Method and apparatus for limited ct image reconstruction based on denoising diffusion model

Abstract

The application discloses a restricted CT image reconstruction method and device based on a denoising diffusion model. The method comprises the following steps: preparing a data set, obtaining an original matched restricted CT image and a clean CT image; constructing a denoising diffusion model; adding Gaussian noise to the original matched image pair in a training process stage to generate a noise target image; and inversely inferring sampling based on the target noise image and the original restricted CT image to obtain a clean image of the original restricted CT image. The denoising diffusion model is applied to the restricted CT image reconstruction, and the denoising diffusion model has good original image structure reservation capability and better reconstruction performance in detail recovery, so that the reconstructed restricted CT image has relatively high subjective and objective quality evaluation.

Description

Technical Field

[0001] This invention belongs to the field of image reconstruction, and in particular relates to a restricted CT image reconstruction method based on a denoising diffusion model. Background Technology

[0002] Computed tomography (CT) is an accurate and non-invasive method for examining internal abnormalities such as tumors and fractures. It is currently widely used by clinicians to diagnose and detect abnormalities in internal organs. However, excessive radiation doses can easily induce metabolic abnormalities and diseases such as cancer, and traditional CT scans are not suitable for sensitive groups such as pregnant women. Issues related to radiation doses have attracted significant public attention.

[0003] Reducing operating current or voltage, or imaging at sparse angles, are common methods to reduce patient radiation exposure. This is referred to as restricted CT imaging. Under restricted conditions, image quality deteriorates, the contrast between the lesion area and the background decreases, severely impacting clinicians' judgment of diseased tissue. Furthermore, the CT imaging mechanism differs from that of ordinary cameras and other equipment, and cannot be simply viewed as a common image pattern recognition problem. Over the past few decades, many methods have been developed to improve the quality of restricted CT images, broadly categorized into (projection domain), (image domain), and (dual-domain) methods. Projection domain methods focus on the original projection data, using filtering to remove artifacts and noise in the projection domain. Iterative reconstruction is also a classic approach. Image domain methods directly post-process the damaged image. Compared to projection domain methods, they do not rely on the vendor's original data acquisition, making data acquisition and integration more convenient. In addition, dual-domain fusion methods combining the projection and image domains have achieved good results.

[0004] However, most of the above methods are discussed under specific constraints, such as reconstruction models specifically designed for sparse angle constraints and low-dose conditions. Furthermore, unlike natural images, medical images place higher demands on the reconstruction process; preserving the original image's structural information and details while removing artifacts and noise is crucial. Current methods are prone to over-smoothing and significant information loss. Therefore, improving the quality of CT reconstructions under different constraints and enhancing detail representation are urgent problems that need to be solved, and existing technologies require further improvement and development. Summary of the Invention

[0005] Based on the above background, it is necessary to provide a CT reconstruction method under constraints to address the aforementioned technical problems. This method can improve the quality of CT reconstruction under different constraints such as low dose and sparse angle, alleviate the problem of loss of structural information in reconstructed images, and better preserve image details.

[0006] To achieve the above objectives, the present invention provides a restricted CT image reconstruction method based on a denoising diffusion model, the restricted CT image reconstruction method based on the denoising diffusion model comprising:

[0007] Step 1, Dataset preparation: Obtain the original matched restricted CT images and clean CT images;

[0008] Step 2: Construct a denoising diffusion model, which includes a forward diffusion structure and a backward inference denoising structure. The forward diffusion structure obtains the target noise image by gradually adding Gaussian noise, and the backward inference denoising structure uses the DiffUnet noise prediction network.

[0009] The Diff UNet noise prediction network includes an encoder and a decoder. The encoder includes multiple downsampling modules (Diff Down) and the decoder includes multiple upsampling modules (Diff Up). The upsampling modules and downsampling modules are connected by a jump connection structure.

[0010] Step 3: During the training process, Gaussian noise is gradually added to the original matched image pairs to generate noisy target images;

[0011] Step 4: Based on the target noisy image and the original restricted CT image, reverse inference sampling is performed to obtain a clean image of the original restricted CT image.

[0012] Further, in step 1, the original clean image slices are obtained and their size is adjusted; under low-dose limiting conditions, the original data is projected and Poisson noise is added to obtain a matching CT image under low-dose limiting conditions; the CT image under sparse angle limiting conditions is simulated by reducing the sampling angle within the scanning range; at this point, the datasets under two different limiting conditions are prepared.

[0013] Furthermore, in step 2, the input to the forward diffusion structure is an image pair, namely, a matched low-dose CT image / sparse angle CT image and a clean CT image. The low-dose image / sparse angle image is referred to as LR, and the normal-dose image is referred to as HR. The forward diffusion process involves progressively adding noise to the image, which is accomplished through a Markov chain. The reverse reasoning denoising process starts from step T and progressively reasons to denoise, with the input being the pure noise image y generated by the forward diffusion process. t Compared to low-quality images, LR images are used as conditionally guided models for generation. The model iteratively removes noise to obtain y. t-1 y t-2 The entire process transforms the standard normal distribution into an empirical distribution, and finally restores the reconstructed CT image; where t is the number of forward diffusions.

[0014] Furthermore, in the original matched image pairs, the variance of the Gaussian noise added at each step of the forward diffusion structure is β.t , where t is a natural number representing the number of times the current forward process diffuses, to obtain the image probability distribution at step t;

[0015] The image at step t is obtained by adding Gaussian noise to the image at step t-1, and so on; the probability distribution expression for the image at step t is:

[0016]

[0017] Where β t Let y be the variance of the Gaussian noise added for the t-th time, where N represents a Gaussian distribution. t For the image at time t, y t-1 For the image at time t-1, q(y) t |y t-1 ) is a known y t-1 In the case of y t Probability distribution;

[0018] By deriving the properties of the Gaussian function, we can conclude that any image at step t can be directly calculated from the original image, as shown in the following expression:

[0019]

[0020]

[0021] Where a t =1-β t , For a t The average; ∈ ~N(0,1), y0 is the original image, y t Image with t-th order Gaussian noise added.

[0022] Furthermore, β t The value of β is obtained through a function whose derivative gradually increases. Specifically, the noise growth rate is slow in the early stage and accelerates in the later stage. d The implementation is achieved through the following expression:

[0023] f d =e xcosx

[0024] Here, x is the independent variable, and the domain range is set through hyperparameters.

[0025] Furthermore, the reverse reasoning denoising structure calculates the t-1 step image from the t-step image using the following formula, and iterates repeatedly to finally obtain the reconstructed clean image:

[0026]

[0027] in For the noise predicted by the Diff UNet network, σ t Let a represent the t-step variance, z ~ N(0,1), and a t =1-β t ,β t Let x be the variance of the Gaussian noise added for the t-th time, and let x be the image under the constraints.

[0028] Furthermore, the downsampling module consists of a normalization function, a swish activation function, a diffusion block, a grouped convolution, and a CrossXY downsampling module; the upsampling module consists of a normalization function, a swish activation function, a diffusion block, and a grouped convolution.

[0029] The input to Diff Down is the noisy image y at the current t-step. t Given an image x under constraints, the image x is processed by a normalization function and a swish activation function; then, a diffusion block iteratively processes the feature information between channels, enabling feature interaction between different order spaces; and the noisy image y is... t The feature maps processed by the Diffusion block are concatenated along the channel dimension; non-linear factors are introduced through normalization and swish activation functions; the feature maps are then processed through grouped convolution; the original input is added to the grouped feature maps, and the advantages of skip connections are utilized to enhance image features; finally, the feature maps are obtained through the CrossXY downsampling module.

[0030] In Diff Up, the feature map obtained from the encoder is processed by a normalization function and a swish activation function; then, the Diffusion block iteratively processes the feature information between channels, enabling feature interaction between different order spaces; and the noisy image y is processed... t The feature maps processed by the Diffusion block are concatenated along the channel dimension; then, non-linear factors are introduced through normalization and swish activation functions; finally, the feature maps are processed through grouped convolution; the original input is added to the feature maps obtained by grouped convolution, and the advantages of skip connections are utilized to enhance image features; finally, upsampling is completed through convolutional layers.

[0031] Furthermore, the Diffusion block structure is as follows: the feature map channels are processed iteratively through splitting-interaction-aggregation diffusion. The input feature channel group X is divided into two parts, denoted as A and B. 1 / 4 of the channels of A and 1 / 4 of the channels of B are connected and then processed by convolution to obtain the fused feature channel Z1. The remaining part is denoted as Id. 1 / 4 of the channels of Id is connected to Z1 and the features are fused by 1x1 convolution to obtain Z2. Id is processed by convolution to obtain Z3. Finally, the channels of Z2 and Z3 are connected to obtain the feature map that incorporates the channel dimension information.

[0032] Furthermore, the CrossXY downsampling module extracts image features from the X and Y spatial dimensions to obtain feature maps at different locations, and finally fuses the X and Y direction feature maps to aggregate spatial features. The module input is the feature map after Diff Down, activation function, regularization, and group convolution processing, which is defined as Z. Spatial dimension partitioning is performed by reducing Z in the X and Y dimensions through a flipping operation to obtain Z0. x With Z y After obtaining paired feature groups of different dimensions by relying on simple spatial transformations, Z... x With Z y Z is obtained by dividing the spatial features along different directions. x1 Z x2 Z y1 Z y2 This achieves feature fusion and segmentation from different orientations. To fully utilize the expressive features in the X and Y directions, the two sets of spatial feature maps are fused to enhance feature information expression. Then, convolution is used to complete information interaction to obtain Out1 and Out2; Out1 = F(Add(Z x1 Z y1 ), Out2 = F(Add(Z x2 Z y2 ), where Add is the feature map addition and fusion operation, and F represents the convolution operation; finally, the two feature maps Out1 and Out2 are fused to obtain the final downsampled feature map.

[0033] The present invention provides a terminal device, wherein the terminal includes a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, complete the restricted CT image reconstruction method based on the denoising diffusion model.

[0034] This invention discloses a method for reconstructing restricted CT images based on a denoising diffusion model. The method includes dataset preparation, obtaining original matched restricted CT images and clean CT images; constructing a denoising diffusion model; progressively adding Gaussian noise to the original matched image pairs during the training phase to generate a noisy target image; and inversely inferring and sampling based on the target noisy image and the original restricted CT image to obtain a clean image of the original restricted CT image. This invention uses a denoising diffusion model for restricted CT image reconstruction. Because the denoising diffusion model has a good ability to preserve the original image structure and has better reconstruction performance in detail recovery, the reconstructed restricted CT image has a high quality evaluation both subjectively and objectively.

[0035] This invention uses a denoising diffusion model for restricted CT image reconstruction, which exhibits good performance, robustness, and wide applicability under different constraints.

[0036] It effectively removes artifacts and noise, and the reconstructed image has good visual effects.

[0037] It has a good ability to preserve the original image structure and better reconstruction performance in detail restoration.

[0038] The computation time is shorter, and the efficiency is greatly improved. The reconstructed CT images have high quality ratings both subjectively and objectively. Attached Figure Description

[0039] Figure 1 This is a flowchart of an embodiment of the present invention.

[0040] Figure 2 This is a schematic diagram of the noise reduction diffusion model structure in an embodiment of the present invention.

[0041] Figure 3 This is a diagram of the Diff UNet noise prediction network structure in an embodiment of the present invention.

[0042] Figure 4 The above diagram shows the results of the upsampling module and the downsampling module in an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0044] like Figure 1 As shown, the restricted CT image reconstruction method based on a denoising diffusion model provided in this embodiment of the invention includes the following steps:

[0045] Acquire CT matching data pairs under constraints, including clean images and CT images with noisy artifacts under constraints, and construct a denoising diffusion model.

[0046] Specifically, the removal of noise artifacts in medical image reconstruction differs significantly from that in natural images, as crucial information from the original image is not lost during reconstruction. Traditional or deep learning methods can effectively remove artifacts from low-dose or sparse-angle images, but they may also cause damage to the original image information and excessive smoothing. To overcome these shortcomings, medical image reconstruction must address these ill-posed inversion problems. The diffusion model defines a forward diffusion stage, gradually introducing Gaussian noise to scramble the original image distribution, ultimately transforming the image into a standard Gaussian normal distribution. The reverse process then progressively removes noise to generate samples.

[0047] The dataset preparation phase specifically includes: acquiring original clean image slices and adjusting their size to 256x256; obtaining matching CT images under low-dose constraints by projecting the original data and adding Poisson noise; simulating CT images under sparse angle constraints by reducing the sampling angle within the scanning range; thus, the preparation of two datasets under different constraints is completed.

[0048] The entire network structure of the denoising diffusion model consists of a forward diffusion structure and a reverse denoising structure, such as... Figure 2 As shown, the forward diffusion structure obtains the target noise image by progressively adding Gaussian noise, while the inverse denoising structure uses Diff UNet. The input to the forward diffusion structure is an image pair, namely a matched low-dose image / sparse angle image and a normal image. Here, we refer to the low-dose image / sparse angle image as LR and the normal-dose image as HR. The forward process involves progressively adding noise to the image, accomplished through a Markov chain. The inverse denoising process starts from step T, p(y t-1 |y t x) represents the noise in the known image y t Inference y in the case of a low-quality original input image x t-1 The distribution of the data allows for stepwise inference denoising. The model input is the pure noise image y generated during the forward process. t Compared to low-quality images, LR images are used as conditionally guided models for generation. The model iteratively removes noise to obtain y. t-1 ,y t-2 The entire process transforms the standard normal distribution into an empirical distribution, ultimately restoring the reconstructed CT image. The variance of the Gaussian noise added at each step is β. t , where t is a natural number representing the number of times the current forward process diffuses, to obtain the image probability distribution at step t;

[0049] The image at step t is obtained by adding Gaussian noise to the image at step t-1, and so on; the probability distribution expression for the image at step t is:

[0050]

[0051] Where β t Let y be the variance of the Gaussian noise added for the t-th time, where N represents a Gaussian distribution. t For the image at time t, y t-1 For the image at time t-1, q(y) t |y t-1 ) is a known y t-1 In the case of y t Probability distribution; β is obtained through a function whose derivative gradually increases. Specifically, the noise growth rate is slow in the early stage and accelerates in the later stage. The value of β is obtained through the function f. d The implementation is achieved through the following expression:

[0052] f d =e xcosx

[0053] x is the independent variable, and its domain is set through hyperparameters. Based on the properties of the Gaussian function, it can be deduced that any image at step t can be directly calculated from the original image, as shown in the following expression:

[0054]

[0055]

[0056] Among them, a t =1-β t ,∈~N(0,1); Therefore, during the forward diffusion process of the model, the CT image under given constraints is stitched together with the matched clean image and Gaussian noise is added to the input iteratively to obtain the noisy image.

[0057] The original image is backsampled based on the target noise to obtain the target clean image of the original image.

[0058] Specifically, the t-step image is calculated based on the formula to obtain the t-step image, and the process is repeated iteratively to finally obtain the reconstructed clean image:

[0059] The formula is:

[0060] in For the noise predicted by the Diff UNet network, σ t Let z represent the t-step variance, and z ~ N(0,1). For example... Figure 3As shown, the Diff UNet noise prediction network includes an encoder and a decoder. The encoder consists of five downsampling modules, and the decoder consists of five upsampling modules. The Skip connection is a skip connection structure that connects the upsampling modules and the downsampling modules. Figure 4 The diagram illustrates the specific components of the upsampling and downsampling modules. The downsampling module consists of a normalization function, a swish activation function, a diffusion block, grouped convolutions, and a CrossXY downsampling module. The upsampling module consists of a normalization function, a swish activation function, a diffusion block, and grouped convolutions. Figure 4 In this context, Identify is a skip connection structure.

[0061] In the encoder structure, the input to Diff Down is the noise image y at the current t-step. t Given an image x under constraints, the model's feature representation ability is improved by applying a normalization function and a swish activation function to x. Then, a diffusion block iteratively processes the feature information between channels, enabling feature interaction between different order spaces. Finally, the noisy image y... t The feature maps processed by the Diffusion block are concatenated along the channel dimension; nonlinear factors are introduced through standardization and swish activation functions to improve the model's expressive power; the feature maps are then processed through grouped convolution with 16 groups; the original input is added to the feature maps obtained in this step, and the advantages of skip connections in the Identify operation are utilized to enhance image features; finally, the feature maps are obtained through the CrossXY downsampling module.

[0062] In the decoder structure, the DiffUp block processes the feature maps obtained from the encoder using a normalization function and a swish activation function to improve the model's feature representation capability; then, the Diffusion block iteratively processes the feature information between channels, enabling feature interaction between different order spaces; and finally, it processes the noisy image y... t The feature maps processed by the Diffusion block are concatenated along the channel dimension. Then, normalization and swish activation functions are used to introduce non-linearity, improving the model's expressive power. Finally, the feature maps are processed through grouped convolutions with 16 groups. The original input is added to the feature maps obtained in this step, leveraging the advantages of skip connections in the Identify operation to enhance image features. Finally, upsampling is completed through a 3x3 convolutional layer.

[0063] Diffusion block structure as follows Figure 4As shown, the feature map channels are processed through a diffusion iteration of "split-interaction-aggregation". The input feature channel group X is divided into two parts, denoted as A and B. 1 / 4 of the channels of A and 1 / 4 of the channels of B are concatenated and then processed by convolution to obtain the fused feature channel Z1. The remaining part is denoted as Id. 1 / 4 of the channels of Id is concatenated with Z1 and fused by 1x1 convolution to obtain Z2. Id is then processed by convolution to obtain Z3. Finally, the channels of Z2 and Z3 are concatenated to obtain a feature map incorporating channel dimension information.

[0064] The CrossXY downsampling module extracts image features from the X and Y spatial dimensions to obtain feature maps at different locations. Finally, it fuses the X and Y feature maps to aggregate spatial features, enriching the detail representation and complementing the details across different dimensions. The model structure can be divided into two parts: spatial dimension partitioning and dimensional information interaction. The module input is the feature map after Diff Down, activation function, regularization, and grouped convolution, defined as Z. Spatial dimension partitioning reduces Z along the X and Y dimensions through a flipping operation. x With Z y After obtaining paired feature groups of different dimensions through simple spatial transformations, Z... x With Z y Z is obtained by dividing the spatial features along different directions. x1 Z x2 Z y1 Z y2 This achieves feature fusion and segmentation across different orientations. To fully utilize the expressive features in the X and Y directions, the two sets of spatial feature maps are fused to enhance feature information expression, and then information interaction is completed through 1x1 convolution to obtain Out1 and Out2. Out1 = F(Add(Z x1 Z y1 ), Out2 = F(Add(Z x2 Z y2 Here, Add represents the feature map addition and fusion operation, and F represents a 1x1 convolution operation. Finally, the two feature maps Out1 and Out2 are fused to obtain the final downsampled feature map. The CrossXY downsampling module compensates for the destruction of image information structure caused by downsampling during image reconstruction. The complementary X and Y spatial dimension features enable the image reconstruction process to better preserve image details.

[0065] Furthermore, based on the above-mentioned restricted CT image reconstruction method based on the denoising diffusion model, this embodiment of the invention also provides a terminal, including a memory and a processor;

[0066] In some embodiments, the memory may be an internal storage unit of the terminal, such as a hard disk or RAM. In other embodiments, the memory may be an external storage device of the terminal, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc., equipped on the terminal. Furthermore, the memory may include both internal storage units and external storage devices of the terminal.

[0067] In some embodiments, the processor may be a central processing unit, a microprocessor, or other data processing chip, used to run program code stored in the memory or process data, such as executing the restricted CT image reconstruction method based on the denoising diffusion model.

[0068] To verify the effectiveness of this invention, 2165 images were randomly selected from the Deep Lesion dataset for training, and 342 images were retained as the validation set. To simulate low-dose images, Poisson noise was added to the projection domain. Furthermore, using the original Deep Lesion data as the ground truth, a parallel beam configuration acquisition scheme was simulated to obtain sparse-view CT images. The scan range was set to [0°, 180°], and 90 sampling angles were acquired to simulate sparse-view CT images.

[0069] Table 1 presents the comparison results between the original UNet method and the Diff UNet model of this invention under low-dose conditions. The model performance is quantitatively evaluated using two common metrics: Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). Higher values ​​indicate better image quality. Table 1 shows that under low-dose constraints, Diff UNet achieves higher accuracy than the UNet model in both metrics.

[0070] Table 1 Comparison of PSNR and SSIM under low-dose conditions

[0071]

[0072] Table 1 presents the comparison results between the original UNet and the Diff UNet method of this invention under low-dose conditions. Under sparse angle constraints, the Diff UNet shows a more significant performance improvement, achieving the best results in both metrics. The Diff UNet method of this invention can adapt to different constraints and has good robustness.

[0073] Table 2 Comparison of PSNR and SSIM under sparse angle conditions

[0074]

[0075] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A limited CT image reconstruction method based on a de-noising diffusion model, characterized in that, The method comprises the following steps: Step 1, data set preparation, obtaining original matched restricted CT images and clean CT images; Step 2, constructing a denoising diffusion model, the denoising diffusion model comprising a forward diffusion structure and a reverse reasoning denoising structure, the forward diffusion structure obtaining a target noise image by adding Gaussian noise step by step, and the reverse reasoning denoising structure adopting a DiffUNet noise prediction network; The DiffUNet noise prediction network comprises an encoder and a decoder, the encoder comprising a plurality of down-sampling modules DiffDown, and the decoder comprising a plurality of up-sampling modules DiffUp, the up-sampling modules and the down-sampling modules being connected through a jump connection structure; The down-sampling module is composed of a normalization function, a swish activation function, a Diffusion block, a grouped convolution and a CrossXY down-sampling module; the up-sampling module is composed of a normalization function, a swish activation function, a Diffusion block and a grouped convolution; The Diffusion block structure is: by splitting-interacting-aggregating diffusion iterative processing of a feature map channel, an input feature channel group X is divided into two parts, denoted as A and B, 1 / 4 channels of A are connected with 1 / 4 channels of B, and then a convolution processing is performed to obtain a fused feature channel Z1, and the remaining part is denoted as Id; 1 / 4 channels of Id are connected with Z1, a 1x1 convolution is performed to fuse the features and obtain Z2, Id is processed by convolution to obtain Z3, and finally the channels of Z2 and Z3 are connected to obtain a feature map with fused channel dimension information; Step 3, in the training process, Gaussian noise is added to the original matched image pair step by step to generate a target noise image; Step 4, based on the target noise image and the original restricted CT image, a clean image of the original restricted CT image is obtained by reverse reasoning sampling.

2. The limited view CT image reconstruction method based on the de-noising diffusion model according to claim 1, characterized in that: In step 1, the original clean image slice is obtained, the size is adjusted, and under the condition of low dose restriction, a matched CT image under the condition of low dose restriction is obtained by projecting the original data and adding Poisson noise; a CT image under the condition of sparse angle restriction is simulated by reducing the sampling angle in the scanning range; thus, the data set preparation under two different restriction conditions is completed.

3. The limited view CT image reconstruction method based on a de-noising diffusion model according to claim 1, wherein: In step 2, the forward diffusion structure input is a pair of images, i.e. a matched low-dose CT image / sparse-angle CT image and a clean CT image, the low-dose image / sparse-angle image is referred to as LR, the normal-dose image is referred to as HR, the forward diffusion process is to add noise to the image step by step, which is completed through a Markov chain, and the reverse inference denoising process starts from T steps, and the input is a pure noise image generated by the forward diffusion process The low-quality image LR is guided to generate the model as a condition, and the model iteratively removes noise step by step to obtain 、 The whole process converts the standard normal distribution into an empirical distribution, and finally restores the reconstructed CT image; wherein t is the number of forward diffusion processes, is the image of the t-1th time, is the image of the t-2th time.

4. The limited view CT image reconstruction method based on a de-noising diffusion model according to claim 1, wherein: In the original matching image pair, the variance of Gaussian noise added in each step of the forward diffusion structure is where t is a natural number, representing the number of current forward process diffusion, and the image probability distribution of t steps is obtained. The image of step t is obtained by adding Gaussian noise to the image of step t-1, and the same is true for the subsequent steps; the probability distribution expression of the image of step t is: ； wherein is the variance of the added Gaussian noise for the tth addition, represents a Gaussian distribution, is the image for the tth addition, is the image for the t-1th addition, is known in the case that the probability distribution; According to the derivation of the Gaussian function, it can be obtained that any t-step image can be directly calculated from the original image, and the expression is as follows: wherein , is the average number; , is the original image, is the image to which t times Gaussian noise is added.

5. The limited view CT image reconstruction method based on a denoising diffusion model according to claim 4, characterized in that: obtained by a function with gradually increasing derivative, specifically, the noise growth speed is slow in the early stage and the noise growth speed is fast in the later stage, the value is obtained by a function is realized, and the expression is: wherein, is the argument, with the domain range defined by the hyperparameter settings.

6. The limited view CT image reconstruction method based on a de-noising diffusion model according to claim 4, wherein: The reverse reasoning denoising structure calculates the image of step t-1 from the image of step t by the following formula, and iteratively obtains the reconstructed clean image: wherein is the noise predicted by the Diff UNet network, represents the variance at step t, , , is the variance of the added Gaussian noise at the t-th step, is the image under the constraint.

7. The limited view CT image reconstruction method based on a de-noising diffusion model according to claim 1, wherein: The input of Diff Down is the noise image of the current t step With the image under the restriction condition With the image under the restriction condition After processing by the standardization function and the swish activation function; then the feature information between the channels is processed by the Diffusion block for iteration, so that the features between different order spaces are interacted; The noise image is obtained by adding the noise image and the image The feature map processed by the Diffusion block is spliced in the channel dimension; a nonlinear factor is introduced by processing through a standardization and a swish activation function; the feature map is processed through grouping convolution; the original input is added to the feature map obtained by grouping, the advantages of the jump connection are used, and the image features are enhanced; finally, the feature map is obtained through a CrossXY downsampling module; In DiffUp, the feature map obtained by the encoder is processed by a normalization function and a swish activation function; then the feature information between channels is iteratively processed by a Diffusion block, so that the features in different order spaces are interacted. The noise image is obtained by adding the noise image to the original image. The feature map processed by the Diffusion block is spliced in the channel dimension; a nonlinear factor is introduced by processing through standardization and a swish activation function; finally, the feature map is processed through grouped convolution; the original input is added to the feature map obtained by grouped convolution, the advantages of jump connection are used to enhance the image features, and finally the upsampling is completed through the convolution layer.

8. The limited view CT image reconstruction method based on a denoising diffusion model according to claim 7, characterized in that: CrossXY down-sampling module extracts image features from two spatial dimensions X and Y to obtain feature maps of different positions, and finally fuses X and Y direction feature maps to aggregate spatial features; the module input is the feature map after Diff Down, activation function, regularization and grouped convolution processing, which is defined as Z; the spatial dimension division reduces X and Y dimensions through folding operation to obtain and , and then and are obtained by dividing the spatial features in different directions , , , , which realizes the fusion and segmentation of different direction features. In order to fully utilize the expression features in X and Y directions, two groups of spatial feature maps are fused to enhance the expression of feature information, and then information interaction is completed through convolution to obtain and ; , , wherein is a feature map addition fusion operation, represents a convolution operation; finally, the two feature maps and are fused to obtain the final down-sampling feature map.

9. A terminal device, characterized by: The terminal comprises a memory and a processor, and computer instructions stored on the memory and running on the processor, when the computer instructions are run by the processor, the computer instructions complete the restricted CT image reconstruction method based on the de-noising diffusion model in any one of claims 1-8.

Images