and the poisson flow generation model, the original generation track continuity is often destroyed, the generation process is unstable, the boundary of a reconstructed image is blurred, the tissue structure is broken, and even forged anatomical details are generated. These false structures may interfere with clinical diagnosis, reducing the reliability of the algorithm. Therefore, how to realize high-quality, physically consistent and efficient CT image reconstruction under the sparse sampling condition is a technical problem that needs to be solved in the prior art. Disclosure of Invention The invention aims to provide a reconstruction method suitable for sparse view CT, which can realize high-fidelity imaging, strong physical consistency constraint and high-efficiency reasoning while remarkably reducing the sampling steps, so as to avoid artifacts and unreal textures under the condition of extremely sparse angles. To achieve the purpose, the invention adopts the following technical scheme: the CT imaging reconstruction method based on the residual poisson flow generation model comprises the following steps: Preliminary image reconstruction step S110: Obtaining a preliminary reconstructed image using conventional computed tomography reconstruction methods ; Conditional poisson flow generation modeling step S120: The preliminary reconstructed image obtained in the step S110 As a condition input, the method is expressed as conditional posterior distribution sampling, so that a generated model can be corrected and refined on the basis of the existing structural refer