A CT image enhancement method and device based on global consistency optimization

Abstract

This invention discloses a CT image enhancement method and apparatus based on global consistency optimization, comprising: performing Gaussian fitting on the pixel intensity distribution of a low-dose CT image to obtain a fitting result containing multiple categories of Gaussian distribution components; calculating the probability of each pixel intensity belonging to each category of the Gaussian distribution component based on the fitting result, and assigning a pixel label to each pixel accordingly; constructing a global optimization objective function containing consistency constraints of adjacent pixel labels, and adjusting the pixel labels through iterative optimization to maximize the objective function, obtaining optimized pixel labels; and performing adaptive deviation reduction processing on the intensity value of each pixel by shrinking it towards the mean of the Gaussian distribution component corresponding to each optimized pixel label, thereby generating an enhanced CT image. This invention can simultaneously achieve noise suppression and detail preservation under low radiation dose conditions, and is applicable to scenarios such as industrial non-destructive testing, medical diagnosis, and high-resolution microscopic imaging.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a CT image enhancement method and apparatus based on global consistency optimization. Background Technology

[0002] In fields such as industrial non-destructive testing and medical imaging, computed tomography (CT) technology is widely used for the internal inspection of complex structures. However, high-quality CT imaging typically requires high radiation doses or long scan times, which increases inspection costs, reduces efficiency, and even poses potential radiation risks to patients in medical applications. To reduce radiation doses, various methods have been proposed, including CT algorithms based on iterative reconstruction and noise reduction methods based on deep learning. The former relies on accurate system modeling, involves high computational costs, and is sensitive to equipment errors; the latter, while highly effective, requires extensive training with labeled data and has limited generalization ability across different data domains.

[0003] In traditional reconstruction optimization approaches, such as the patent application with publication number CN110136217A, a method and system for enhancing liver CT images are provided. This method combines filtered backprojection with adaptive statistical iterative reconstruction, processing and integrating images from different plain and enhanced phases to construct a three-dimensional liver model. However, the filtered backprojection step in this method is sensitive to noise and prone to artifacts, while the adaptive statistical iterative reconstruction process can suppress noise, but it is computationally time-consuming, and its overall efficiency and detail accuracy cannot meet the real-time clinical requirements.

[0004] In the deep learning approach, patent application CN114913107A provides a method for enhancing low-dose CT images in children. It utilizes an improved dual-channel Transformer network, combining image decomposition and segmented reconstruction modules, to enhance low-dose CT images in children to a clinical diagnostic level. This model performs well visually on specific datasets, but its detail restoration capability is highly dependent on well-registered low-dose-high-dose image pairs, limiting its generalization ability.

[0005] Therefore, developing an algorithm that can achieve adaptive noise suppression and detail enhancement directly on a single reconstructed image without modifying CT hardware or requiring registration training data has significant research value and application implications. Summary of the Invention

[0006] In view of the above, the purpose of this invention is to provide a CT image enhancement method and apparatus based on global consistency optimization. By constraining the pixel intensity distribution globally and optimizing the variance, the signal-to-noise ratio and structural clarity of low-dose CT images can be significantly improved without increasing the radiation dose, achieving a balance between noise suppression and detail preservation. It can be widely used in industrial non-destructive testing, medical diagnosis, and high-resolution microscopy.

[0007] To achieve the above-mentioned objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a CT image enhancement method based on global consistency optimization, comprising the following steps: The low-dose CT image to be enhanced is acquired, and Gaussian fitting is performed based on its pixel intensity distribution to obtain the fitting result containing Gaussian distribution components of multiple categories. Based on the fitting results, the probability of each pixel intensity belonging to each Gaussian distribution component is calculated pixel by pixel, and a pixel label is assigned to each pixel based on the category probability; A global optimization objective function containing constraints on the consistency of adjacent pixel labels is constructed. The pixel labels are adjusted through iterative optimization to maximize the objective function, resulting in the optimized pixel labels. Based on the mean of the Gaussian distribution component corresponding to each optimized pixel label, the intensity value of each pixel is subjected to adaptive bias reduction processing to shrink towards the mean, thereby generating an enhanced CT image.

[0008] Preferably, the step of performing Gaussian fitting based on the pixel intensity distribution to obtain a fitting result containing Gaussian distribution components of multiple categories includes: A pixel intensity distribution histogram is constructed based on the intensity values ​​of all pixels in the low-dose CT image to be enhanced. A one-dimensional Gaussian fit is then performed on the pixel intensity distribution histogram to obtain a combination of peaks containing multiple classes of Gaussian distribution components, as well as the mean and standard deviation of each class of Gaussian distribution components.

[0009] Preferably, the step of calculating the probability of each pixel intensity belonging to each Gaussian distribution component based on the fitting results pixel by pixel includes: For any pixel, its intensity value is Then the pixel belongs to the first... Class probabilities of Gaussian-like distribution components Calculate using the following formula: , , in, To determine the total number of categories in the fitted Gaussian distribution components, Intensity value In the The probability density function under the Gaussian-like distribution components, and The first The mean and standard deviation of the Gaussian-like distribution components.

[0010] Preferably, assigning pixel labels to each pixel based on category probability includes: Based on the probability of each pixel belonging to each Gaussian distribution component, a pixel label is assigned to each pixel through probabilistic random sampling, and all pixel labels together constitute the initial label map of the entire image.

[0011] Preferably, the global optimization objective function is expressed as: , in, Represents the set of all pixels. This represents the set of pixel adjacency pairs defined by the selected neighborhood structure. Indicates an indicator function, and Representing pixels pixel labels and pixels pixel tags This represents the total number of partitions in the pixel intensity distribution histogram, which is the number of partitions specified by the user. The CT image pixels were arranged from smallest to largest and equally divided into... Each interval This indicates that it belongs to a pixel intensity distribution histogram partition. The set of pixels, This indicates the partitioning of the pixel intensity distribution histogram. The middle should be assigned to a category The number of pixels, This represents the total number of categories of the fitted Gaussian distribution components.

[0012] Preferably, the iterative optimization is implemented using a greedy algorithm, by arbitrarily selecting a partition of the pixel intensity distribution histogram. For two pixels within the same area, swap their pixel labels. If this increases the global objective function value, keep the swapped label; otherwise, restore the original label. Repeat the swapping process until the global objective function value converges to a local optimum or reaches the preset number of iterations.

[0013] Preferably, the step of performing adaptive bias reduction processing on the intensity value of each pixel by shrinking it towards the mean based on the mean of the Gaussian distribution component corresponding to each optimized pixel label to generate an enhanced CT image includes: For any pixel, its original intensity value is The mean of the Gaussian distribution components corresponding to the optimized pixel labels is The updated strength value Calculate using the following formula: , in, Indicates the first The mean of the Gaussian-like distribution components, Indicates the preset shrinkage factor; By iterating through all pixels in the image and performing the above processing, the intensity values ​​of all pixels have been updated, resulting in the enhanced CT image.

[0014] Secondly, embodiments of the present invention also provide a CT image enhancement device based on global consistency optimization, which is implemented using the above-mentioned CT image enhancement method based on global consistency optimization, including: an intensity distribution fitting module, a probability label allocation module, a global consistency optimization module, and an adaptive deviation shrinkage module; The intensity distribution fitting module is used to acquire the low-dose CT image to be enhanced, and perform Gaussian fitting based on its pixel intensity distribution to obtain a fitting result containing a combination of peaks with multiple Gaussian distribution components. The probability label allocation module is used to calculate the class probability of each pixel intensity belonging to each Gaussian distribution component based on the fitting results, and to assign a pixel label to each pixel based on the class probability. The global consistency optimization module is used to construct a global optimization objective function that includes consistency constraints on adjacent pixel labels. The pixel labels are adjusted through iterative optimization to maximize the objective function, resulting in optimized pixel labels. The adaptive bias shrinkage module is used to perform adaptive bias shrinkage processing on the intensity value of each pixel by shrinking it towards the mean of the Gaussian distribution component corresponding to each optimized pixel label, thereby generating an enhanced CT image.

[0015] Thirdly, embodiments of the present invention also provide an electronic device, including a memory and one or more processors, wherein the memory is used to store a computer program, and the processor is used to implement the above-described CT image enhancement method based on global consistency optimization when executing the computer program.

[0016] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a computer, implements the above-described CT image enhancement method based on global consistency optimization.

[0017] Compared with the prior art, the beneficial effects of the present invention include at least the following: (1) This invention effectively suppresses random noise in low-dose CT images while preserving detailed structures through Gaussian fitting and adaptive deviation reduction processing with mean shrinkage, avoiding over-smoothing and significantly improving the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of the images. It achieves better results than traditional algorithms on multiple public datasets. Furthermore, the intensity deviation reduction of each label pixel in this invention is performed independently, and the intensity calculation of adjacent pixels does not affect each other, thus maintaining edge sharpness better than existing traditional methods.

[0018] (2) This invention optimizes global consistency by using a global maximization of neighborhood label consistency constraints and a greedy swapping iteration. In the pixel label optimization process, it balances the intensity distribution fitting degree and spatial continuity. While maintaining edge sharpness and fine texture structure, it can eliminate artifacts caused by label inconsistency, achieving a synergistic optimization of noise suppression and detail preservation. Because this method is based on global analysis rather than local similarity analysis, it does not produce artifacts due to the interconnected coupling between different locations in the image, unlike traditional methods.

[0019] (3) This invention is model-driven rather than data-driven. The processing does not rely on any training data and can be enhanced directly with a single low-dose CT reconstructed image as input. It is computationally stable, highly interpretable, and compatible with various existing CT imaging systems. It can be directly used in different types of CT image data without modifying hardware or scanning protocols. It can be used as a general post-processing module and directly integrated into clinical or industrial testing processes. Compared with traditional deep learning methods, it significantly reduces the need for expensive training data. Experimental verification shows that in actual lithium battery sample testing, the method of this invention achieves clarity comparable to conventional imaging with one-fifth of the radiation dose, significantly reducing radiation risk and scanning costs. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic flowchart of the CT image enhancement method based on global consistency optimization provided in an embodiment of the present invention; Figure 2 These are examples of image processing procedures and schematic diagrams of the final results provided in embodiments of the present invention; Figure 3 These are comparison images of image enhancement results provided in embodiments of the present invention; Figure 4This is a schematic diagram of the structure of a CT image enhancement device based on global consistency optimization provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0023] The inventive concept of this invention is as follows: Addressing the problems of existing technologies where noise suppression and detail preservation in low-dose CT images are difficult to balance, and which rely heavily on training data and have poor generalization ability, this invention provides a CT image enhancement method and apparatus based on global consistency optimization. It involves Gaussian fitting of pixel intensity distribution, assigning pixel labels according to category probability, constructing a global optimization objective function containing neighborhood label consistency constraints, iteratively optimizing to achieve global adjustment of pixel labels, and finally, adaptively shrinking pixel intensity based on the optimized pixel labels and Gaussian fitting parameters. The entire process requires no training data and can be directly embedded into existing CT systems, achieving image enhancement effects that balance image detail and noise consistency even at low radiation doses.

[0024] like Figure 1 As shown in the embodiment, a CT image enhancement method based on global consistency optimization is provided, including the following steps: S1. Acquire the low-dose CT image to be enhanced, and perform Gaussian fitting based on its pixel intensity distribution to obtain the fitting result containing Gaussian distribution components of multiple categories.

[0025] In this embodiment, the low-dose CT image to be enhanced is acquired, specifically the input image is an image matrix IMG (e.g., ...). Figure 1 The noise map shown by I in the image is then used to obtain the intensity distribution envelope of the image based on the intensity distribution of each pixel (e.g., the noise map shown by I in the image). Figure 1 The pixel intensity distribution histogram shown in II is then fitted to a graph containing... The combination of peaks of the Gaussian distribution components of each category (such as...) Figure 1 (As shown in III).

[0026] Specifically, a one-dimensional Gaussian fitting is used, obtained by directly calling functions in software such as MATLAB. For example, the MATLAB function `fitgmdist` is expressed as `gm = fitgmdist(IMG(:), C)`, and the output is a `gm` cell, where the `gm.mean` value contains the mean value. The mean of a Gaussian distribution, gm.Sigma in the cell contains The standard deviation of a Gaussian distribution.

[0027] S2, based on the fitting results, calculate the class probability of each pixel intensity belonging to each Gaussian distribution component, and assign a pixel label to each pixel based on the class probability.

[0028] In the embodiments, based on such Figure 1 The fitting results shown in III in the figure demonstrate that hard classification is performed pixel by pixel according to image intensity, which can ensure that the overall intensity distribution remains basically unchanged.

[0029] Specifically, for any pixel, its intensity value is Then the pixel is randomly classified into the first... Class probabilities of Gaussian-like distribution components Calculate using the following formula: , , in, To determine the total number of categories in the fitted Gaussian distribution components, Intensity value In the The probability density function under the Gaussian-like distribution components, and The first The mean and standard deviation of the Gaussian-like distribution components.

[0030] like Figure 1 As shown in IV, based on the probability of each pixel belonging to each Gaussian distribution component, the probability (i.e., obtained by the above formula) is used to calculate the probability. The random sampling method assigns a pixel label to each pixel, and all pixel labels are used to form an initial label map for the entire image (e.g., ...). Figure 1 (as shown by V in the diagram).

[0031] S3. Construct a global optimization objective function that includes constraints on the consistency of adjacent pixel labels. Adjust the pixel labels through iterative optimization to maximize the objective function and obtain the optimized pixel labels.

[0032] In this embodiment, a global optimization objective function is constructed to maximize the consistency of adjacent pixel labels, expressed as: , in, Represents the set of all pixels. This represents the set of pixel adjacency pairs defined by the selected neighborhood structure. Indicates an indicator function, and Representing pixels pixel labels and pixels pixel tags, This represents the total number of partitions in the pixel intensity distribution histogram, which is the number of partitions specified by the user. The CT image pixels were arranged from smallest to largest and equally divided into... Each interval This indicates that it belongs to a pixel intensity distribution histogram partition. The set of pixels, This indicates the partitioning of the pixel intensity distribution histogram. The middle should be assigned to a category The number of pixels, This represents the total number of categories of the fitted Gaussian distribution components. The neighborhood structure specifically uses 26-adjacency (face, edge, vertex adjacency), which means that all voxels except the center in a 3×3×3 cube are neighborhoods.

[0033] Specifically, the iterative optimization is implemented using a greedy algorithm, which arbitrarily selects partitions of the pixel intensity distribution histogram. Swap the pixel labels of two pixels within the same area. and If the global objective function value increases, the swap is retained; otherwise, the original label is restored, and the above swap process is repeated until the global objective function value converges to a local optimum or reaches the preset number of iterations.

[0034] After iterative optimization, a new label map was obtained, such as Figure 1 As shown in VI.

[0035] S4. Based on the mean of the Gaussian distribution component corresponding to each optimized pixel label, the intensity value of each pixel is subjected to adaptive bias reduction processing to shrink towards the mean, thereby generating an enhanced CT image.

[0036] In this embodiment, the pixel intensity is redistributed based on the fitting result of step S1 and the pixel label obtained by optimization in step S3.

[0037] Specifically, for any pixel, its original intensity value is After optimization, it belongs to category pixel tags The mean of the corresponding Gaussian distribution components is The updated strength value Calculate using the following formula: , in, Indicates the first The mean of the Gaussian-like distribution components; This represents the preset shrinkage factor, with a preferred value range of 3 to 8, which can be specified by the user.

[0038] By iterating through all pixels in the image and performing the above processing, the intensity values ​​of all pixels have been updated (the intensity distribution histogram results are shown below). Figure 1 As shown in VII), the final enhanced CT image is obtained (as shown in VII). Figure 1 (as shown in VIII). Figure 1 IX in the diagram represents the ground truth.

[0039] To further verify the experimental effects of the present invention, such as Figure 3 As shown, the method of this invention is compared with existing traditional methods. These existing methods include BM3D (reference title: Image denoising by sparse 3-D transform-domain collaborative filtering) and Collaborative filtering of correlated noise: Exact transform-domain variance for improved shrinkage and patch matching), NLM (reference title: Non-local means denoising), and BF (reference title: Bilateral filtering for gray and color images). It can be seen that the enhancement results of this method are significantly better than those of existing methods.

[0040] Based on the same inventive concept, such as Figure 4 As shown, this embodiment of the invention also provides a CT image enhancement device 400 based on global consistency optimization, including: an intensity distribution fitting module 410, a probability label allocation module 420, a global consistency optimization module 430, and an adaptive deviation shrinkage module 440.

[0041] The intensity distribution fitting module 410 is used to acquire the low-dose CT image to be enhanced, and perform Gaussian fitting based on its pixel intensity distribution to obtain the fitting result containing a combination of peaks with multiple Gaussian distribution components.

[0042] The probability label assignment module 420 is used to calculate the class probability of each pixel intensity belonging to each Gaussian distribution component based on the fitting results, and assign a pixel label to each pixel based on the class probability.

[0043] The global consistency optimization module 430 is used to construct a global optimization objective function that includes consistency constraints on adjacent pixel labels. The pixel labels are adjusted through iterative optimization to maximize the objective function and obtain the optimized pixel labels.

[0044] The adaptive bias shrinkage module 440 is used to perform adaptive bias shrinkage processing on the intensity value of each pixel by shrinking it towards the mean of the Gaussian distribution component corresponding to each optimized pixel label, thereby generating an enhanced CT image.

[0045] Based on the same inventive concept, embodiments of the present invention also provide an electronic device, including a memory and one or more processors, wherein the memory is used to store a computer program, and the processor is used to implement the above-described CT image enhancement method based on global consistency optimization when executing the computer program.

[0046] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a computer, implements the above-described CT image enhancement method based on global consistency optimization.

[0047] It should be noted that the CT image enhancement device, electronic device, and computer-readable storage medium based on global consistency optimization provided in the above embodiments all belong to the same inventive concept as the CT image enhancement method based on global consistency optimization. For details of their specific implementation process, please refer to the embodiments of the CT image enhancement method based on global consistency optimization, which will not be repeated here.

[0048] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A CT image enhancement method based on global consistency optimization, characterized in that, Includes the following steps: The low-dose CT image to be enhanced is acquired, and Gaussian fitting is performed based on its pixel intensity distribution to obtain the fitting result containing Gaussian distribution components of multiple categories. Based on the fitting results, the probability of each pixel intensity belonging to each Gaussian distribution component is calculated pixel by pixel, and a pixel label is assigned to each pixel based on the category probability; A global optimization objective function containing constraints on the consistency of adjacent pixel labels is constructed. The pixel labels are adjusted through iterative optimization to maximize the objective function, resulting in the optimized pixel labels. Based on the mean of the Gaussian distribution component corresponding to each optimized pixel label, the intensity value of each pixel is subjected to adaptive bias reduction processing to shrink towards the mean, thereby generating an enhanced CT image.

2. The CT image enhancement method based on global consistency optimization according to claim 1, characterized in that, The Gaussian fitting based on the pixel intensity distribution yields a fitting result containing Gaussian distribution components of multiple categories, including: A pixel intensity distribution histogram is constructed based on the intensity values ​​of all pixels in the low-dose CT image to be enhanced. A one-dimensional Gaussian fit is then performed on the pixel intensity distribution histogram to obtain a combination of peaks containing multiple classes of Gaussian distribution components, as well as the mean and standard deviation of each class of Gaussian distribution components.

3. The CT image enhancement method based on global consistency optimization according to claim 1 or 2, characterized in that, The step of calculating the probability of each pixel intensity belonging to each Gaussian distribution component based on the fitting results includes: For any pixel, its intensity value is Then the pixel belongs to the first... Class probabilities of Gaussian-like distribution components Calculate using the following formula: ， ， in, To determine the total number of categories in the fitted Gaussian distribution components, Intensity value In the The probability density function under the Gaussian-like distribution components, and The first Mean and standard deviation of Gaussian-like distribution components.

4. The CT image enhancement method based on global consistency optimization according to claim 1, characterized in that, The process of assigning pixel labels to each pixel based on class probability includes: Based on the probability of each pixel belonging to each Gaussian distribution component, a pixel label is assigned to each pixel through probabilistic random sampling, and all pixel labels together constitute the initial label map of the entire image.

5. The CT image enhancement method based on global consistency optimization according to claim 1, characterized in that, The global optimization objective function is expressed as: ， in, Represents the set of all pixels. This represents the set of pixel adjacency pairs defined by the selected neighborhood structure. Indicates an indicator function, and Representing pixels pixel labels and pixels pixel tags, This represents the total number of partitions in the pixel intensity distribution histogram, which is the number of partitions specified by the user. The CT image pixels were arranged from smallest to largest and equally divided into... Each interval This indicates that it belongs to a pixel intensity distribution histogram partition. The set of pixels, This indicates the partitioning of the pixel intensity distribution histogram. The middle should be assigned to a category The number of pixels, This represents the total number of categories of the fitted Gaussian distribution components.

6. The CT image enhancement method based on global consistency optimization according to claim 5, characterized in that, The iterative optimization is implemented using a greedy algorithm, which arbitrarily selects partitions of the pixel intensity distribution histogram. For two pixels within the same area, swap their pixel labels. If this increases the global objective function value, keep the swapped label; otherwise, restore the original label. Repeat the swapping process until the global objective function value converges to a local optimum or reaches the preset number of iterations.

7. The CT image enhancement method based on global consistency optimization according to claim 1, characterized in that, The step of adaptively reducing the intensity value of each pixel by shrinking it towards the mean of the Gaussian distribution component corresponding to each optimized pixel label to generate an enhanced CT image includes: For any pixel, its original intensity value is The mean of the Gaussian distribution components corresponding to the optimized pixel labels is The updated strength value Calculate using the following formula: ， in, Indicates the first The mean of the Gaussian-like distribution components, Indicates the preset shrinkage factor; By iterating through all pixels in the image and performing the above processing, the intensity values ​​of all pixels have been updated, resulting in the enhanced CT image.

8. A CT image enhancement device based on global consistency optimization, implemented using the CT image enhancement method based on global consistency optimization as described in any one of claims 1 to 7, characterized in that, include: The module includes an intensity distribution fitting module, a probability label allocation module, a global consistency optimization module, and an adaptive bias shrinkage module. The intensity distribution fitting module is used to acquire the low-dose CT image to be enhanced, and perform Gaussian fitting based on its pixel intensity distribution to obtain a fitting result containing a combination of peaks with multiple Gaussian distribution components. The probability label allocation module is used to calculate the class probability of each pixel intensity belonging to each Gaussian distribution component based on the fitting results, and to assign a pixel label to each pixel based on the class probability. The global consistency optimization module is used to construct a global optimization objective function that includes consistency constraints on adjacent pixel labels. The pixel labels are adjusted through iterative optimization to maximize the objective function, resulting in optimized pixel labels. The adaptive bias shrinkage module is used to perform adaptive bias shrinkage processing on the intensity value of each pixel by shrinking it towards the mean of the Gaussian distribution component corresponding to each optimized pixel label, thereby generating an enhanced CT image.

9. An electronic device comprising a memory and one or more processors, the memory for storing a computer program, characterized in that, The processor is used to implement the CT image enhancement method based on global consistency optimization as described in any one of claims 1 to 7 when executing a computer program.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by a computer, it implements the CT image enhancement method based on global consistency optimization as described in any one of claims 1 to 7.

Images