A pre-logarithmic domain voronoi decomposition assisted low-dose ct reconstruction method

Abstract

This invention provides a low-dose CT reconstruction method assisted by pre-log domain Voronoi decomposition. The method includes obtaining a pre-log sinusoidal graph of low-dose CT, performing Voronoi decomposition on the pre-log sinusoidal graph using the K-means clustering algorithm to obtain several feature clusters, mapping each feature cluster to its corresponding latent space to obtain a multi-channel input feature map, inputting the multi-channel input feature map into a pre-trained diffusion transformer model for iterative denoising and optimization to obtain optimized cluster features, fusing and performing logarithmic transformation on the optimized cluster features to obtain a post-log sinusoidal graph, and reconstructing the post-log sinusoidal graph using a filtered back-projection algorithm to obtain a high-quality CT image. This invention solves the problems of large dynamic range, uneven gradient, and progressively amplified noise in the pre-log domain by decoupling features from the pre-log sinusoidal graph, modeling the latent space, and optimizing the diffusion transformer, thus achieving high-fidelity low-dose CT image reconstruction.

Description

Technical Field

[0001] This invention belongs to the technical field of medical image processing, specifically relating to a low-dose CT reconstruction method and system assisted by pre-logarithmic domain Voronoi decomposition. Background Technology

[0002] Computed tomography (CT) technology uses X-rays to penetrate the human body from multiple angles and reconstructs tomographic images of the body's internal structures using projection data, making it an indispensable tool in modern clinical diagnosis. However, the X-ray radiation used during CT examinations may pose potential health risks to patients, thus reducing radiation dose has become an important direction for the development of CT technology. Low-dose CT (LDCT) technology has emerged to address this need, but its core challenge lies in the fact that reducing the radiation dose significantly increases quantum noise in the projection data, leading to severe fringe artifacts and structural blurring in the reconstructed images, thereby affecting the accuracy and reliability of diagnosis.

[0003] Existing low-dose CT reconstruction methods can be divided into three main categories according to the different data domains processed:

[0004] The first category is image domain methods. These methods directly post-process the reconstructed CT images, using convolutional neural networks, generative adversarial networks, or unfolded optimization frameworks to remove noise and artifacts from the images. The advantage of image domain methods is that they process clinically common image formats and are easy to deploy. However, their fundamental limitation is that noise and artifacts in the image domain are deeply integrated with the anatomical structures during the reconstruction process, making it impossible to fundamentally correct the physical noise introduced by the projection data, especially global stripe artifacts.

[0005] The second category is the post-logarithmic projection domain method. This type of method performs denoising after the projected data undergoes a logarithmic transformation; that is, it suppresses noise in the sinusoidal domain before image reconstruction. Compared to image domain methods, post-logarithmic domain processing can correct the projected data before reconstruction, resulting in better physical consistency. However, the logarithmic transformation alters the statistical properties of the noise, converting Poisson noise into signal-dependent approximate Gaussian noise. Furthermore, residual noise in the sinusoidal graph is amplified and propagates globally during the filtered backprojection process, thus requiring extremely high denoising accuracy.

[0006] The third category is the pre-log projection domain method. This type of method directly processes the original photon count data before logarithmic transformation, preserving the original Poisson statistical characteristics of the noise and possessing optimal physical modeling potential. However, the pre-log domain faces three major challenges: First, the dynamic range of the pre-log sine curve is extremely large, with high photon counts corresponding to background regions and extremely low photon counts corresponding to tissue decay regions, causing the network gradient to be dominated by the background and making it difficult to learn fine structures; second, the gradient scale in the pre-log domain changes exponentially with the decay coefficient, exhibiting severe non-stationarity; third, small prediction errors in the pre-log domain are exponentially amplified after logarithmic transformation and propagate globally during reconstruction, resulting in an overall bias in the final image.

[0007] Therefore, there is an urgent need for a low-dose CT reconstruction method that can effectively decouple different intensity components in the prelog domain, stabilize gradient statistical properties, and suppress noise amplification step by step. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a low-dose CT reconstruction method and system assisted by prelog domain Voronoi decomposition. By performing feature decoupling, latent space modeling, and diffusion transformer optimization on the prelog sine curve, the problems of large dynamic range, uneven gradient, and progressively amplified noise in the prelog domain are solved, thereby achieving high-fidelity low-dose CT image reconstruction.

[0009] In a first aspect, the present invention provides the following technical solution: a low-dose CT reconstruction method assisted by pre-logarithmic domain Voronoi decomposition, comprising:

[0010] The pre-log sinusoidal plot of low-dose CT was obtained, and the Voronoi decomposition of the pre-log sinusoidal plot was performed using the K-means clustering algorithm to obtain several feature clusters.

[0011] Each feature cluster is mapped to its corresponding latent space to obtain a multi-channel input feature map;

[0012] The multi-channel input feature map is input into the pre-trained diffusion transformer model for iterative denoising and optimization to obtain optimized clustering features;

[0013] The optimized clustering features are fused and subjected to logarithmic transformation to obtain a post-logarithmic sine curve.

[0014] A filtered back-projection algorithm is used to reconstruct the post-logarithmic sine curve to obtain a high-quality CT image.

[0015] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention introduces a Voronoi decomposition mechanism into the pre-log projection domain for the first time. Through K-means clustering, the pre-log sine wave is divided into different clusters according to intensity. Within each cluster, the numerical scale tends to be balanced and the gradient remains stable, effectively solving the problems of large dynamic range and non-stationary gradients in the pre-log domain. This allows deep learning models to focus on the features themselves rather than being dominated by the background. Furthermore, the invention employs a latent space mapping strategy, mapping different clusters to independent latent spaces, achieving effective decoupling of features and avoiding information aliasing between regions of different intensities, laying the foundation for subsequent accurate denoising. This invention incorporates latent space representation into a transformer-enhanced diffusion model, utilizes multi-head transpose attention to capture global dependencies, and employs a gated feedforward network to enhance feature discriminability, ensuring structural fidelity while maintaining global consistency. Through mathematical proof and experimental verification, this invention determines the optimal number of clusters for Voronoi decomposition to be 3, a choice that achieves an optimal balance between reconstruction noise and model complexity, demonstrating theoretical completeness and sufficient experimental results. This invention effectively suppresses the stepwise amplification effect of pre-logarithmic domain noise during logarithmic transformation and filtered backprojection, significantly improving the peak signal-to-noise ratio and structural similarity index of low-dose CT reconstructed images.

[0016] Preferably, the step of obtaining the pre-log sinusoidal plot of low-dose CT and performing Voronoi decomposition on the pre-log sinusoidal plot using the K-means clustering algorithm to obtain several feature clusters includes:

[0017] Obtain raw photon count data from low-dose CT scans, and determine the photon intensity data based on the raw photon count data:

[0018] ;

[0019] In the formula, For projection angle and detector position The photon intensity at that location was measured. The intensity of the incident photon. Integrating for noise-free tissue attenuation lines, Poisson measurement noise inherent in photon counting;

[0020] The photon intensity data are arranged according to the projection angle and detector channel to obtain a pre-logarithmic sine curve for low-dose CT.

[0021] The K-means clustering algorithm is used to perform Voronoi decomposition on the prelog sine graph to obtain several feature clusters, including high-intensity features, medium-intensity features, and low-intensity features.

[0022] Preferably, the step of mapping each feature cluster to its corresponding latent space to obtain a multi-channel input feature map includes:

[0023] Each feature cluster is input into the hidden encoder. Mapping to the corresponding latent space to obtain a low-dimensional vector representation. :

[0024] ;

[0025] In the formula, For the first Clustering based on features;

[0026] The low-dimensional vector representations are concatenated along the channel dimensions to obtain a multi-channel input feature map. :

[0027] ;

[0028] In the formula, For splicing operations, These are the low-dimensional vector representations corresponding to the first, second, and third feature clusters, respectively.

[0029] Preferably, the step of inputting the multi-channel input feature map into the pre-trained diffusion transformer model for iterative denoising and optimization to obtain optimized clustering features includes:

[0030] The multi-channel input feature map is input into the pre-trained diffusion transformer model, and then processed by the noise estimation network in the pre-trained diffusion transformer model. Estimating the noise of the multi-channel input feature map :

[0031] ;

[0032] In the formula, For the first Each feature cluster at time step The noisy characteristics, The first of the multi-channel input feature maps One channel;

[0033] According to time steps Perform reverse denoising updates sequentially:

[0034] , ;

[0035] In the formula, For the first Each feature cluster at time step The noisy characteristics, These are the signal preservation coefficient and the noise variance, respectively.

[0036] Repeat the inverse denoising process until a clean latent vector is output. ;

[0037] The clean hidden vector In the input multi-head transpose attention module, for the clean latent vector Query is generated using 1×1 pointwise convolution and 3×3 depthwise convolution. ,key ,value ;

[0038] Query ,key ,value Divided into There are 1 attention head, and a scaled dot product attention is computed for each attention head. :

[0039] ;

[0040] In the formula, The first The query features, key features, and value features corresponding to each attention head. The dimension of the attention head;

[0041] The scaled dot product attentions of all attention heads are concatenated to obtain the MTA output. :

[0042] ;

[0043] In the formula, For splicing operations;

[0044] Output the MTA The input is gated in a gated feedforward network to obtain optimized clustering features:

[0045] ;

[0046] In the formula, These are 1×1 pointwise convolution kernels for the first and second groups, respectively. These are 3×3 depth convolution kernels for the first and second groups, respectively. For element-wise multiplication, is the activation function for the Gaussian error linear unit.

[0047] Preferably, the step of fusing the optimized clustering features and performing a logarithmic transformation to obtain the post-logarithmic sine curve includes:

[0048] Optimize clustering features The fusion is performed to obtain a prelogarithmic sine curve. :

[0049] ;

[0050] Perform a logarithmic transformation on the pre-logarithmic sine graph to obtain the post-logarithmic sine graph. :

[0051] ;

[0052] In the formula, The intensity of the incident photon.

[0053] Preferably, in the step of reconstructing the post-logarithmic sine wave using a filtered back-projection algorithm to obtain a high-quality CT image, the high-quality CT image... The output process is as follows:

[0054] ;

[0055] In the formula, This indicates the filter used in the FBP. For the set of projection angles, This is for filtered back projection. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention, 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.

[0057] Figure 1 This is a flowchart of low-dose CT reconstruction assisted by pre-logarithmic domain Voronoi decomposition in an embodiment of the present invention;

[0058] Figure 2 This is a silhouette coefficient analysis diagram for different numbers of clusters in an embodiment of the present invention;

[0059] Figure 3 The images shown are reconstruction results of the CHAOS dataset at a noise level of 1e4, using different methods applied in this embodiment of the invention. (a) Reference image, (b) LDCT, comparison image, (c) Pix2Pix, (d) NCSN++, (e) CaGAN, (f) Re-UNet, (g) WiTUnet, (h) SDCNN, (i) PLOT-CT. The display window width is [-185, 10] HU. The second row shows the reconstructed residual image, and the third row gives a magnified view of the ROI region marked by the red box in the first row.

[0060] Figure 4The images show the reconstruction results of the CHAOS dataset at a noise level of 5e3, using different methods applied in this embodiment of the invention. (a) Reference image, (b) LDCT, comparison image, (c) Pix2Pix, (d) NCSN++, (e) CaGAN, (f) Re-UNet, (g) WiTUnet, (h) SDCNN, (i) PLOT-CT. The display window width is [-185, 10] HU. The second row shows the reconstructed residual image, and the third row gives a magnified view of the ROI marked by the red box in the first row.

[0061] Figure 5 The images shown are reconstruction results of the CHAOS dataset at a noise level of 1e3, using different methods applied in this embodiment of the invention. (a) Reference image, (b) LDCT, comparison image, (c) Pix2Pix, (d) NCSN++, (e) CaGAN, (f) Re-UNet, (g) WiTUnet, (h) SDCNN, (i) PLOT-CT. The display window width is [-185, 10] HU. The second row shows the reconstructed residual image, and the third row gives a magnified view of the ROI region marked by the red box in the first row.

[0062] Figure 6 The images show the reconstruction results of different methods applied to the AAPM challenge dataset at a noise level of 1e3 in this embodiment of the invention. (a) Reference image, (b) Low-dose CT (LDCT, contrast image), (c) Pix2Pix, (d) NCSN++, (e) CaGAN, (f) Re-UNet, (g) WiTUnet, (h) SDCNN, (i) PLOT-CT. The display window width is [-85, 14] HU. The second row shows the reconstructed residual map, and the third row gives a magnified view of the ROI.

[0063] The embodiments of the present invention will be further described below with reference to the accompanying drawings. Detailed Implementation

[0064] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain embodiments of the present invention, and should not be construed as limiting the present invention.

[0065] Example 1

[0066] In Embodiment 1 of the present invention, as Figure 1 As shown, a low-dose CT reconstruction method assisted by pre-logarithmic domain Voronoi decomposition includes:

[0067] S1. Obtain the pre-log sinusoidal plot of low-dose CT, and use the K-means clustering algorithm to perform Voronoi decomposition on the pre-log sinusoidal plot to obtain several feature clusters;

[0068] Specifically, step S1 includes:

[0069] S11. Obtain raw photon count data from low-dose CT scans, and determine the measured photon intensity data based on the raw photon count data:

[0070] ;

[0071] In the formula, For projection angle and detector position The photon intensity at that location was measured. The intensity of the incident photon. Integrating for noise-free tissue attenuation lines, This is the Poisson measurement noise inherent in photon counting.

[0072] S12. Arrange the photon intensity data according to the projection angle and detector channel to obtain a pre-logarithmic sine curve for low-dose CT.

[0073] S13. The prelog sine graph is decomposed into Voronoi decomposition using the K-means clustering algorithm to obtain several feature clusters, wherein the feature clusters include high-intensity features, medium-intensity features and low-intensity features.

[0074] Specifically, high-intensity features correspond to high-intensity regions, mainly including background regions and regions with high photon counts; medium-intensity features correspond to intermediate-intensity regions, mainly including transition regions; and low-intensity features correspond to low-intensity regions, mainly including tissue attenuation regions and regions with low photon counts.

[0075] In this application, the number of clustering features is 3, and the corresponding proof is given below:

[0076] Assume the data space of the prelog sine graph is a compact set. ,set up for The above describes the probability measure of data distribution, which has generator The decomposition of Volonoi will Divided into One Volonoi unit Each unit is defined as:

[0077] ;

[0078] For a given Volonoi decomposition, the reconstructed noise is quantized into random data points. The expected squared distance to its nearest generator:

[0079] ;

[0080] For a fixed number of clusters The optimal generator configuration minimizes this noise:

[0081] ;

[0082] To evaluate clustering quality, a silhouette coefficient is introduced. This coefficient measures both intra-cluster cohesion and inter-cluster separation. For any data point , Intraclass mean Euclidean distance ,as well as The minimum average distance to all non-self Volonoi units is:

[0083] ;

[0084] ;

[0085] In the formula, For Volonoi Unit The total number of data points included. For the first Data points, For the first One Volonoi unit;

[0086] The silhouette coefficient quantifies the similarity of a point to its neighboring clusters compared to its own cluster, ranging from -1 to 1. A score close to 1 indicates high-quality clustering, while a score close to -1 suggests the point may have been assigned to an incorrect cluster. The contour coefficient is defined as:

[0087] ;

[0088] The overall silhouette coefficient of the cluster is the average of all data points:

[0089] , ;

[0090] along with The increase in the profile coefficient Usually in a specific The value reaches its maximum at a certain point, and this peak reflects the optimal balance between intra-cluster cohesion and inter-cluster separation.

[0091] To balance these trade-offs, a regularization objective function is introduced. This combines reconstruction noise with complexity penalties and clustering quality:

[0092] ;

[0093] In the formula, and This is a regularization parameter used to control the trade-off between reconstruction accuracy, model complexity, and clustering quality;

[0094] According to quantization theory, for a smooth probability density on a tight support... Data distribution, optimal quantization noise Satisfy the following asymptotic relations:

[0095] ;

[0096] in Represents the spatial dimension of the data. This is a constant that depends on the dimension and data distribution. This gives a sufficiently large... Approximate time:

[0097] ;

[0098] Substituting this approximation into the regularization objective function yields:

[0099] ;

[0100] For typical data distributions For a convex function, at a certain... The maximum value is found at the value, which is near the optimal value. Confirmed for The local minimum value.

[0101] The proof theoretically demonstrates that, for any given dataset and regularization strength, there exists an optimal number of Volonoi clusters. This minimizes the combination of reconstruction noise and model complexity. This provides the mathematical basis for our choice of tri-cluster decomposition. Figure 2 The experimental results show that the tri-clustering partitioning is superior to other partitioning schemes in all evaluation indicators, verifying the consistency between theoretical analysis and experimental results.

[0102] S2. Map each feature cluster to its corresponding latent space to obtain a multi-channel input feature map;

[0103] Specifically, step S2 includes:

[0104] S21. Input each feature cluster into the hidden encoder separately. Mapping to the corresponding latent space to obtain a low-dimensional vector representation. :

[0105] ;

[0106] In the formula, For the first Clustering based on features;

[0107] S22. Concatenate the low-dimensional vector representation along the channel dimension to obtain a multi-channel input feature map. :

[0108] ;

[0109] In the formula, For splicing operations, These are the low-dimensional vector representations corresponding to the first, second, and third feature clusters, respectively.

[0110] S3. Input the multi-channel input feature map into the pre-trained diffusion transformer model for iterative denoising and optimization to obtain optimized clustering features;

[0111] Specifically, step S3 includes:

[0112] S31. Input the multi-channel input feature map into the pre-trained diffusion transformer model, and then use the noise estimation network in the pre-trained diffusion transformer model. Estimating the noise of the multi-channel input feature map :

[0113] ;

[0114] In the formula, For the first Each feature cluster at time step The noisy characteristics, The first of the multi-channel input feature maps One channel;

[0115] Specifically, the actual diffusion transformer model also includes a forward diffusion process, through which noise is gradually added to the initial clean latent variables. The reverse process trains the diffusion transformer to recover clean clusters by estimating the noise. The noisy features here are pure random noise. The model only depends on low-dose input and pre-trained network parameters, and does not require any normal dose data as supervision or reference.

[0116] S32, according to time steps Perform reverse denoising updates sequentially:

[0117] , ;

[0118] In the formula, For the first Each feature cluster at time step The noisy characteristics, These are the signal preservation coefficient and the noise variance, respectively.

[0119] S33. Repeat the reverse denoising process until a clean latent vector is output. .

[0120] S34, the clean hidden vector In the input multi-head transpose attention module, for the clean latent vector Query is generated using 1×1 pointwise convolution and 3×3 depthwise convolution. ,key ,value .

[0121] S35, query ,key ,value Divided into There are 1 attention head, and a scaled dot product attention is computed for each attention head. :

[0122] ;

[0123] In the formula, The first The query features, key features, and value features corresponding to each attention head. For the dimension of attention head.

[0124] S36. Concatenate the scaled dot product attention values ​​of all attention heads to obtain the MTA output. :

[0125] ;

[0126] In the formula, This is for splicing operations.

[0127] S37, Output the MTA The input is gated in a gated feedforward network to obtain optimized clustering features:

[0128] ;

[0129] In the formula, These are 1×1 pointwise convolution kernels for the first and second groups, respectively. These are 3×3 depth convolution kernels for the first and second groups, respectively. For element-wise multiplication, The activation function for the Gaussian error linear unit;

[0130] Both scaled dot product attention and gated feedforward network are integrated into the Transformer module. The Transformer module is trained in a supervised manner to learn cross-cluster feature mapping relationships. The core components of the Transformer module include MTA (scaled dot product attention) for capturing long-distance dependencies between different regions in the Voronoi space, and GFN (gated feedforward network) for improving feature discriminativeness.

[0131] Specifically, during the pre-training process of the diffusion converter model, normal dose CT images need to be acquired simultaneously. After processing them through step S2, the same feature clusters are obtained. The normal dose and low dose feature clusters are then concatenated in step S2 and input into the model for pre-training. The model uses the low dose feature cluster as input and the normal dose feature cluster as the true label to learn the cross-cluster feature mapping from low dose to normal dose.

[0132] S4. The optimized clustering features are fused and logarithmically transformed to obtain the post-logarithmic sine curve;

[0133] Specifically, step S4 includes:

[0134] S41, Optimize clustering features The fusion is performed to obtain a prelogarithmic sine curve. :

[0135] ;

[0136] Specifically, the three optimized clustering features do not overlap and correspond to different regions. By directly summing and fusing elements one by one, a complete pre-log sine curve is reconstructed.

[0137] S42. Perform a logarithmic transformation on the pre-logarithmic sine graph to obtain the post-logarithmic sine graph. :

[0138] ;

[0139] In the formula, The intensity of the incident photon;

[0140] Specifically, the reconstructed pre-logarithmic sine wave is transformed into the post-logarithmic domain by a logarithmic transformation. This step converts the multiplicative Poisson noise in the pre-logarithmic domain into approximately additive Gaussian noise in the post-logarithmic domain.

[0141] S5. The back-projection algorithm is used to reconstruct the logarithmic sine curve to obtain a high-quality CT image;

[0142] In step S5, high-quality CT images The output process is as follows:

[0143] ;

[0144] In the formula, This indicates the filter used in the FBP. For the set of projection angles, This is for filtered back projection.

[0145] To verify the beneficial effects of the method of the present invention, the following auxiliary experiments are provided:

[0146] The diffusion converter model in this application is denoted as the PLOT-CT model. It is implemented using ODL and PyTorch, and runs on a workstation equipped with two NVIDIA RTX 3090 GPUs to ensure computational power and stability for model training and inference. To simulate X-ray computed tomography, Poisson noise generated by photon counting statistics is introduced into the projection data, and X-ray sources of different intensities are configured to match different clinical low-dose acquisition conditions. A two-stage training strategy balances convergence speed and performance. The first stage uses the Adam optimizer with an initial learning rate of 1.5e. -4 CosineAnnealingRestartLR is used to prevent premature convergence. The second stage employs the Adam optimizer with an initial learning rate of 2e^(-1 / 2). -4 A Multi-Step LR scheduler was used, halving the learning rate at 300k steps to optimize parameters. To evaluate reconstruction quality, three standard LDCT metrics were employed: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Mean Squared Error (MSE). All metrics were averaged over test samples to ensure reliability. Generally, higher PSNR / SSIM and lower MSE indicate better reconstruction quality.

[0147] It is worth noting that PLOT-CT generates images with fewer artifacts and lower noise compared to baseline methods, while preserving anatomical structures that are crucial for clinical analysis.

[0148] PLOT-CT was compared with six baseline LDCT reconstruction methods, including Pix2Pix, NCSN++, CaGAN, Re-UNet, WiTUnet, and SDCNN, to evaluate the performance of each method. 2200 images were selected from the AAPMchallenge dataset, a commonly used benchmark dataset in LDCT research, to ensure the validity and comparability of the experiments. Three noise levels were added, corresponding to 1e4, 5e3, and 1e3 photons per X-ray path, respectively, to simulate different low-dose scenarios. After adding noise, a total of 6600 images were obtained, and PLOT-CT was trained based on these log-before-sine curve data. The hyperparameters of all baseline methods were set according to the guidelines in the original study to ensure a fair comparison environment. The X-ray system was configured in the experiment to emit 1e4, 5e3, and 1e3 photons per path, replicating the LDCT acquisition conditions. Table 1 presents the quantitative evaluation results, including the PSNR, SSIM, and MSE of the reconstructed images; the optimal values ​​are marked in bold.

[0149] Table 1: PSNR / SSIM / MSE of different methods for reconstructing AAPM challenge data at different noise levels

[0150]

[0151] To comprehensively evaluate the proposed method, this study assesses the performance of each algorithm from both visual and quantitative dimensions. Table 1 presents detailed evaluation results under different noise levels, recording PSNR, SSIM, and MSE. PLOT-CT achieved the highest PSNR and SSIM, and the lowest MSE across all scenarios. Specifically, Figures 3 to 5 The diagram presents reconstructed images and their corresponding residual maps under multiple noise levels. These residual maps reflect the distribution of reconstructed noise and capture performance differences when processing complex textures or edge regions. This format allows for direct comparison of noise suppression and detail preservation capabilities, enabling researchers to observe the balance between noise reduction and feature preservation for each method without relying solely on quantitative metrics. This visual insight, combined with the quantitative advantages shown in Table 1, validates the strengths of the experimental results and provides a reference for subsequent algorithm optimization.

[0152] Among all the comparative algorithms tested in this study, Pix2Pix exhibited significant structural defects even at relatively low noise levels. NCSN++ often over-smoothed edges during denoising, rendering it almost ineffective at removing noise under extremely high noise levels and frequently introducing stripe artifacts in high-contrast regions. SDCNN achieved only moderate denoising, failing to fully recover fine textures and potentially leaving subtle artifacts in the reconstructed image. Re-UNet maintained basic structural integrity under moderate noise, but still couldn't fully capture texture details. CaGAN struggled to recover complex structural details and couldn't accurately reproduce complex textures and hierarchical structures. WiTUnet performed exceptionally well, ranking second only to PLOT-CT, but still exhibited some noise sensitivity, resulting in slight artifacts and structural incompleteness in the reconstruction, failing to accurately recover fine image details. Under all tested noise levels and scenarios, PLOT-CT consistently outperformed the comparative methods in both quantitative metrics and visual quality, demonstrating robust stability.

[0153] To better explore the generalization ability of PLOT-CT, various comparison methods were applied to the CHAOS dataset, a widely used benchmark dataset for medical imaging covering diverse clinical abdominal imaging scenarios to ensure comprehensive evaluation. Furthermore, as an unsupervised model, NCSN++ was trained on 2200 images from the AAPM challenge dataset, while all other comparison models were trained at noise levels of 1e4, 5e3, and 1e3 to simulate real-world low-dose CT acquisition conditions. Table 2 presents detailed quantitative evaluation results for Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Mean Squared Error (MSE). PLOT-CT outperformed all comparison methods across all key performance metrics, demonstrating robust generalization and excellent performance. Moreover, Figure 6 Representative reconstruction results and noise maps of PLOT-CT, SDCNN and other comparative methods are presented to visually verify the effectiveness of the proposed method.

[0154] Table 2: PSNR / SSIM / MSE of CHAOS data reconstructed using different methods at different noise levels

[0155]

[0156] The pre-log domain Voronoi decomposition-assisted low-dose CT reconstruction method provided in Embodiment 1 of this invention is the first to introduce a Voronoi decomposition mechanism in the pre-log projection domain. By using K-means clustering, the pre-log sine wave is divided into different clusters according to intensity. Within each cluster, the numerical scale tends to be balanced and the gradient remains stable, effectively solving the problems of large dynamic range and non-stationary gradients in the pre-log domain. This allows the deep learning model to focus on the features themselves rather than being dominated by the background. The invention employs a latent space mapping strategy, mapping different clusters to independent latent spaces, achieving effective decoupling of features and avoiding information aliasing between different intensity regions, thus facilitating subsequent accurate reconstruction. Denoising lays the foundation; this invention incorporates latent space representation into a transformer-enhanced diffusion model, utilizes multi-head transpose attention to capture global dependencies, and employs a gated feedforward network to enhance feature discriminability, ensuring structural fidelity while maintaining global consistency; through mathematical proof and experimental verification, this invention determines that the optimal number of clusters for Voronoi decomposition is 3, a choice that achieves the optimal balance between reconstruction noise and model complexity, demonstrating theoretical completeness and sufficient experimental results; this invention effectively suppresses the stepwise amplification effect of pre-logarithmic domain noise during logarithmic transformation and filtered backprojection, significantly improving the peak signal-to-noise ratio and structural similarity index of low-dose CT reconstructed images.

[0157] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0158] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A low-dose CT reconstruction method assisted by pre-logarithmic domain Voronoi decomposition, characterized in that, include: The pre-log sinusoidal plot of low-dose CT was obtained, and the Voronoi decomposition of the pre-log sinusoidal plot was performed using the K-means clustering algorithm to obtain several feature clusters. Each feature cluster is mapped to its corresponding latent space to obtain a multi-channel input feature map; The multi-channel input feature map is input into the pre-trained diffusion transformer model for iterative denoising and optimization to obtain optimized clustering features; The optimized clustering features are fused and subjected to logarithmic transformation to obtain a post-logarithmic sine curve. A filtered back-projection algorithm is used to reconstruct the post-logarithmic sine curve to obtain a high-quality CT image; The step of inputting the multi-channel input feature map into the pre-trained diffusion transformer model for iterative denoising and optimization to obtain optimized clustering features includes: The multi-channel input feature map is input into the pre-trained diffusion transformer model, and then processed by the noise estimation network in the pre-trained diffusion transformer model. Estimating the noise of the multi-channel input feature map : ； In the formula, For the first Each feature cluster at time step The noisy characteristics, The first of the multi-channel input feature maps One channel; According to time steps Perform reverse denoising updates sequentially: ， ； In the formula, For the first Each feature cluster at time step The noisy characteristics, These are the signal preservation coefficient and the noise variance, respectively. Repeat the inverse denoising process until a clean latent vector is output. ; The clean hidden vector In the input multi-head transpose attention module, for the clean latent vector Query is generated using 1×1 pointwise convolution and 3×3 depthwise convolution. ,key ,value ; Query ,key ,value Divided into There are 1 attention head, and a scaled dot product attention is computed for each attention head. : ； In the formula, The first The query features, key features, and value features corresponding to each attention head. The dimension of the attention head; The scaled dot product attentions of all attention heads are concatenated to obtain the MTA output. : ； In the formula, For splicing operations; Output the MTA The input is gated in a gated feedforward network to obtain optimized clustering features: ； In the formula, These are 1×1 pointwise convolution kernels for the first and second groups, respectively. These are 3×3 depth convolution kernels for the first and second groups, respectively. For element-wise multiplication, is the activation function for the Gaussian error linear unit.

2. The low-dose CT reconstruction method assisted by pre-logarithmic domain Voronoi decomposition according to claim 1, characterized in that, The steps of obtaining the pre-log sinusoidal plot of low-dose CT and performing Voronoi decomposition on the pre-log sinusoidal plot using the K-means clustering algorithm to obtain several feature clusters include: Obtain raw photon count data from low-dose CT scans, and determine the photon intensity data based on the raw photon count data: ； In the formula, For projection angle and detector position The photon intensity at that location was measured. The intensity of the incident photon. Integrating for noise-free tissue attenuation lines, Poisson measurement noise inherent in photon counting; The photon intensity data are arranged according to the projection angle and detector channel to obtain a pre-logarithmic sine curve for low-dose CT. The K-means clustering algorithm is used to perform Voronoi decomposition on the prelog sine graph to obtain several feature clusters, including high-intensity features, medium-intensity features, and low-intensity features.

3. The low-dose CT reconstruction method assisted by pre-logarithmic domain Voronoi decomposition according to claim 1, characterized in that, The step of mapping each feature cluster to its corresponding latent space to obtain a multi-channel input feature map includes: Each feature cluster is input into the hidden encoder. Mapping to the corresponding latent space to obtain a low-dimensional vector representation. : ； In the formula, For the first Clustering based on features; The low-dimensional vector representations are concatenated along the channel dimensions to obtain a multi-channel input feature map. : ； In the formula, For splicing operations, These are the low-dimensional vector representations corresponding to the first, second, and third feature clusters, respectively.

4. The low-dose CT reconstruction method assisted by pre-logarithmic domain Voronoi decomposition according to claim 1, characterized in that, The step of fusing the optimized clustering features and performing a logarithmic transformation to obtain the post-logarithmic sine curve includes: Optimize clustering features The fusion is performed to obtain a prelogarithmic sine curve. : ； Perform a logarithmic transformation on the pre-logarithmic sine graph to obtain the post-logarithmic sine graph. : ； In the formula, The intensity of the incident photon.

5. The low-dose CT reconstruction method assisted by pre-logarithmic domain Voronoi decomposition according to claim 1, characterized in that, In the step of reconstructing the post-logarithmic sine wave using a filtered back-projection algorithm to obtain a high-quality CT image, the high-quality CT image... The output process is as follows: ； In the formula, This indicates the filter used in the FBP. For the set of projection angles, This is for filtered back projection.

Images