CT image reconstruction method based on progressive texture perception diffusion model

Abstract

The invention relates to the technical field of medical image processing, in particular to a CT image reconstruction method based on a progressive texture perception diffusion model, and the method comprises the steps: inputting a to-be-reconstructed low-dose CT image or a sparse view angle image and projection image data into a trained progressive texture perception diffusion model, and outputting a reconstructed CT image; the progressive texture perception diffusion model carries out preliminary reconstruction on a to-be-reconstructed low-dose CT image or a sparse view angle image to obtain a low-frequency structure of the image; based on the low-frequency structure, high-frequency details are generated through conditional Schrodinger bridge diffusion network iteration, multi-scale high-frequency features in the high-frequency details are extracted, and a reconstructed CT image is generated based on the low-frequency structure of the image and the multi-scale high-frequency features. The invention provides a CT (Computed Tomography) reconstruction method based on coarse-fine segmentation, so that artifacts are removed and more tiny tissue textures are reserved on the premise of reserving global structures such as key bones and blood vessels.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method and apparatus for CT image reconstruction using a progressive texture-aware diffusion model. Background Technology

[0002] In modern medicine, computed tomography (CT) is widely used in clinical settings such as tumor diagnosis, cardiovascular imaging, and trauma assessment due to its high resolution and rapid scanning capabilities. However, conventional CT scans using high-dose X-rays can lead to excessive cumulative radiation doses for patients, increasing the risk of cancer. To reduce this dose, low-dose CT and sparse projection CT techniques have been proposed. Low-dose CT refers to computed tomography that significantly reduces the radiation dose received by patients. Sparse projection CT is a CT technique that reconstructs images of the internal structure of an object even when projection data (i.e., X-ray images taken from different angles) is incomplete or severely lacking. Here, "sparse" refers to the fact that the number of projection angles is much smaller than in traditional CT, or that projection data is missing at certain angles. The challenge of sparse projection CT lies in the small number of projection angles (e.g., only a few dozen or even a few), resulting in severely insufficient data. If the FBP algorithm is used directly for reconstruction, it will produce a large number of artifacts (stripes, distortions, etc. in the image), leading to poor image quality. The commonality between low-dose CT and sparse projection CT techniques is that, after the projection dose or the number of viewing angles is reduced, strong noise and streak artifacts will appear in the reconstructed images, which will seriously affect the diagnostic results.

[0003] Traditional imaging model-based reconstruction methods (such as FBP and iterative reconstruction) can partially suppress noise in low-dose scenarios, but their recovery of structural and texture details is limited. In recent years, deep learning methods (CNN, GAN) have been introduced into the field of CT reconstruction: CNN-based denoising methods (such as RED-CNN) can effectively reduce noise, but due to pixel-level loss constraints, they often lose tissue details; GAN-based methods (such as WGAN-VGG) generate visually more realistic textures, but training is unstable and prone to producing false structures; although diffusion model-based methods have shown excellent detail generation capabilities in natural and medical imaging, their direct application to low-dose CT reconstruction still faces a dual bottleneck: first, the randomness of the generation process can easily lead to structural instability phenomena such as blood vessel drift and cortical bone fracture; second, iterative sampling requires dozens of steps to obtain one image, and the computational time is difficult to meet the real-time clinical needs. Therefore, there is an urgent need to construct a reconstruction paradigm that can quickly inject texture details while preserving a faithful global anatomical framework, in order to meet the stringent clinical requirements for high-quality low-dose CT. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies in balancing structural fidelity and texture detail, this invention proposes a CT image reconstruction method and apparatus based on a progressive texture-aware diffusion model.

[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0006] The CT image reconstruction method based on the progressive texture-aware diffusion model provided in this application includes the following steps: The low-dose CT image to be reconstructed, the sparse-view CT image, or the sparsely sampled sinusoidal data is input into the trained progressive texture-aware diffusion model, and the reconstructed CT image is output. The progressive texture-aware diffusion model performs preliminary reconstruction on the low-dose CT image, sparse-view CT image, or sparsely sampled sinusoidal data to be reconstructed, obtaining the low-frequency structure of the image. Based on the low-frequency structure, high-frequency details are generated iteratively through a conditional Schrödinger bridge diffusion network, multi-scale high-frequency features are extracted from the high-frequency details, and a reconstructed CT image is generated based on the low-frequency structure of the image and the multi-scale high-frequency features.

[0007] This invention employs a two-stage "coarse-fine" strategy: first, deterministic reconstruction is performed on low-dose or sparse view data through preliminary reconstruction to accurately restore the low-frequency structure of the image; then, conditional diffusion is performed to generate high-frequency details. This divide-and-conquer approach ensures stable reconstruction of key structures such as bones and blood vessels, while enhancing texture representation through the diffusion process, overcoming the limitations of a single network in simultaneously achieving global consistency and local detail.

[0008] Preferably, the network used in the initial reconstruction includes CNN network, ResNet network, VGG network, DenseNet network, EfficientNet network, RED-CNN network, or LEARN network.

[0009] Preferably, the extraction of multi-scale high-frequency features from the high-frequency details specifically includes: The high-frequency details are downsampled, and the sampling results are input into a lightweight super-resolution U-shaped network to predict the high-frequency residual features at each scale. The texture features output from the intermediate layer of the lightweight super-resolution network are aligned by convolution and then output to the corresponding layer of the encoder or decoder of the lightweight super-resolution U-shaped network.

[0010] Preferably, the lightweight super-resolution U-shaped network has a structure of 3 layers of 3×3 convolutions, uses the ReLU activation function, and the last layer outputs a high-frequency residual with the same dimension as the downsampled image.

[0011] Preferably, the lightweight super-resolution U-shaped network uses downsampled MSE to calculate the guiding loss. MSE stands for Mean Square Error.

[0012] Preferably, the progressive texture-aware diffusion model is trained in a supervised manner, using the MSE loss function.

[0013] Preferably, during the training process of the progressive texture-aware diffusion model, a training dataset and a validation dataset are constructed. The training dataset includes a low-dose CT training dataset and a sparse-view CT training dataset; the validation dataset includes a low-dose CT validation dataset and a sparse-view CT validation dataset. The low-dose CT training dataset includes multiple sets of low-dose and normal-dose CT images; the sparse-view CT training dataset includes sparse-view CT images, sparsely sampled sine wave data, and full-view sampled CT images. The low-dose CT validation dataset includes multiple sets of low-dose CT images; the sparse-view CT validation dataset includes sparsely sampled sine wave data and full-view sampled CT images.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention provides a CT reconstruction method based on coarse-fine treatment. By first performing deterministic reconstruction on low-dose / sparse view data, and then refining high-frequency details with conditional diffusion, the structural prior is effectively constrained in a regularized manner to limit the reconstruction solution space, thereby removing artifacts and preserving more micro-tissue textures while retaining key global structures such as bones and blood vessels. Attached Figure Description

[0015] Figure 1 This is a flowchart of the CT image reconstruction method using the progressive texture-aware diffusion model in Example 1; Figure 2 The CT image reconstruction device is the progressive texture-aware diffusion model described in Example 1. Figure 3 This is the schematic diagram of Example 2 and the corresponding PTDrec and PTDdiff modules; Figure 4 The images are CT images taken at 5% dose in Example 3, along with the reconstruction results of each method. Figure 5 The evaluation indicators of the reconstruction results of the method of the present invention in Example 2 and other low-dose CT methods are compared. Figure 6 The images shown are 32-view sparse-view CT images from Example 3 and the reconstruction results of each method. Figure 7This is a comparison of the evaluation metrics of the reconstruction results of Test Example 2 in Example 3 with those of other sparse view CT methods. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to specific embodiments. However, this should not be construed as limiting the scope of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0017] Applying diffusion models to natural and medical imaging can reconstruct good details, but they still face computational challenges and struggle to meet the demands of real-time clinical image generation. A balance between structural fidelity and texture detail cannot be achieved. Therefore, a progressive texture-aware diffusion model has been proposed to further improve the diffusion model. This model performs preliminary reconstruction on low-dose CT images, sparse-view CT images, or sparsely sampled sinusoidal data to obtain the low-frequency structure of the image. Based on this low-frequency structure, a conditional Schrödinger bridge diffusion network iteratively generates high-frequency details, extracts multi-scale high-frequency features from these details, and generates a reconstructed CT image based on the low-frequency structure and these multi-scale high-frequency features. First, a preliminary reconstruction is performed on the low-dose or sparse-view data to perform deterministic reconstruction, accurately restoring the low-frequency structure of the image. Then, conditional diffusion is performed to generate high-frequency details. This divide-and-conquer approach ensures stable reconstruction of key structures such as bones and blood vessels while enhancing texture representation through the diffusion process, overcoming the limitations of a single network in balancing global consistency and local detail. Based on this, further optimizations were proposed in several aspects, including the selection of the optimal network for initial reconstruction, the extraction of multi-scale high-frequency features from the high-frequency details, and the construction of training and validation datasets. This makes the scheme not only theoretically feasible but also quickly applicable to clinical practice, meeting the stringent clinical requirements for high-quality, low-dose CT.

[0018] Example 1 In the first aspect, a CT image reconstruction method based on a progressive texture-aware diffusion model is described in the flowchart below. Figure 1 As shown, it includes the following steps: The low-dose CT image to be reconstructed, the sparse-view CT image, or the sparsely sampled sinogram data (i.e., sinogram data) is input into the trained progressive texture-aware diffusion model, and the reconstructed CT image is output. The progressive texture-aware diffusion model performs preliminary reconstruction on the low-dose CT image, sparse-view CT image, or sparsely sampled sinusoidal data to be reconstructed, obtaining the low-frequency structure of the image. Based on the low-frequency structure, high-frequency details are generated iteratively through a conditional Schrödinger bridge diffusion network, multi-scale high-frequency features are extracted from the high-frequency details, and a reconstructed CT image is generated based on the low-frequency structure of the image and the multi-scale high-frequency features.

[0019] Furthermore, training the progressive texture-aware diffusion model includes the following steps: S1: Collect multiple sets of low-dose CT images, normal-dose images, sparse-view CT images, sparsely sampled sinogram data, and full-view sampled CT images, and construct training and validation datasets. The collected images are used to train low-dose and sparse-view CT reconstruction models, respectively. The low-dose and sparse-view CT reconstruction models belong to the progressive texture-aware diffusion model, which is a further subdivision of the progressive texture-aware diffusion model for the input image type. The low-dose and sparse-view CT reconstruction models are consistent in model structure.

[0020] The low-dose CT training dataset includes multiple sets of low-dose and normal-dose CT images, the low-dose CT validation dataset includes multiple sets of low-dose CT images, the sparse-view CT training dataset includes sparse-view CT images, sparsely sampled sinogram data, and full-view sampled CT images, and the sparse-view CT validation dataset includes sparsely sampled sinogram data and full-view sampled CT images.

[0021] Sparse-view CT images refer to images with severe artifacts reconstructed due to insufficient sampling (sparse sampling). Sparsely sampled sinogram data refers to the original data actually acquired by the CT scanning device with insufficient projection angle. The progressive texture-aware diffusion model of the present invention can reconstruct the original data actually acquired by the CT scanning device with insufficient projection angle (i.e., sparsely sampled sinogram data) to generate full-view sampled CT images, and it can also reconstruct the sparse-view CT images that have already been generated and contain severe artifacts due to insufficient sampling.

[0022] S2: Construct a progressive texture-aware diffusion model. The progressive texture-aware diffusion model first performs preliminary reconstruction to obtain the low-frequency structure of the image; based on the low-frequency structure, high-frequency details are generated iteratively through a conditional Schrödinger bridge diffusion network, multi-scale high-frequency features are extracted from the high-frequency details, and a reconstructed CT image is generated based on the low-frequency structure of the image and the multi-scale high-frequency features.

[0023] The Progressive Texture-Aware Diffusion (PTD) model consists of two modules: the preliminary reconstruction module (PTDrec) and the diffusion refinement module (PTDdiff).

[0024] S3: Train the low-dose CT image diffusion reconstruction model and the sparse-view CT image diffusion reconstruction model using the image groups in the training dataset, evaluate the model using the validation dataset, and adjust the hyperparameters to obtain the trained image reconstruction model.

[0025] This invention employs a two-stage "coarse-fine" strategy: first, a preliminary reconstruction module (PTDrec) performs deterministic reconstruction on low-dose CT images or sparsely sampled sinogram data to accurately recover the low-frequency structure of the image; then, a diffusion refinement module (PTDdiff) performs conditional diffusion on this basis to generate high-frequency details. This method ensures stable reconstruction of key structures such as bones and blood vessels, while enhancing texture representation through the diffusion process, overcoming the limitation of a single network in simultaneously achieving global consistency and local detail.

[0026] Reconstruction models based on low-dose and sparse perspectives were set up separately, making the models more closely resemble real-world application scenarios and more targeted and practical. Since both the training and validation sets are constructed based on low-dose and sparse perspective projection images, the models can learn during training how to reconstruct high-quality CT images under different noise conditions. This gives the models better generalization ability and robustness when processing real-world CT data.

[0027] Secondly, CT image reconstruction devices based on a progressive texture-aware diffusion model, such as... Figure 2 As shown, it includes a data input module and a model processing module. The data input module is used to input the low-dose CT image to be reconstructed, the sparse-view CT image, or the sparsely sampled sinogram data into the model module. The model processing module stores a trained progressive texture-aware diffusion model and outputs reconstructed CT images. The progressive texture-aware diffusion model performs preliminary reconstruction on the low-dose CT image, sparse-view CT image, or sparsely sampled sinusoidal data to be reconstructed, obtaining the low-frequency structure of the image. Based on the low-frequency structure, high-frequency details are generated iteratively through a conditional Schrödinger bridge diffusion network, multi-scale high-frequency features are extracted from the high-frequency details, and a reconstructed CT image is generated based on the low-frequency structure of the image and the multi-scale high-frequency features.

[0028] The modular design described above ensures high fidelity of the overall model structure, effectively restores detailed textures, and meets the needs of real-time clinical reconstruction with minimal diffusion steps.

[0029] Preferably, the model processing module includes a preliminary reconstruction module and a conditional diffusion refinement module; the preliminary reconstruction module includes a CNN network for handling different noise levels; the conditional diffusion refinement module includes a conditional diffusion module and a multi-scale high-frequency texture guidance module; the conditional diffusion module iteratively generates high-frequency details through a conditional Schrödinger bridge diffusion network. The multi-scale high-frequency texture guidance module extracts multi-scale high-frequency features from the high-frequency details.

[0030] The model processing module is explained in detail below: The model processing module consists of two main parts: a preliminary reconstruction module and a conditional diffusion refinement module. First, the PTDrec (preliminary reconstruction module) uses a deterministic CNN network (such as RED-CNN / LEARN) to perform preliminary reconstruction on low-dose or sparsely sampled sinogram data, accurately restoring the low-frequency structure of the image. Then, the PTDdiff (conditional diffusion refinement module), within the Diffusion Schrödinger Bridge framework, uses the results of the preliminary reconstruction and its multi-scale high-frequency texture features as dual-domain conditions, employing a lightweight U-Net network-based multi-scale high-frequency texture guidance module (MHFG) to generate high-frequency details. By simultaneously optimizing the reconstruction loss, texture guidance loss, and diffusion loss, a balanced reconstruction of high-fidelity structure and rich details is achieved.

[0031] In terms of detail generation, the PTD model includes a multi-scale high-frequency texture guidance (MHFG) module. This module injects texture features from the initial reconstruction at different scales as "implicit cues" into the MHFG module, extracting multi-scale features to preserve realism while suppressing noise. End-to-end joint training combines reconstruction loss, texture guidance loss, and diffusion loss to achieve an optimal balance between structural fidelity and visual quality. Furthermore, the efficient sampling of only 1, 5, or 8 steps significantly reduces computational costs, meeting the needs of real-time clinical reconstruction.

[0032] Preferably, the proposed preliminary reconstruction module (PTDrec module) is a deterministic predictor used to directly recover the low-frequency signals (i.e., smooth structure and texture) of an image from low-dose / sparsely sampled sinogram data. Existing CNN networks (such as ResNet, VGG, DenseNet, EfficientNet, RED-CNN, or LEARN) can be flexibly selected to obtain preliminary estimates. , Input to the network: low-dose or sparsely sampled sinogram data (including noise / artifacts); () represents a deterministic CNN network with parameter θ, where θ is a symbol. The unified name for all parameters in a function; X C The initial reconstruction result is a "clean" image containing only low-frequency structures and smooth textures. Mean squared error is used to ensure high fidelity of the overall structure and texture, providing a reliable prior for subsequent high-frequency generation. The mean squared error is... Here, X0 is the target image, or a standard-dose CT image in CT reconstruction, which serves as a key reference point during the generation process; Xc (i.e., the initial coarse prediction image) is a coarse approximation of X0 generated by the reconstruction module PTDrec. This initial prediction Xc provides a starting point for the subsequent diffusion process, helping to generate finer high-frequency details and ultimately obtain a result that is closer to X0.

[0033] Furthermore, after the initial reconstruction module, the CT image first enters the Conditional Diffusion Refinement (PTDdiff) module and then the Multi-Scale High-Frequency Texture Guidance (MHFG) module, forming a two-stage processing. The PTDdiff module is essentially a lightweight U-Net network with four layers each for the encoder and decoder, and the number of channels is [32, 64, 128, 256]. A MHFG module is inserted at each encoding / decoding stage to extract texture biases (texture biases being multi-scale high-frequency texture features) based on the output of the initial reconstruction module. Within the Diffusion Schrödinger Bridge (DSB) framework, PTDdiff introduces a dual-domain (pixel domain + feature domain) condition, guiding the model to focus on high-frequency detail reconstruction.

[0034] First, the initial stage involves the CT image entering the Conditional Diffusion Refinement Module (PTDdiff module). The core idea of ​​a typical diffusion model is to train a network to predict noise, thereby gradually denoising and restoring the image. This invention extends the traditional diffusion model into a conditional diffusion model, introducing multiple conditional information to guide the noise prediction process, thereby improving the structural consistency and texture details of the reconstructed image. In the conditional diffusion model, noise is predicted and removed through a rewritten noise function. During training, the rewritten noise function helps reduce the image quality degradation caused by low-dose noise while restoring image details and structure. By rewriting the noise function, the model can more accurately predict and eliminate noise in the image, improving the quality of the reconstructed image. The output of the rewritten noise function is the "denoising result" at a certain time t during the diffusion process. This output is used in the denoising step and is input into the next stage of the diffusion model, namely, output to the Multi-Scale High-Frequency Texture Guidance Module (MHFG) to generate high-frequency details.

[0035] X C Together with the original low-dose image, the noise prediction function is rewritten as follows: ; X C This is the preliminary reconstruction result, where t is time. For the image with noise added at time t, the network focuses on removing noise and artifacts, and further refines it towards high-frequency details. This is an image at a certain time step during the diffusion process. As time step t progresses, it gradually moves from its initial noisy state towards the target image X0. The target image and... They are directly related. It gradually approaches the intermediate state of X0 during the diffusion process, while X... C This is a preliminary estimate of X0 obtained through the reconstruction module.

[0036] Secondly, the processing results of the conditional diffusion refinement module are input into the multi-scale high-frequency texture guidance module to refine the preliminary reconstruction result X. C Perform separately Downsampling (k=1, 2, 3, 4) is performed and the input is fed into a lightweight super-resolution subnet. Each subnetwork learns the downsampled high-frequency texture estimate and adds its extracted intermediate features to the corresponding encoder / decoder layer of the diffusion network (aligning channels via 1×1 convolutions) as "implicit cues". The supervision loss for texture at each scale is: ; Where k is the scale index, k = 1,2,3,4, corresponding to 1 / 2, 1 / 4, 1 / 8, 1 / 16 downsampling; ()for Double downsampling operation; () represents a lightweight image supramolecular network that predicts high-frequency textures at this scale. X0 is the target image, which is the normal-dose CT image corresponding to the low-dose CT image.

[0037] Ultimately, the multi-scale guided loss is: ; Under the above conditions, the denoising loss (i.e., diffusion loss) of the diffusion model is defined as: ; Where t is the diffusion time step, and samples are uniformly taken in [0,1]. The image after adding noise at time t; The input contains noise for the network's prediction. Time t, preliminary reconstruction result Xc, and multi-scale texture features; This is the noise scheduling coefficient, which controls the noise intensity in this step. The denoising loss of this diffusion model is used to enhance the model's high-frequency generation capability.

[0038] Finally, after end-to-end joint training and inference, a fusion loss function is used: ; in, It is the fusion loss function. It is a reconstruction loss. It is texture-guided loss. It is a diffusion loss.

[0039] The three types of losses are clearly defined: reconstruction losses It only ensures that the initial reconstruction module PTDrec correctly reconstructs the low-frequency skeleton; texture-guided loss. A 4-scale lightweight super-resolution network was used during the training phase. The "standard high-frequency residual" is calculated and inserted into U-Net as an implicit cue, then discarded during inference; diffusion loss. Using the coarse reconstruction result Xc and the already generated texture features, the conditional diffusion refinement module PTDdiff is trained to predict noise during training. During inference, only 1 / 5 / 8 sampling steps are needed to restore the noisy map into a detailed and structurally accurate CT image. Directly discarding during inference means that the network has already been trained and learned how to recover these high-frequency details during the inference stage, so it no longer needs to rely on additional texture-guided loss (i.e., the MHFG module) to help it generate images. The goal of the inference stage is to directly obtain the final reconstruction result from the network. At this point, the texture-guided module no longer plays a role. To reduce computation and improve inference efficiency, the texture cues used to guide the network during training can be discarded.

[0040] This fusion loss function is used to simultaneously optimize the initial reconstruction module PTDrec and the conditional diffusion refinement module PTDdiff, enabling synergistic gains in the "coarse reconstruction first, then refined generation" workflow. Through this modular design, the model can ensure high fidelity of the overall structure while effectively restoring detailed textures, and meet the needs of real-time clinical reconstruction with minimal diffusion steps.

[0041] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention provides a CT reconstruction method based on coarse-fine treatment. It first performs deterministic reconstruction on low-dose sparse-view CT images or sparsely sampled sinogram data, and then refines high-frequency details with conditional diffusion. This allows the structural prior to effectively limit the reconstruction solution space in a regularized manner, thereby removing artifacts and preserving more micro-tissue textures while retaining key global structures such as bones and blood vessels.

[0042] 2. This invention provides a multi-scale texture cueing strategy. By extracting high-frequency texture features from the preliminary reconstruction results at different downsampling scales and injecting them layer by layer into the encoder-decoder module of the diffusion network, the deep feature interaction and progressive transmission of prior information between iterative blocks are realized, which significantly enhances the cueing effect and makes fuller use of the prior noise distribution, thereby improving the overall reconstruction performance of the model.

[0043] 3. This invention provides a lightweight cross-conditional diffusion module that, by fusing low-frequency reconstruction results with multi-scale texture cues in U-Net, not only accurately captures local-global dependencies in the feature dimension, but also significantly reduces the number of parameters and computational load, enabling prior information to be more efficiently integrated into image features, and significantly improving noise artifact removal capabilities and inference efficiency.

[0044] 4. This invention provides an end-to-end joint optimization method that incorporates reconstruction loss, texture guidance loss, and diffusion denoising loss into a unified objective, and collaboratively trains the PTDrec and PTDdiff modules. This simplifies the traditional multi-stage training pipeline and promotes mutual priors and collaborative gains among the sub-modules, achieving the best trade-off between structural fidelity and rich detail.

[0045] 5. This invention provides a modular and replaceable overall architecture. PTDrec can be replaced with any advanced CNN reconstruction network, and the depth and channels of PTDdiff and multi-scale subnetworks can also be flexibly adjusted, which facilitates rapid integration and customized optimization with existing CT imaging systems to meet the real-time high-quality reconstruction needs of different clinical scenarios.

[0046] Example 2 A CT image reconstruction method based on a progressive texture-aware diffusion model includes the following steps: A1: Collect multiple sets of low-dose CT images, normal-dose images, sparse-view CT images, sparsely sampled sinogram data, and full-view sampled CT images to train low-dose and sparse-view CT reconstruction models, respectively, and construct training and validation datasets. The low-dose CT training dataset includes multiple sets of low-dose and normal-dose CT images, and the low-dose CT validation dataset includes multiple sets of low-dose CT images. The sparse-view CT training dataset includes sparse-view CT images, sparsely sampled sinogram data, and full-view sampled CT images, and the sparse-view CT validation dataset includes sparsely sampled sinogram data and full-view sampled CT images.

[0047] A2: Construct a progressive texture-aware diffusion reconstruction model, which consists of a preliminary reconstruction module and a conditional diffusion refinement module. The preliminary reconstruction module consists of a CNN network that processes different noise levels. The conditional diffusion refinement module consists of a conditional diffusion module and a multi-scale high-frequency texture guidance module. The conditional diffusion module uses a conditional Schrödinger Bridge diffusion network to iteratively generate high-frequency details to achieve texture enhancement. The multi-scale high-frequency texture guidance module extracts high-frequency textures at each scale through a lightweight super-resolution subnet U-Net network and injects intermediate-scale features from the multi-scale high-frequency features into the encoder and decoder layers of the U-Net network as "implicit cues".

[0048] A3: The model is trained in a supervised manner, using the MSE loss function to ensure that the error between the model output and the real CT image is minimized.

[0049] A4: Input the low-dose CT image to be reconstructed or the sparsely sampled sinogram data to be reconstructed to obtain the final reconstructed CT image.

[0050] In an optional implementation, step A1 can further optimize the dataset selection and preprocessing process to enhance the model's generalization ability and robustness. Specifically, this can be achieved by enriching the diversity and scale of the dataset, or by introducing various data augmentation techniques to expand and transform the original data.

[0051] In an optional implementation, step A2 can further adjust the model structure and parameters. For example, the CNN network model of the basic reconstruction network can be replaced to adapt to different noise patterns, or parameters such as the activation function can be adjusted to optimize the model's performance. Furthermore, other types of reconstruction modules or regularization methods can be explored to further improve the reconstruction effect. According to the above description, the CNN network of the initial reconstruction module can be set to either RED-CNN or LEARN network, respectively, to adapt to low-dose or sparse-view CT images, conforming to the following formula: ; in, It is to recover the low-frequency components of an image from the original projection data of a low-dose or sparse view, X N These are low-dose CT images or sparsely sampled sinogram data to be reconstructed. This represents the CNN network in the initial reconstruction module.

[0052] Figure 3 It is the principle framework diagram and the corresponding PTDrec and PTDdiff modules, in Figure 3In this module, the conditional diffusion refinement module consists of two important sub-modules: the conditional diffusion module and the multi-scale high-frequency texture guidance module. The conditional diffusion module, also known as PTDdiff, operates within the improved Diffusion Schrödinger Bridge framework, from... Initially, noise is gradually added to obtain a series of intermediate variables: ; The above process is the forward process of the conditional diffusion model, where { } is the preset noise scheduling. For the image after adding noise at time t, The image after adding noise at time t-1.

[0053] Correspondingly, the inverse process is the model learning the inverse distribution of parameters: ; The training is performed by minimizing the MSE of the reconstructed real noise.

[0054] By combining the conditional information from the pixel domain and the feature domain, high-frequency details are supplemented to the output of the initial reconstruction module PTDrec, and the noise prediction model is modified as follows: ; X C This is the preliminary reconstruction result, where t is time. The image after adding noise at time t.

[0055] Multi-scale high-frequency texture guidance module Downsampling is performed proportionally, as follows: The values ​​(k=1, 2, 3, 4) are input to their respective lightweight super-resolution networks (texture subnets). The structure consists of 3 layers of 3×3 convolutions + ReLU, with [16, 32, 16] channels. The last layer outputs a high-frequency residual of the same dimension as the downsampled image, predicting the high-frequency residual features at this scale. The texture features output from the intermediate layers of the lightweight super-resolution network are aligned through 1×1 convolutions and then concatenated to the corresponding layers of the U-Net encoder / decoder; achieving dual cues of "pixel domain + feature domain." This module significantly improves the recovery quality of high-frequency edges and fine blood vessels. Each texture sub-network... Calculate the guiding loss by pressing the sampled MSE: ; Under the above conditions, PTDdiff learns to predict noise: ; This MSE loss-driven network progressively recovers clean images at different noise levels t. High-quality images can be generated in just 1, 5, or 8 steps of inverse diffusion, significantly reducing inference time compared to traditional diffusion models (which require tens to hundreds of steps).

[0056] The sampling process can be summarized as: from pure noise (T can be set to 100) Start; iterate in reverse according to the set number of steps. Ultimately, a high-quality reconstruction was achieved. .

[0057] Example 3 To verify the effectiveness and superiority of the reconstruction method proposed in Example 1, this example selected six other existing reconstruction methods for comparative experiments. Each method was trained according to the experimental settings of this invention to ensure the fairness of the comparison between the methods.

[0058] Figure 3 In this model, XT represents the initial input, which is then gradually updated through diffusion at each step to approach the target image X0. Therefore, the state of the image at different time steps is labeled Xt, showing its evolution throughout the generation process. In other words, Xt represents an intermediate state. Figure 3 middle Compared with Example 2 They are the same object; the conditions are Xc.

[0059] The six comparison methods are: Method 1, FBP proposed by Ramachandran et al., Reference Ramachandran, GN, & Lakshminarayanan, AVThree-dimensional reconstruction from radiographs andelectron micrographs: Application of convolutions instead of Fouriertransforms. Proceedings of the National Academy of Sciences, 68(9), 2236–2240(1971); Hu Chen, Yi Zhang, Mannudeep KKalra, Feng Lin, Yang Chen, Peixi Liao, Jiliu Zhou, and Ge Wang, “Low-dose ctwith residual neural encoder evolution network,”IEEEtransactions on medical imaging, vol. 36, no. 12, pp. 2524–2535(2017); Qingsong Yang, Pingkun Yan, Yanbo Zhang, Hengyong Yu, Yongyi Shi, Xuanqin Mou, Mannudeep K Kalra, YiZhang and Ling Sun, WGAN-VGG, WGAN-VGG, c. denoising using a generativeadversarial network with wasserstein distance and perceptual loss,”IEEEtransactions on medical imaging, vol. 37, no. 6, pp. 1348–1357(2018); Mufeng Geng, Xiangxi Meng, JiangyuanYu, Lei Zhu, Lujia Jin, Zhe Jiang, Bin Qiu, Hui Li, Hanjing Kong, JianminYuan, et al. medical imagedenoising,”IEEE transactions on medical imaging, vol. 41, no. 2, pp. 407–419(2021); Method 5, CoreDiff proposed by Gao et al., reference Qi Gao, Zilong Li, Junping Zhang, Yi Zhang, and Hongming Shan, “Corediff: Contextual error-modulated generalized diffusion model for low-dose ct denoising and generalization,” IEEE Transactions on Medical Imaging (2023); Method 6, SDM proposed by Du et al., reference Wenchao Du, HuanHuan Cui, LinChao He, Hu Chen, Yi Zhang, and Hongyu Yang, “Structure-aware diffusion for low-dose ctimaging,” Physics in Medicine & Biology, vol. 69, no. 15, pp. 155008 (2024).

[0060] The datasets used for training and testing in this invention are the Mayo dataset, which is widely used in the field of CT image reconstruction. The datasets include 5% dose CT images and reconstruction results from various methods, as shown below. Figure 4 As shown. Method 1 is FBP, Method 2 is RED-CNN, Method 3 is WGAN-VGG, Method 4 is CNCL, Method 5 is CoreDiff, and Method 6 is SDM. Through analysis of... Figure 4 The analysis of the results shows that the method proposed in this invention achieves good results in both PSNR and SSIM image reconstruction evaluation metrics, and also achieves the best results in LPIPS, a metric that is more in line with human visual perception. This indicates that the method proposed in this invention generates reconstruction results that are more in line with human visual perception while maintaining image reconstruction quality.

[0061] This test case uses 5% ultra-low dose CT images for a practical test, in which... Figure 4 This is a comparison of test results for corresponding low-dose CT images. Based on... Figure 4The green and yellow arrows indicate that the method proposed in this invention possesses superior detail reconstruction and artifact elimination capabilities, thanks to the conditional diffusion and progressive texture perception strategies proposed in this invention. In summary, compared with recent low-dose CT image reconstruction methods, the method proposed in this invention not only achieves excellent results in quantitative evaluation metrics (PSNR and SSIM) but also outperforms other methods in visual perception. Furthermore, the model complexity of the proposed method is lower than other depth model-based methods, exhibiting a significant advantage in inference efficiency and enabling better application in medical practice. Therefore, this invention is an effective and efficient low-dose CT image reconstruction method.

[0062] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0063] Example 4 To verify the effectiveness and superiority of the reconstruction method proposed in Example 1, this test case selected six other existing reconstruction methods for comparative experiments. Each method was trained according to the experimental setup of this invention to ensure fairness in the comparison. The six comparative methods are: Method 1, FBP proposed by Ramachandran et al., reference Ramachandran, GN, & Lakshminarayanan, AVThree-dimensional reconstruction from radiographs and electron micrographs: Application of convolutions instead of Fouriertransforms. Proceedings of the National Academy of Sciences, 68(9), 2236–2240(1971); Method 2, FBPConvNet proposed by Jin et al., reference Jin, KH, McCann, MT, Froustey, E., Unser, M.: Deep convolutional neural network for inverse problems in imaging. IEEE transactions on image processing 26(9), 4509–4522(2017); Method 3, DuDoTrans proposed by Wang et al., reference: Wang, C., Shang, K., Zhang, H., Li, Q., Zhou, SK: Dudotrans: Dual-domain transformer for sparse-view ctreconstruction. In: International Workshop on Machine Learning for MedicalImage Reconstruction. pp. 84–94. Springer (2022); Method 4, LEARN proposed by Chen et al., reference Chen, H., Zhang, Y., Chen, Y., Zhang, J., Zhang, W., Sun, H., Lv, Y., Liao, P., Zhou, J., Wang, G.: Learn: Learned experts' assessment-based reconstruction network for sparse-data ct. IEEE transactions on medical imaging 37(6), 1333–1347 (2018); Method 5, RegFormer proposed by Xia et al., see reference Xia, W., Yang, Z., Lu, Z., Wang, Z., Zhang, Y.: Regformer: A local-nonlocal regularization-based model for sparse-view ct reconstruction. IEEE Transactions on Radiation and Plasma Medical Sciences (2023); Method 6, DDDM proposed by Yang et al., refer to Chun Yang, Dian Sheng, Bo Yang, Wenfeng Zheng, and Chao Liu, "A dual-domain diffusion model for sparse-view ctreconstruction," IEEE Signal Processing Letters (2024).

[0064] The datasets used for training and testing in this invention are both the Mayo dataset, which is widely used in the field of CT image reconstruction. Experimental results comparing the reconstruction results of the method of this invention with those of other low-dose CT methods are shown below. Figure 5 As shown. Method 1 is FBP, Method 2 is FBPConvNet, Method 3 is DuDoTrans, Method 4 is LEARN, Method 5 is RegFormer, and Method 6 is DDDM. The 32-view sparse-view CT images and the reconstruction results of each method are shown below. Figure 6 As shown, it can be done through Figure 6 It can be seen that this invention also achieves a balance between traditional image measurement indicators and indicators that conform to human visual perception in the field of sparse view CT image reconstruction.

[0065] This embodiment uses 32-view sparse-view CT images, and the present invention achieves significant improvements in several aspects compared to the original LEARN model. The original LEARN achieved higher PSNR and SSIM metrics compared to RegFormer, and also had a lower LPIPS value using only one sampling step. Furthermore, Figure 7 The visualization results are presented. Most methods result in severe oversmoothing in the reconstructed images, while DDDM preserves more texture details. However, undesirable artifacts and inconsistent textures also lead to lower PSNR and SSIM scores. In contrast, this invention achieves high-fidelity image reconstruction by increasing the sampling steps to eight. Similarly, this invention is an effective and efficient method for sparse-view CT image reconstruction.

[0066] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A CT image reconstruction method based on a progressive texture-aware diffusion model, characterized in that, Includes the following steps: The low-dose CT image to be reconstructed, the sparse-view CT image, or the sparsely sampled sinusoidal data is input into the trained progressive texture-aware diffusion model, and the reconstructed CT image is output. The progressive texture-aware diffusion model performs preliminary reconstruction on the low-dose CT image, sparse-view CT image, or sparsely sampled sinusoidal data to be reconstructed, obtaining the low-frequency structure of the image. Based on the low-frequency structure, high-frequency details are generated iteratively through a conditional Schrödinger bridge diffusion network, multi-scale high-frequency features are extracted from the high-frequency details, and a reconstructed CT image is generated based on the low-frequency structure of the image and the multi-scale high-frequency features.

2. The CT image reconstruction method based on a progressive texture-aware diffusion model as described in claim 1, characterized in that, The networks used in the initial reconstruction include CNN, ResNet, VGG, DenseNet, EfficientNet, RED-CNN, or LEARN.

3. The CT image reconstruction method based on a progressive texture-aware diffusion model as described in claim 2, characterized in that, Extracting multi-scale high-frequency features from the high-frequency details specifically includes: The high-frequency details are downsampled, and the sampling results are input into a lightweight super-resolution U-shaped network to predict the high-frequency residual features at each scale. The texture features output from the intermediate layer of the lightweight super-resolution U-shaped network are aligned by convolution and then output to the corresponding layer of the encoder or decoder of the lightweight super-resolution U-shaped network.

4. The CT image reconstruction method based on a progressive texture-aware diffusion model as described in claim 3, characterized in that, The lightweight super-resolution U-shaped network has a structure of 3 layers of 3×3 convolutions, uses the ReLU activation function, and the output of the last layer is a high-frequency residual with the same dimension as the downsampled image.

5. The CT image reconstruction method based on a progressive texture-aware diffusion model as described in claim 3, characterized in that, The lightweight super-resolution U-shaped network uses downsampled MSE to calculate the guiding loss.

6. The CT image reconstruction method based on a progressive texture-aware diffusion model as described in any one of claims 1-5, characterized in that, The progressive texture-aware diffusion model is trained in a supervised manner, using the MSE loss function.

7. The CT image reconstruction method based on a progressive texture-aware diffusion model as described in claim 6, characterized in that, The progressive texture-aware diffusion model was trained using training and validation datasets. The training dataset includes a low-dose CT training dataset and a sparse-view CT training dataset; the validation dataset includes a low-dose CT validation dataset and a sparse-view CT validation dataset. The low-dose CT training dataset includes multiple sets of low-dose and normal-dose CT images; the sparse-view CT training dataset includes sparse-view CT images, sparsely sampled sine wave data, and full-view sampled CT images. The low-dose CT validation dataset includes multiple sets of low-dose CT images; the sparse-view CT validation dataset includes sparsely sampled sine wave data and full-view sampled CT images.

Images