Low dose ct image denoising method and system based on fast diffusion model

Abstract

The application discloses a kind of low-dose CT image denoising method and system based on fast diffusion model.There is including: scanning scheme design and experimental phantom preparation, obtain the CT image under different doses and matched normal dose CT image;Fast diffusion model is constructed;Training process phase gradually diffuses noise to the matched image pair, generates noise target image;Based on the target noise image and low-dose CT image is condition block processing, while removing noise gradually to obtain clear CT image by reverse inference.This application uses high-performance fast diffusion model to be applied to low-dose CT image reconstruction, by gradually diffusing and removing noise, model can capture the complex structure and details of image, while removing image noise and artifact in better degree, maximum degree reserves CT image original structure.In addition to this, the present application and the processing result of existing advanced method are given to radiology doctor blind evaluation processing, the method of the present application obtains optimal result.

Description

Technical Field

[0001] This invention relates to the field of medical CT image processing, specifically to a method and system for denoising low-dose CT images based on a high-performance fast diffusion model. Background Technology

[0002] CT scans are widely used in clinical practice as a non-invasive imaging modality for auxiliary diagnostic examinations. However, discussions regarding dosage issues have been ongoing, as radiation doses exceeding a certain threshold pose a risk of inducing diseases such as cancer. Effective dose control can reduce the risk of patient exposure to CT radiation and extend the lifespan of the X-ray tube. However, it can also lead to increased noise and artifacts, affecting doctors' diagnoses. Therefore, researching high-performance low-dose denoising and reconstruction methods for medical CT to improve diagnostic reliability is of great significance. However, due to its unique imaging mechanism, noise non-uniformity, and the higher demand for detail information in medical images, this is a challenging problem. Deep learning methods offer a novel denoising approach, relying on convolutional neural networks to learn non-linear mappings from low to high doses. These methods have achieved impressive results, effectively suppressing noise and artifacts. However, for medical CT image denoising tasks, there are higher requirements for preserving the original detail information of the denoised image. Excessive smoothing can lead to the loss of lesion information, resulting in undiagnostic problems or even misdiagnosis. The aforementioned methods still have shortcomings in preserving texture and original detail information.

[0003] Diffusion models are generative models that have emerged in recent years. By gradually diffusing noise, these models can capture the complex structure and details of an image, reconstructing a more realistic image. Diffusion models are widely used in the field of natural images, and compared to natural images, they have a natural advantage in CT denoising and reconstruction tasks; however, efficiency and time costs are also significant issues. Therefore, how to efficiently improve the denoising effect of low-dose CT and enhance detail representation is an urgent problem to be solved, and existing technologies still need improvement and development. Summary of the Invention

[0004] In view of the foregoing background, and with the goal of solving the aforementioned technical problems, this paper proposes a low-dose CT image denoising method based on a high-performance fast diffusion model to improve denoising efficiency and effect and alleviate the loss of important details of lesions in denoised images under different low-dose conditions.

[0005] To achieve the above objectives, the present invention relates to a low-dose CT image denoising method based on a high-performance fast diffusion model, which includes the following steps:

[0006] Step 1: Obtain CT images at different low-dose levels and normal-dose CT images;

[0007] Step 2: Build a fast diffusion model, including forward diffusion structure and reverse denoising inference;

[0008] Step 3: Input the matched low-dose and normal-dose images into the forward diffusion structure, and gradually add noise in the manner of a Markov chain to finally obtain the target noise image;

[0009] Step 4: Based on the target noisy image and the low-dose CT image, reverse denoising inference is performed using the LUNet model to gradually remove noise and obtain a clear CT image.

[0010] The LUNet model consists of an encoder, intermediate layers, and a decoder. The encoder includes a ResNet block, an efficient extended perceptual downsampling module (EEP), and a symmetric perceptual self-attention mechanism (SASelf-attention). The intermediate layers include convolutional layers and SASelf-attention. The encoder includes a ResNet block, SASelf-attention, and convolutional upsampling layers.

[0011] Furthermore, in step three, the variance of the Gaussian noise added at each step of the forward diffusion structure is β. t Where t represents the number of forward diffusion steps, 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:

[0012]

[0013] Where β t The variance of adding Gaussian noise for the t-th time, where N represents x. t The mean that follows the formula is The variance is β t Gaussian distribution, x t For the image at time t, x t-1 For the (t-1)th time graph, q(x) t |x t-1 (x) is a known value. t-1 In the case of x t Probability distribution; therefore, the image at any step t can be obtained from the initial x0, where T represents the total number of steps, expressed by the following formula:

[0014]

[0015] x 1:T Images representing steps 1 to T.

[0016] Furthermore, in step four, after obtaining the pure target noise image, the pure target noise is progressively denoising under the condition of inputting low-dose CT to obtain a clear CT image, specifically as follows:

[0017] The image at step t-1 is calculated from the image at step t using the following formula, and is represented as follows:

[0018]

[0019] Where ∈ θ Predicted using the LUNet model Representing low-dose CT images, a t =1-β t .

[0020] Furthermore, the specific processing procedure of the LUNet model is as follows;

[0021] The encoder's input is the noisy image x at the current t-step. t Images under low-dose conditions The ResNet blocks compute rich feature information, which is then passed through the EEP module to obtain smaller feature maps, thereby reducing spatial resolution. After passing through multiple ResNet blocks and the EEP module, attention feature maps are obtained through SA self-attention. Finally, intermediate layers learn higher-level feature representations x. mid The decoder's input is the feature x processed by the intermediate layer. mid After upsampling and SASelf-attention processing, the image details are restored through multiple ResNet blocks and upsampling convolutional layers. In addition, the feature maps of the encoder and the corresponding layers of the decoder are connected through a skip connection structure.

[0022] Furthermore, the ResNet block includes Norm normalization layers, Swish activation functions, and 3x3 convolutions.

[0023] Furthermore, the EEP module consists of two parts: an augmented perception module and a feature selection module. In the augmented perception module, feature transformation technology F is used, as specifically shown below;

[0024] F(X)=f i,j =X[i:H:2,j:W:2](i,j∈(0,1))

[0025] Where X[i:H:2,j:W:2] represents that the numerical values ​​in the H and W directions of the feature map X are spaced 2 apart, and i and j represent the initial values ​​of H and W;

[0026] Assume the input feature map X of the EEP module has a shape of H×W×C. in H and W are the height and width of the feature map, respectively, and C... in Representing the number of input channels, setting the downsampling scale to 2 allows X to be mapped to obtain four sets of downsampled feature maps f.00 ,f 01 ,f 10 ,f 11 Each set of features has a shape of (H / 2, W / 2, C). in By fusing the features that are furthest apart, a wider perceptual field of view can be obtained. 00 with f 11 Adding them together gives f up f 01 with f 10 Adding them together gives f down ;

[0027] After completing the perceptual augmentation, the feature selection module adaptively selects features. Specifically, it first aggregates global information using global pooling to obtain C. up With C down ;

[0028]

[0029] The global feature vector C up With C down The weight scores for different channels are obtained using the following formula:

[0030]

[0031] The adaptively optimized feature vectors α1 and α2 represent the original feature input weights. Under the guidance of the adaptive feature weights, the upper and lower features are fused to obtain the final downsampling result X. down ;

[0032] X down =α1·F up +α2·F down .

[0033] Furthermore, SA self-attention will input the feature image X r Q, K, and V are obtained through 1x1 convolution. A weight-sharing strategy is used in the calculation of K and Q, sharing a single 1x1 convolution to reduce computation. The channels are split to half their original size to obtain Q and K. Subsequently, the ReLU activation function is used to sparsify the K and Q features, yielding K. sp and Q sp Meanwhile, V is divided equally along the channel dimension to obtain V1 and V2; K sp Q sp VQ is obtained by cross-addition and reconstruction with V1 and V2 respectively. sp and VK sp This further enhances and highlights the information, improving the information flow exchange between channel groups; finally, the cross-reconstructed feature group VQ sp and VKsp The refined feature X is obtained by concatenating along the channel dimension. cross X cross With attention map M att After multiplication, a 1x1 convolution is performed to obtain the self-attention-enhanced feature X. att Finally, the original input features X r With X att X is obtained by fusing and adding points one by one. out .

[0034] Furthermore, the attention map M is obtained through the feature scaling module. att The specific processing procedure is as follows;

[0035] Perform feature transformation technique F on the two sets of features Q and K respectively, assuming that Q and K have the same size. D represents the number of channels for the Q and K features, and Q can be obtained through transformation techniques. 00 Q 01 Q 10 Q 11 ,K 00 ,K 01 ,K 10 ,K 11 ;

[0036] Enhance the interaction of Q and K information flows through interleaving and fusion: QK1 = Q 00 +K 11 QK2=Q 01 +K 10 QK3 = Q 10 +K 01 QK4 = Q 11 +K 00 ;

[0037] Finally, the feature P is obtained by concatenating along the channel dimension. The shape is H×W×2D, and Cat represents the connection along the channel dimension.

[0038] P = Cat(QK1,QK2,QK3,QK4)

[0039] A convolution is added after the P feature to enhance the feature information representation, resulting in the cross-scale intermediate feature P. att Then Q and K are respectively compared with P att Multiplying them yields the corresponding vertical and horizontal attention maps H. att With K att Finally, multiply the two to calculate the attention map M. att :

[0040]

[0041]

[0042] M att =W att ·H att .

[0043] This invention also provides a low-dose CT image denoising system based on a fast diffusion model, comprising the following units:

[0044] The image acquisition unit is used to acquire CT images of different low-dose levels and normal-dose CT images;

[0045] The fast diffusion model building unit is used to build fast diffusion models, including forward diffusion structure and reverse denoising inference;

[0046] The forward diffusion unit is used to input matched low-dose and normal-dose images into the forward diffusion structure and gradually add noise in a Markov chain manner to finally obtain the target noise image.

[0047] The inverse denoising inference unit is used to perform inverse denoising inference based on the target noisy image and the low-dose CT image, and gradually remove noise to obtain a clear CT image.

[0048] The LUNet model consists of an encoder, intermediate layers, and a decoder. The encoder includes a ResNet block, an efficient extended perceptual downsampling module (EEP), and a symmetric perceptual self-attention mechanism (SASelf-attention). The intermediate layers include convolutional layers and SASelf-attention. The encoder includes a ResNet block, SASelf-attention, and convolutional upsampling layers.

[0049] The present invention also provides a terminal device, comprising: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the low-dose CT image denoising method based on the fast diffusion model described in the above technical solution.

[0050] This invention discloses a low-dose CT image denoising method based on a high-performance fast diffusion model. First, a scanning scheme is designed and an experimental phantom is prepared to acquire CT images at different doses and matching normal-dose CT images. A fast diffusion model is then constructed. During the training phase, noise is progressively diffused onto the matched image pairs to generate a noisy target image. Based on the target noisy image and the low-dose CT image, the image is processed in blocks, and noise is progressively removed through reverse reasoning to obtain a clear CT image. This invention uses a high-performance fast diffusion model for low-dose CT image reconstruction. By progressively diffusing and removing noise, the model can capture the complex structure and details of the image, effectively removing image noise and artifacts while preserving the original structure of the CT image to the greatest extent. Furthermore, the results, processed with existing advanced methods, were subjected to blind review by radiologists, and the method described in this invention achieved optimal results. Attached Figure Description

[0051] Figure 1 This forms the overall framework of the method of the present invention;

[0052] Figure 2 This is a diagram showing the data of the phantom.

[0053] Figure 3 This is a structural diagram of the Efficient Extended Perception Downsample (EEP) in this invention;

[0054] Figure 4 This is a diagram of the Symmetric Aware Self-attention (SA self-attention) structure in this invention. Detailed Implementation

[0055] 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.

[0056] like Figure 1 As shown, the low-dose CT image denoising method based on a high-performance fast diffusion model includes the following steps:

[0057] Step 1: Design a CT scan scheme to obtain low-dose and normal-dose images of different locations and at different doses, obtain matching image domain CT data pairs, and use these data as the basis to build a high-performance rapid diffusion model.

[0058] During the scanning scheme design and phantom data preparation phase, a simulated head and chest phantom (GE, General Electric Company, USA) was used. The volume, size, and density of the brain, lungs, vertebrae, chest wall, mediastinum, and bones in the phantom are identical to the soft tissues and bones of the corresponding parts of the human body, thus accurately simulating the organs and tissues of the head and chest. During CT scans, the phantom has the same X-ray attenuation characteristics as the human body, allowing for the acquisition of CT values ​​for various organs similar to those in the human body. The head and chest phantoms are shown below. Figure 2 As shown:

[0059] Both the head and chest phantoms were scanned on a Revolution CT scanner (GE Healthcare). After locating the head and chest phantoms, standard dose images of the head were obtained at 120 kV and 300 mA. Standard dose images of the chest were obtained at 120 kV and 320 mA. Using the CTDI of these standard dose images as a standard, and keeping the tube voltage constant in principle, low-dose head and chest images with 25% and 10% CTDI were obtained by changing the tube current.

[0060] The head CT scan used axial scanning mode with a detector width of 40 mm, a scan time of 1 second, and a slice thickness of 0.625 mm. Low-dose phantom data were obtained by controlling the CTDI value. The CTDI value for the head phantom at the standard dose was set to 46.06, corresponding to 11.51 for a 25% dose and 4.61 for a 10% dose. Specifically, the tube voltage was kept constant while the tube current was adjusted to obtain images at different doses; all other acquisition parameters remained unchanged. A total of 512 phantom slices were obtained per scan, in standard 512x512 format. The images were stored in DICOM format.

[0061] The chest scan used a conventional helical scanning method with a detector width of 80 mm, a pitch of 1.375 mm, a scan time of 1 second, and a head scan slice thickness of 0.625 mm. The CTDI value of the head phantom at standard dose was set to 7.86, corresponding to a 25% dose CTDI of 1.96 and a 10% dose CTDI of 0.74. A total of 481 phantom slices were generated per scan, all in a standard 512x512 format. The storage format was DICOM.

[0062] Step 2: Build a high-performance fast diffusion model, including a forward diffusion structure and a reverse denoising inference process. The forward process takes matched low-dose and normal-dose images as input and adds noise progressively according to a Markov chain approach. The final result is a purely noisy image. The reverse inference process is performed sequentially, with each input consisting of a low-dose image and a noisy image. After inference through the LUNet model, the noise is gradually reduced until a clear image is obtained.

[0063] Step 3: Input the matched low-dose and normal-dose images into the forward diffusion structure, and gradually add noise in the manner of a Markov chain to finally obtain the target noise image;

[0064] The variance of the Gaussian noise added at each step of the forward diffusion structure is β. t , where t represents the number of forward diffusion steps. 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:

[0065]

[0066] Where β t The variance of adding Gaussian noise for the t-th time, where N represents x. t The mean that follows the formula is The variance is β t Gaussian distribution, x t For the image at time t, x t-1 For the (t-1)th time graph, q(x) t |x t-1 (x) is a known value. t-1 In the case of x t Probability distribution; therefore, the image at any step t can be obtained from the initial x0, where T represents the total number of steps, expressed by the following formula:

[0067]

[0068] x 1:T Images representing steps 1 to T.

[0069] Furthermore, in step four, after obtaining the pure noise, a clear CT image is obtained by progressively inferring and denoising the pure noise using a low-dose CT input as a condition. Specifically:

[0070] The image at step t-1, calculated from the image at step t based on the above formula, can be represented as follows:

[0071]

[0072] Where ∈ θ Predicted using the LUNet network Representing low-dose CT images, a t =1-β t .

[0073] Step 4: Based on the target noisy image and the low-dose CT image, reverse denoising inference is performed using the LUNet model to gradually remove noise and obtain a clear CT image.

[0074] The LUNet model consists of an encoder, intermediate layers, and a decoder. The encoder includes a ResNet block, an efficient extended perceptual downsampling module (EEP), and a symmetric perceptual self-attention mechanism (SASelf-attention). The intermediate layers include convolutional layers and SASelf-attention. The encoder includes a ResNet block, SASelf-attention, and convolutional upsampling layers.

[0075] The encoder's input is the noisy image x at the current t-step. t Images under low-dose conditions Rich feature information is computed by ResNet blocks; smaller feature maps are obtained through the Efficient Extended Perception Downsample (EEP) module, thus reducing spatial resolution; after passing through multiple ResNet blocks and EEP modules, attention feature maps are obtained through Symmetric Aware Self-attention (SA self-attention); then, a higher-level feature representation x is learned by the intermediate Midblock layer. mid The decoder's input is the feature x processed by the intermediate layer. mid After upsampling and SASElf-attention processing, the image details are restored through multiple ResNet blocks and upsampling convolutional layers. A single pass of LUNet denoises the input noisy image with the image under low-dose conditions to obtain a less noisy image; subsequent iterations of denoising yield a clearer image. LUNet introduces skip connections, linking feature maps from a specific layer of the encoder with the feature maps of the corresponding layer of the decoder. This allows the network to simultaneously focus on local details and global contextual information.

[0076] EEP such as structure Figure 3 As shown, it consists of two parts: an augmented perception module and a feature selection module. In the augmented perception part, we used feature transformation technique F, specifically:

[0077] F(X)=f i,j =X[i:H:2,j:W:2](i,j∈(0,1))

[0078] X[i:H:2,j:W:2] represents a value interval of 2 in the H and W directions, where i and j represent the initial values ​​of H and W.

[0079] Assume the EEP input feature map X has a shape of H×W×C. in H and W are the height and width of the feature map, respectively, and C... inThis represents the number of input channels. Setting the downsampling scale to 2 will map X to obtain four sets of downsampled feature maps f. 00 ,f 01 ,f 10 ,f 11 Each set of features has a shape of (H / 2, W / 2, C). in We fuse the features that are furthest apart to gain a wider perceptual field of view. 00 with f 11 Adding them together gives f up f 01 with f 10 Adding them together gives f down .

[0080] After perceptual augmentation is completed, the feature selection module adaptively selects features, preserving the feature information of the original channels while reducing computational costs. A simple and effective adaptive feature selection method is used to merge f. up with f down The characteristics of [the system / entity] are as follows. Specifically, the operation first aggregates global information using a global pooling method to obtain C. up With C down .

[0081]

[0082] The global feature vector C up With C down The weight scores for different channels are obtained using the following formula:

[0083]

[0084] The adaptively optimized feature vectors α1 and α2 represent the original feature input weights. Under the guidance of the adaptive feature weights, the upper and lower features are fused to obtain the final downsampling result X. down This is the output of the EEP module.

[0085] SA self-attention structure Figure 4 As shown, SA self-attention is added in the middle layer, the last layer between the encoder and decoder. SA self-attention will input the feature image X... r Q, K, and V are obtained through 1x1 convolution. For the calculation of K and Q, a weight-sharing strategy is used, sharing a single 1x1 convolution to reduce computation. The channels are split into half of their original values ​​to obtain Q and K. This reduces channel information redundancy. Subsequently, the K and Q features are sparsified using the ReLU activation function to obtain K. sp and Q sp Meanwhile, V is divided equally along the channel dimension to obtain V1 and V2.sp ,K sp VQ is obtained by cross-addition and reconstruction with V1 and V2 respectively. sp and VK sp This further enhances and highlights the information, improving the information flow exchange between channel groups. Finally, the cross-reconstructed feature group VQ sp and VK sp The refined feature X is obtained by concatenating along the channel dimension. cross X cross With attention map M att After multiplication, a 1x1 convolution is performed to obtain the self-attention-enhanced feature X. att Finally, the original input features X r With X att X is obtained by fusing and adding points one by one. out .

[0086] In the network configuration, SA self-attention is added to the last layer of the encoder. Since the feature maps exhibit cross-scale similarity, the size of Matt is halved (8x8) to improve computational efficiency. The feature scaling module is the Cross-Aware and Scale Block (CSPBlock), which references the design principles of EEP (Enhanced Feature Processing), performing feature transformation techniques F on the two sets of features Q and K respectively. Assume that the sizes of Q and K are both... D represents the number of channels for the Q and K features, from which Q can be obtained. 00 Q 01 Q 10 Q 11 ,K 00 ,K 01 ,K 10 ,K 11 .

[0087] QK1 = Q 00 +K 11 QK2=Q 01 +K 10 QK3 = Q 10 +K 01 QK4 = Q 11 +K 00

[0088] The interaction of Q and K information flows is enhanced by interleaving and fusion. Finally, the feature P is obtained by splicing in the channel dimension. The feature P has a shape of H×W×2D, and Cat represents the connection in the channel dimension.

[0089] P = Cat(QK1,QK2,QK3,QK4)

[0090] Adding a 1x1 convolution (stride=1) after the P feature enhances the feature information representation, resulting in the cross-scale intermediate feature P. att Medical images exhibit greater symmetry; therefore, we rely on only half of the features for global information utilization and modeling, and perform the modeling separately along the horizontal and vertical symmetry directions. Specifically, we correlate Q and K with P respectively. att Multiplying them yields the corresponding vertical and horizontal attention maps H. att With K att Finally, multiply the two to calculate the attention map M. att .

[0091]

[0092]

[0093] M att =W att ·H att

[0094] This invention also provides a low-dose CT image denoising system based on a fast diffusion model, comprising the following units:

[0095] The image acquisition unit is used to acquire CT images of different low-dose levels and normal-dose CT images;

[0096] The fast diffusion model building unit is used to build fast diffusion models, including forward diffusion structure and reverse denoising inference;

[0097] The forward diffusion unit is used to input matched low-dose and normal-dose images into the forward diffusion structure and gradually add noise in a Markov chain manner to finally obtain the target noise image.

[0098] The inverse denoising inference unit is used to perform inverse denoising inference based on the target noisy image and the low-dose CT image, and gradually remove noise to obtain a clear CT image.

[0099] The LUNet model consists of an encoder, intermediate layers, and a decoder. The encoder includes a ResNet block, an efficient extended perceptual downsampling module (EEP), and a symmetric perceptual self-attention mechanism (SA). The intermediate layers include convolutional layers and SA self-attention. The encoder includes a ResNet block, SA self-attention, and convolutional upsampling layers.

[0100] The present invention provides a method and system for denoising low-dose CT images based on a high-performance fast diffusion model, and also provides a corresponding terminal device including one or more processors.

[0101] A storage device is provided for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the low-dose CT image denoising method based on the fast diffusion model described above. In some embodiments, the storage device may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the storage device 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. Further, the storage device may include both internal and external storage units of the terminal.

[0102] 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 low-dose CT image denoising method based on a high-performance fast diffusion model.

[0103] To assess the complexity of the modules involved in the invention, the table below shows the FLOPs and number of parameters. Raw_Down represents the original downsampling method, and Self-attention represents the original self-attention mechanism. Our designed downsampling module does not increase the number of parameters or FLOPs, while achieving superior performance. Furthermore, the SA Self-attention mechanism involved in the invention has lower parameters and FLOPs than the original. The module designed in this invention effectively reduces FLOPs and the number of parameters.

[0104] Table 1 Comparison of Parameter Complexity of Invention Methods

[0105] Method Params(M) FLOPs(M) Raw_Down 0.5 151.13 EEP 0 0 Self-attention 0.62 16.91 SA self-attention 0.59 14.34

[0106] Ten sets of head and ten sets of chest image slices were randomly selected from the test set and blindly interpreted by radiologists. Each set of images included informed low-dose CT (LDCT) and normal-dose CT (NDCT) images. The denoising results of other methods were blindly evaluated, including RED-CNN, EDCNN, WGAN, and the method of this invention. Referring to the standard EU CT evaluation guidelines, CT image noise, structural clarity, artifacts, and diagnosticity were evaluated using a 5-point evaluation method (1 being the worst, 5 the best). Table 2 shows the blind evaluation results of the head phantom test data. The method of this invention has the highest diagnosticity and the highest preservation of image detail and structure. Although the REDCNN method has good denoising and artifact removal effects, its excessive blurring results in the worst diagnosticity. Table 3 shows the blind evaluation results of the chest phantom test data. In the chest test data, the results of this invention achieved the most balanced effect, with all indicators close to the standard values.

[0107] Table 2. Head Phantom Test Data and Doctor's Blind Assessment Results

[0108]

[0109] Table 3. Chest Phantom Test Data and Doctor's Blind Assessment Results

[0110]

[0111] 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 low-dose CT image denoising method based on a fast diffusion model, characterized in that, Includes the following steps: Step 1: Obtain CT images at different low-dose levels and normal-dose CT images; Step 2: Build a fast diffusion model, including forward diffusion structure and reverse denoising inference; Step 3: Input the matched low-dose and normal-dose images into the forward diffusion structure, and gradually add noise in the manner of a Markov chain to finally obtain the target noise image; Step 4: Based on the target noisy image and the low-dose CT image, reverse denoising inference is performed using the LUNet model to gradually remove noise and obtain a clear CT image. The LUNet model consists of an encoder, intermediate layers, and a decoder. The encoder includes a ResNet block, an efficient extended perceptual downsampling module (EEP), and a symmetric perceptual self-attention mechanism (SA). The intermediate layers include convolutional layers and SASelf-attention. The encoder includes a ResNet block, SA Self-attention, and convolutional upsampling layers. The EEP module consists of two parts: an augmented perception module and a feature selection module. The augmented perception module utilizes feature transformation techniques. The specific details are as follows; in, Represented in feature map X and The numerical values ​​in the direction are spaced 2. and represent and Initial value; Assuming the EEP module inputs a feature map Shape , , For the height and width of the feature map, This represents the number of input channels. Setting the downsampling scale to 2 will reduce the input channel size. The mapping yields four sets of downsampled feature maps f 00 ,f 01 ,f 10 ,f 11 Each set of feature shapes is ( , , ), f), f fuses the features that are furthest apart to obtain a larger perceptual field of view. 00 with f 11 Adding them together gives f up f 01 with f 10 Adding them together gives f down ; After completing the perceptual augmentation, the feature selection module adaptively selects features. Specifically, it first aggregates global information using global pooling to obtain C. up With C down ; global feature vectors and The weight scores for different channels are obtained using the following formula: Adaptive optimized feature vector and The weights represent the original feature input weights. Under the guidance of adaptive feature weights, the upper and lower features are fused to obtain the final downsampling result. ; SA self-attention will input feature image The corresponding query Q, key K, and value V are obtained through 1x1 convolution. A weight-sharing strategy is used in the calculation of K and Q, sharing a single 1x1 convolution to reduce computation. The channels are split to half their original size to obtain Q and K. Subsequently, the ReLU activation function is used to sparsify the features of K and Q to obtain K. sp and Q sp Meanwhile, V is divided equally along the channel dimension to obtain V1 and V2; K sp Q sp VQ is obtained by cross-addition and reconstruction with V1 and V2 respectively. sp and VK sp This further enhances and highlights the information, improving the information flow exchange between channel groups; finally, the cross-reconstructed feature group VQ sp and VK sp The refined feature X is obtained by concatenating along the channel dimension. cross X cross With attention map M att After multiplication, the features are processed by 1x1 convolution to obtain the self-attention-enhanced features. Finally, the original input features and The result is obtained by merging and adding points one by one. .

2. The low-dose CT image denoising method based on the fast diffusion model as described in claim 1, characterized in that: In step three, the variance of the Gaussian noise added at each step of the forward diffusion structure is... Where t represents the number of forward diffusion steps, 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: ； in Let the variance of adding Gaussian noise for the t-th time be... represent The mean that follows the formula is The variance is Gaussian distribution, For the image at time t, For the image of the (t-1)th iteration, Known In this case Probability distribution; therefore, the image at any t-step can be obtained from the initial... We obtain T, which represents the total number of steps, expressed by the following formula: Representing 1 to Image of the step.

3. The low-dose CT image denoising method based on the fast diffusion model as described in claim 2, characterized in that: In step four, after obtaining the pure target noise image, a clear CT image is obtained by progressively reasoning and denoising the pure target noise using low-dose CT as the input condition. Specifically: The image at step t-1 is calculated from the image at step t using the following formula, and is expressed as follows: in Predicted using the LUNet model Represents low-dose CT images, .

4. The low-dose CT image denoising method based on the fast diffusion model as described in claim 1, characterized in that: The specific processing steps of the LUNet model are as follows; The encoder's input is the noisy image at the current t-step. Images under low-dose conditions The ResNet blocks compute rich feature information, which is then processed by the EEP module to obtain smaller feature maps, thereby reducing spatial resolution. After passing through multiple ResNet blocks and the EEP module, attention feature maps are obtained through SA self-attention. Finally, intermediate layers learn higher-level feature representations. ; The input to the decoder is the features processed by the intermediate layer. After upsampling and SASelf-attention processing, the image details are restored through multiple ResNet blocks and upsampling convolutional layers. In addition, the feature maps of the encoder and the corresponding layers of the decoder are connected through a skip connection structure.

5. The low-dose CT image denoising method based on the fast diffusion model as described in claim 4, characterized in that: The ResNet block includes Norm normalization layers, Swish activation functions, and 3x3 convolutions.

6. The low-dose CT image denoising method based on the fast diffusion model as described in claim 1, characterized in that: Attention map M is obtained through the feature scaling module. att The specific processing procedure is as follows; Perform feature transformation techniques on the two sets of features Q and K respectively. Assume that both Q and K are of size 1. D represents the number of channels for the Q and K features, and Q can be obtained through transformation techniques. 00 Q 01 Q 10 Q 11 , K 00 , K 01 , K 10 , K 11 ; Enhance the interaction of Q and K information flows through interleaving and fusion: , , , ; Finally, the features are obtained by concatenating the data along the channel dimension. , shape is , This indicates a connection along the channel dimension; A convolution is added after the P feature to enhance the feature information representation, resulting in the cross-scale intermediate feature P. att Then Q and K are respectively compared with P att Multiplying them yields the corresponding vertical and horizontal attention maps H. att With K att Finally, multiply the two to calculate the attention map M. att : 。 7. A low-dose CT image denoising system based on a fast diffusion model, characterized in that, Includes the following units: The image acquisition unit is used to acquire CT images of different low-dose levels and normal-dose CT images; The fast diffusion model building unit is used to build fast diffusion models, including forward diffusion structure and reverse denoising inference; The forward diffusion unit is used to input matched low-dose and normal-dose images into the forward diffusion structure and gradually add noise in a Markov chain manner to finally obtain the target noise image. The inverse denoising inference unit is used to perform inverse denoising inference based on the target noisy image and the low-dose CT image, and gradually remove noise to obtain a clear CT image. The LUNet model consists of an encoder, intermediate layers, and a decoder. The encoder includes a ResNet block, an efficient extended perceptual downsampling module (EEP), and a symmetric perceptual self-attention mechanism (SA). The intermediate layers include convolutional layers and SASelf-attention. The encoder includes a ResNet block, SA Self-attention, and convolutional upsampling layers. The EEP module consists of two parts: an augmented perception module and a feature selection module. The augmented perception module utilizes feature transformation techniques. The specific details are as follows; in, Represented in feature map X and The numerical values ​​in the direction are spaced 2. and represent and Initial value; Assuming the EEP module inputs a feature map Shape , , For the height and width of the feature map, This represents the number of input channels. Setting the downsampling scale to 2 will reduce the input channel size. The mapping yields four sets of downsampled feature maps f 00 ,f 01 ,f 10 ,f 11 Each set of feature shapes is ( , , ), f), f fuses the features that are furthest apart to obtain a larger perceptual field of view. 00 with f 11 Adding them together gives f up f 01 with f 10 Adding them together gives f down ; After completing the perceptual augmentation, the feature selection module adaptively selects features. Specifically, it first aggregates global information using global pooling to obtain C. up With C down ; global feature vectors and The weight scores for different channels are obtained using the following formula: Adaptive optimized feature vector and The weights represent the original feature input weights. Under the guidance of adaptive feature weights, the upper and lower features are fused to obtain the final downsampling result. ; SA self-attention will input feature image The corresponding query Q, key K, and value V are obtained through 1x1 convolution. A weight-sharing strategy is used in the calculation of K and Q, sharing a single 1x1 convolution to reduce computation. The channels are split to half their original size to obtain Q and K. Subsequently, the ReLU activation function is used to sparsify the features of K and Q to obtain K. sp and Q sp Meanwhile, V is divided equally along the channel dimension to obtain V1 and V2; K sp Q sp VQ is obtained by cross-addition and reconstruction with V1 and V2 respectively. sp and VK sp This further enhances and highlights the information, improving the information flow exchange between channel groups; finally, the cross-reconstructed feature group VQ sp and VK sp The refined feature X is obtained by concatenating along the channel dimension. cross X cross With attention map M att After multiplication, the features are processed by 1x1 convolution to obtain the self-attention-enhanced features. Finally, the original input features and The result is obtained by merging and adding points one by one. .

8. A terminal device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the low-dose CT image denoising method based on the fast diffusion model as described in any one of claims 1 to 6.

Images