A low-dose CT denoising method based on tight-frame wavelet residual diffusion model

Abstract

The present invention relates to a low-dose computed tomography (CT) denoising method based on a tight frame wavelet, the method comprising: S1: collecting low-dose CT images and conventional-dose CT images to construct a CT data set and converting it into a Numpy matrix; S2: performing a GTF (Geometric Tight Framelet) transform on the image data to obtain 1 low-frequency component and 8 high-frequency components, grouping the high-frequency components, and the number of high-frequency components in each group is 2, 3, and 3; S3: creating a tight frame diffusion model (Tight Frame Diffusion Model) to denoise the low-frequency components; S4: creating a Grouped High-Frequency Enhancement Network (GHFEN) to restore the high-frequency components; S5: finally performing an inverse GTF transform on the enhanced and denoised components to obtain a denoised image. By using the present invention, a CT image denoising service that is more in line with the human eye observation effect can be provided for the low-dose CT obtained by the user through a CT scanner. As a multi-level feature-perceived low-dose CT denoising method, the present invention can be widely used in the field of CT denoising.

Description

Technical Field

[0001] The present invention relates to the technical field of image denoising, and in particular to a low-dose CT image denoising method. Background Art

[0002] In the field of medical imaging, computed tomography (CT) plays a vital role in modern clinical medicine. However, X-rays can cause irreversible damage to the human body, including an increased risk of cancer. During X-ray scans, medical professionals develop strategies to reduce X-ray dose based on actual conditions. For patients, lower X-ray intensity means a lower risk of cancer. However, for medical professionals, the resulting degradation of CT images requires more accurate diagnoses. Therefore, it is crucial to maximize the quality of LDCT images while reducing X-ray dose.

[0003] Existing LDCT image reconstruction methods are primarily categorized into three types: denoising based on sinusoidal filtering, iterative reconstruction denoising, and image post-processing denoising. The first two methods require access to difficult-to-obtain raw CT sinusoidal data, limiting their applicability. Denoising methods based on image post-processing are primarily categorized into traditional machine learning and deep learning (DL). In recent years, deep learning (DL) has enabled significant advances in image processing, achieving results far exceeding those of traditional machine learning algorithms in many respects.

[0004] In recent years, many studies on low-dose CT reconstruction have shown that the diffusion model using a denoising codec network has attracted much attention for its ability to map high-quality images from randomly sampled Gaussian noise through an iterative diffusion process. However, the reverse process starts with full-noise Gaussian noise, which makes the image restoration results diverse, while image denoising requires determinism. Secondly, the traditional wavelet transform has certain shortcomings when processing non-stationary signals or strong noise interference. Its fixed wavelet basis function is difficult to adaptively adjust according to the specific characteristics of the image, resulting in less refined denoising results. With the development of deep learning technology, wavelet-related methods based on deep learning have gradually attracted attention. By combining the multi-scale characteristics of the wavelet transform and the powerful feature learning capabilities of deep neural networks, more flexible and accurate denoising effects can be achieved in complex image features and noisy environments.

[0005] In summary, in the existing technology, traditional wavelet functions are difficult to meet the growing demand for high CT image quality, and need to be combined with deep learning networks; at the same time, the diversity of the inverse process of the diffusion model causes the restored image to contain a lot of interfering information out of thin air. Summary of the Invention

[0006] In order to solve the above technical problems, the present invention provides a low-dose CT denoising method based on a tight-frame wavelet residual diffusion model, which can provide users with a CT image denoising service that is more consistent with the observation effect of the human eye.

[0007] Specifically, the method comprises the following steps:

[0008] S1: Collect low-dose CT images and conventional-dose CT images to construct CT datasets and convert them into Numpy matrix data;

[0009] S2: Perform GTF (Geometric Tight Framelet) transformation on the image data to obtain 1 low-frequency component and 8 high-frequency components. The high-frequency components are grouped, and the number of high-frequency components in each group is 2, 3, and 3;

[0010] S3: Create a tight frame diffusion model to denoise the low-frequency components;

[0011] S4: Create a group high-frequency enhancement network to restore high-frequency components;

[0012] S5: Finally, the enhanced and denoised components are subjected to inverse GTF transformation to obtain a denoised image.

[0013] Preferably, S1 comprises the following steps:

[0014] S1.1: Acquire DICOM format CT raw data containing normal dose and low dose;

[0015] S1.2: Label the normal-dose CT data with the “target” label and the low-dose CT data with the “input” label.

[0016] S1.3: Convert the labeled normal-dose CT and low-dose CT in DICOM format into NPY matrix format data.

[0017] Preferably, S2 includes the following steps:

[0018] S2.1: Construct GTF transform and its inverse transform filter;

[0019] S2.2: Perform GTF transformation on the Numpy matrix to obtain one low-frequency component and eight high-frequency components;

[0020] S2.3: Group the transformed components. The high-frequency components are divided into three groups, with the numbers of components in each group being 2, 3, and 3 respectively.

[0021] Preferably, S2.1 includes the following steps:

[0022] Construct the basic matrices W0, W1 and W2, which will serve as the key starting modules for generating a series of complex matrix structures. The three basic matrices are as follows:

[0023]

[0024] The size of the matrix is ​​the same as the slice size of the image.

[0025] Current definition Where i, j = 0, 1, 2, Represents the Kronecker product, which is combined to generate the GTF transform filter matrix group:

[0026] D0=W 0,0 ,

[0027] D5=W 1,2

[0028] D7=W 2,1 ,

[0029] Where D0 is a low-pass filter, and the remaining components are high-pass filters obtained at different angles, and both satisfy:

[0030] D L T D H =E

[0031] D L Represents a low-pass filter, D H represents a high-pass filter, and E represents the identity matrix.

[0032] Preferably, S2.2 includes the following steps:

[0033] Given a single-channel preprocessed Numpy image data matrix X, after GTF transformation, one low-frequency component and eight high-frequency components are obtained. The description of GTF transformation is as follows:

[0034] [X L ,X H ]=Conv([D L ,D H ],X)

[0035] X L is the low-frequency component, X H is the high frequency component. The inverse transform of GTF is now defined as IGTF, which is described as follows:

[0036] X′=Conv(Transposed[D L ,DH ],[X L ,X H ])

[0037] X′ is the Numpy matrix of the image output after IGTF, Conv is the convolution operation, and Transposed is the transposition operation.

[0038] Preferably, S2.3 includes the following steps:

[0039] According to the connection between high-frequency components, the transformation components with the same first digit of the basic matrix subscript number obtained by Kronecker product are grouped into one group. The grouping is as follows:

[0040] D1 and D2 are grouped together; D3, D4 and D5 are grouped together; D6, D7 and D8 are grouped together.

[0041] Preferably, S3 includes the following steps:

[0042] S3.1: Introduce a tight frame diffusion model to perform noise diffusion and residual diffusion on low-frequency components;

[0043] S3.2: Construct residual prediction network and noise prediction network;

[0044] S3.3: reverse sampling the degraded image after diffusion to gradually remove the residual and noise in the components;

[0045] Preferably, S3.1 includes the following steps:

[0046] Order I t-1 is the low-frequency component of the previous state, I t is the low-frequency component in the current state, T is the total time step of diffusion, and the diffusion process is as follows:

[0047]

[0048] In the above formula, I res is the low-frequency component of the low-dose image after GTF transformation I in The difference between the low-frequency component I0 and the normal dose image after GTF transformation (i.e. I res =I in -I0), α t and β t Control residual diffusion and noise diffusion respectively, t∈{1,…,T}, ∈ t-1 ,…,∈~N(0,1), when t=T,

[0049] Preferably, S3.2 includes the following steps:

[0050] Use U-Net network to build residual prediction network and noise prediction network ∈ θ (I t ,t,I in ), use the L1 loss function to guide the learning of the network, the loss functions of the two networks are as follows:

[0051]

[0052] L ∈ (θ):=E[λ ∈ |∈-∈ θ (I t ,t,I in )| 2 ]

[0053] The residual loss is L res (θ), the noise loss is L ∈ (θ). λ res ,λ ε ∈{0,1}, when λ res =1 and λ ε When =1, both residual and noise are predicted.

[0054] Preferably, S3.3 includes the following steps:

[0055] I at each time step t It has been obtained from the diffusion process. Difference between them and the trained prediction network can get the reverse sampling process as follows:

[0056]

[0057] Finally, the restored low-frequency component and the low-frequency component of the normal dose after GTF transformation are subjected to MSE loss:

[0058]

[0059] is the low-frequency component obtained by the model, low is the low-frequency component obtained by GTF transformation of normal dose, L low is the MSE loss of the two, and N is the number of images.

[0060] Preferably, S4 includes the following steps:

[0061] S4.1: Construct a shallow feature extraction network for high-frequency components;

[0062] S4.2: Construct a spatial frequency enhancement module for high-frequency components;

[0063] S4.3: Construct a group-specific high-frequency enhancement network to enhance each high-frequency group component;

[0064] Preferably, S4.1 includes the following steps:

[0065] A High-Frequency Feature Extraction Module (HFEM) was designed. This network module has four branches. The first three branches are multi-scale convolutions of 1×1, 3×3, and 5×5, which are used to extract features of high-frequency components. The last branch is a learnable Sobel operator (LSO) that extracts edge information of high-frequency components. The features extracted by the shallow network are then flattened and normalized. It is described as follows:

[0066]

[0067] Relu(·) represents the Relu activation function, Conv n×n (·) represents the convolution kernel of size n, represents the splicing operation, LSO(·) represents the edge detection algorithm, Flatten(·) represents the linear representation of flattening the two-dimensional space into one dimension, LayerNorm(·) represents layer normalization, and f g Represents the local features extracted from multiple scales, f e Represents the edge features extracted by the learnable Sobel operator, f m It is a combination of local features and edge features, and finally f m Flatten and normalize to get the output f of the shallow network o .

[0068] Preferably, S4.2 includes the following steps:

[0069] Construct a spatial-frequency enhancement module Spatial-Frequency Enhance Block (SFEB), and use the State Spatial Model (SSM) and convolution operation to extract the features of the component spatial domain. It is described as follows:

[0070] f ssm1 =SiLU(x)

[0071] SiLU(·) represents the SiLU activation function, LayerNorm(·) represents layer normalization, SSM(·) represents the spatial state model, DWConv(·) represents the 1×1 separable depth convolution, It is a point-by-point matrix multiplication operation.

[0072] The phase spectrum and amplitude obtained by two-dimensional Fourier transform of the high-frequency component are convolved to obtain the frequency domain features, which are described as follows:

[0073] f input =Conv 1×1 (x)

[0074] Conv n×n Represents an n-dimensional convolution operation, FFT represents a two-dimensional Fourier transform, Amp and Phase represent amplitude and phase spectra, and IFFT represents an inverse two-dimensional Fourier transform.

[0075] Preferably, S4.3 includes the following steps:

[0076] A Grouped High-Frequency Enhancement Network (GHFEN) is designed to extract the features of each high-frequency component, fuse them with other high-frequency components, and enhance the components. Its structure is as follows:

[0077] When the number of high-frequency components within a group is 2, HFEM is used to extract the features of each component separately, and then the extracted features are spliced ​​and fused with the other component corresponding to each component, and then enhanced by a U-shaped network structure of SFFB to obtain the enhanced high-frequency components. Figure 3 .

[0078] When the number of high-frequency components within a group is 3, HFEM is used to extract the features of each component separately, and then the extracted features are spliced ​​and fused with the other two components corresponding to each component, and then enhanced by a U-shaped network structure of SFFB to obtain the enhanced high-frequency components. Figure 4 .

[0079] Finally, MSE loss is used to guide GHFEN to learn, and the expression is as follows:

[0080]

[0081] N represents the number of images, M represents the number of high-frequency components in each group, high i,j Represents the high-frequency component extracted from the normal dose, Represents the learning results of GHFEN.

[0082] Preferably, S5 includes the following steps:

[0083] S5.1: merging the enhanced high-frequency component and the enhanced low-frequency component;

[0084] S5.2: Perform an inverse GTF transform on the combined components so that the components are transformed back into the image domain;

[0085] S5.3: Evaluate the denoising effect of the model based on the calculated quantitative indicators PSNR, SSIM and qualitative human eye visual effects.

[0086] The second technical solution adopted by the present invention is: a low-dose CT denoising device based on a tight-frame wavelet residual diffusion model, comprising:

[0087] Preprocessing module: used to collect data, label the data and convert it into Numpy matrix;

[0088] GTF transformation module: used to perform wavelet transform on the image. The low-frequency component represents the image contour, and the high-frequency component represents the image edge and details.

[0089] Tight frame diffusion model module: used to enhance low-frequency components;

[0090] GHFEN: used to perform fusion enhancement on high-frequency components;

[0091] HFEM: used for feature extraction of high-frequency components;

[0092] SFEB: used to enhance the features of high-frequency components;

[0093] Image reconstruction module: used to perform inverse GTF transformation on the enhanced low-frequency and high-frequency components to generate a reconstructed CT image.

[0094] The beneficial effects of the method of the present invention are as follows: the present invention first obtains one low-frequency component and eight high-frequency components by performing a GTF transform on a Numpy matrix, restores and enhances the low-frequency components using a tight-frame diffusion model, groups the high-frequency components and fuses and enhances them separately, and finally merges the components and performs an inverse GTF transform to obtain the denoised result. At the same time, the present invention can effectively address the challenges commonly faced by existing neural network models in CT denoising. On the one hand, combining the tight-frame wavelet transform with a deep learning network not only retains the advantages of the wavelet transform in multi-scale feature extraction and frequency analysis, but also utilizes the powerful nonlinear modeling capabilities of deep learning to process complex image features and noise patterns. On the other hand, introducing residuals into the diffusion model guides the denoising of the diffused image, improves the certainty of denoising, reduces the possibility of excess interference information after denoising, and is beneficial for noise removal in CT images. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the prior art and the drawings required for use in the embodiments. The following drawings are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0096] Figure 1 This is a flow chart of a low-dose CT denoising method based on a tight-frame wavelet residual diffusion model according to the present invention;

[0097] Figure 2 2 is a schematic diagram of a specific embodiment of a low-dose CT denoising method based on a tight-frame wavelet residual diffusion model according to the present invention;

[0098] Figure 3 Schematic diagram of the structure of a three-branch GHFEN module of a low-dose CT denoising method based on a tight-frame wavelet residual diffusion model according to the present invention;

[0099] Figure 4 Schematic diagram of the dual-branch GHFEN module structure of a low-dose CT denoising method based on a tight-frame wavelet residual diffusion model of the present invention;

[0100] Figure 5 This is a schematic diagram of the HFEM module structure of a low-dose CT denoising method based on a tight-frame wavelet residual diffusion model according to the present invention;

[0101] Figure 6 Schematic diagram of the SFEB module structure of a low-dose CT denoising method based on a tight-frame wavelet residual diffusion model according to the present invention; Specific implementation plan

[0102] In order to make the purpose, features, and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making creative work are within the scope of protection of the present invention.

[0103] The embodiment of the present invention provides a method for LDCT denoising based on a tight-frame wavelet residual diffusion model, which is used to reduce the noise level caused by photon starvation in CT at the algorithm level, thereby improving the visualization effect of low-dose CT.

[0104] A typical embodiment of the present invention, taking the animal data set as an example, refers to Figure 1 、 Figure 2 , the method comprises the following steps:

[0105] S1: Collect low-dose CT images and conventional-dose CT images to construct CT datasets and convert them into Numpy matrix data;

[0106] S2: Perform GTF (Geometric Tight Framelet) transformation on the image data to obtain 1 low-frequency component and 8 high-frequency components. The high-frequency components are grouped, and the number of high-frequency components in each group is 2, 3, and 3;

[0107] S3: Create a tight frame diffusion model to denoise the low-frequency components;

[0108] S4: Create a group high-frequency enhancement network to restore high-frequency components;

[0109] S5: Finally, the enhanced and denoised components are subjected to inverse GTF transformation to obtain a denoised image.

[0110] Furthermore, as a preferred embodiment of the present method, S1 includes the following steps:

[0111] S1.1: Acquire DICOM format CT raw data containing normal dose and low dose;

[0112] Furthermore, the step of acquiring the DICOM format CT raw data including normal dose and low dose specifically includes:

[0113] Animal data sets were acquired at a source voltage of 100 kVp and a slice thickness of 0.625 mm, with the tube current adjusted from 300 mAs to as low as 15 mAs.

[0114] S1.2: Label the normal-dose CT data with the “target” label and the low-dose CT data with the “input” label.

[0115] Furthermore, the step of labeling the normal-dose CT data with a "target" label and the low-dose CT data with an "input" label in the CT raw data specifically includes:

[0116] The CT images acquired under the imaging conditions of tube current 300mAs and source voltage 100kVp were labeled with the label “target”;

[0117] The CT image acquired under the imaging conditions of tube current 15mAs and source voltage 100kVp is labeled with the label "input";

[0118] Convert the labeled normal-dose CT and low-dose CT in dicom format into NPY matrix format data.

[0119] S1.3: Convert the labeled normal-dose CT and low-dose CT in dicom format into Numpy matrix format data.

[0120] Further, refer to Figure 2 , S2 includes the following steps:

[0121] S2.1: Construct GTF transform and its inverse transform filter;

[0122] S2.2: Perform GTF transformation on the Numpy matrix to obtain one low-frequency component and eight high-frequency components;

[0123] S2.3: Group the transformed components. The high-frequency components are divided into three groups, with the numbers of components in each group being 2, 3, and 3 respectively.

[0124] Furthermore, S2.1 includes the following steps:

[0125] Construct the basic matrices W0, W1 and W2, which will serve as the key starting modules for generating a series of complex matrix structures. The three basic matrices are as follows:

[0126]

[0127] The size of the matrix is ​​the same as the slice size of the image.

[0128] Current definition Where i, j = 0, 1, 2, Represents the Kronecker product, which is combined to generate the GTF transform filter matrix group:

[0129] D0=W 0,0 ,

[0130] D5=W 1,2

[0131] D7=W 2,1 ,

[0132] Where D0 is a low-pass filter, and the remaining components are high-pass filters obtained at different angles, and both satisfy:

[0133] D L T D H =E

[0134] D L Represents a low-pass filter, D H represents a high-pass filter, and E represents the identity matrix.

[0135] Furthermore, S2.2 includes the following steps:

[0136] Given a single-channel preprocessed Numpy image data matrix X, after GTF transformation, one low-frequency component and eight high-frequency components are obtained. The description of GTF transformation is as follows:

[0137] [X L ,X H ]=Conv([D L ,D H ],X)

[0138] X L is the low-frequency component, X H is the high frequency component. The inverse transform of GTF is now defined as IGTF, which is described as follows:

[0139] X′=Conv(Transposed[D L ,D H ],[X L ,X H ])

[0140] X′ is the Numpy matrix of the image output after IGTF, Conv is the convolution operation, and Transposed is the transposition operation.

[0141] Furthermore, S2.3 includes the following steps:

[0142] According to the connection between high-frequency components, the transformation components with the same first digit of the basic matrix subscript number obtained by Kronecker product are grouped into one group. The grouping is as follows:

[0143] D1 and D2 are grouped together; D3, D4 and D5 are grouped together; D6, D7 and D8 are grouped together.

[0144] Further, refer to Figure 2 , S3 includes the following steps:

[0145] S3.1: Introduce a tight frame diffusion model to perform noise diffusion and residual diffusion on low-frequency components;

[0146] S3.2: Construct residual prediction network and noise prediction network;

[0147] S3.3: Perform reverse sampling on the diffused degraded image to gradually remove the residuals and noise in the components.

[0148] Furthermore, S3.1 includes the following steps:

[0149] Order I t-1 is the low-frequency component of the previous state, I tis the low-frequency component in the current state, T is the total time step of diffusion, and the diffusion process is as follows:

[0150]

[0151] In the above formula, I res is the low-frequency component of the low-dose image after GTF transformation I in The difference between the low-frequency component I0 and the normal dose image after GTF transformation (i.e. I res =I in -I0). α t and β t Control residual diffusion and noise diffusion, t∈{1,…,T}, ∈ t-1 ,…,∈~N(0,1), when t=T,

[0152] Furthermore, S3.2 includes the following steps:

[0153] Use U-Net network to build residual prediction network and noise prediction network ∈ θ (I t ,t,I in ), use the L1 loss function to guide the learning of the network, the loss functions of the two networks are as follows:

[0154]

[0155] L ∈ (θ):=E[λ ∈ |∈-∈ θ (I t ,t,I in )| 2 ]

[0156] The residual loss is L res (θ), the noise loss is L ∈ (θ). λ res ,λ ε ∈{0,1}, when λ res =1 and λ ε When =1, both residual and noise are predicted.

[0157] Furthermore, S3.3 includes the following steps:

[0158] I at each time step t It has been obtained from the diffusion process. Difference between them and the trained prediction network can get the reverse sampling process as follows:

[0159]

[0160] Finally, the restored low-frequency component and the low-frequency component of the normal dose after GTF transformation are subjected to MSE loss:

[0161]

[0162] is the low-frequency component obtained by the model, low is the low-frequency component obtained by GTF transformation of normal dose, L low is the MSE loss of the two, and N is the number of images.

[0163] Further, refer to Figure 3 、 Figure 4 and Figure 5 , S4 includes the following steps:

[0164] S4.1: Construct a feature extraction module for high-frequency components;

[0165] S4.2: Construct a spatial frequency enhancement module for high-frequency components;

[0166] S4.3: Construct a high-frequency feature enhancement and fusion module to enhance each high-frequency group component;

[0167] Furthermore, S4.1 includes the following steps:

[0168] A High-Frequency Feature Extraction Module (HFEM) was designed. This network module has four branches. The first three branches are multi-scale convolutions of 1×1, 3×3, and 5×5, which are used to extract features of high-frequency components. The last branch is a learnable Sobel operator (LSO) that extracts edge information of high-frequency components. The features extracted by the shallow network are then flattened and normalized. It is described as follows:

[0169]

[0170] Relu(·) represents the Relu activation function, Conv n×n (·) represents the convolution kernel of size n, represents the splicing operation, LSO(·) represents the edge detection algorithm, Flatten(·) represents the linear representation of flattening the two-dimensional space into one dimension, LayerNorm(·) represents layer normalization, and f g Represents the local features extracted from multiple scales, f e Represents the edge features extracted by the learnable Sobel operator, f m It is a combination of local features and edge features, and finally fm Flatten and normalize to get the output f of the shallow network o .

[0171] Furthermore, S4.2 includes the following steps:

[0172] Construct a spatial-frequency enhancement module Spatial-Frequency Enhance Block (SFEB), and use the State Spatial Model (SSM) and convolution operation to extract the features of the component spatial domain. It is described as follows:

[0173] f ssm1 =SiLU(x)

[0174] SiLU(·) represents the SiLU activation function, LayerNorm(·) represents layer normalization, SSM(·) represents the spatial state model, DWConv(·) represents the 1×1 separable depth convolution, It is a point-by-point matrix multiplication operation.

[0175] The phase spectrum and amplitude obtained by two-dimensional Fourier transform of the high-frequency component are convolved to obtain the frequency domain features, which are described as follows:

[0176] f input =Conv 1×1 (x)

[0177] Conv n×n Represents an n-dimensional convolution operation, FFT represents a two-dimensional Fourier transform, Amp and Phase represent amplitude and phase spectra, and IFFT represents an inverse two-dimensional Fourier transform.

[0178] Furthermore, S4.3 includes the following steps:

[0179] A Grouped High-Frequency Enhancement Network (GHFEN) is designed to extract the features of each high-frequency component, fuse them with other high-frequency components, and enhance the components. Its structure is as follows:

[0180] When the number of high-frequency components within a group is 2, HFEM is used to extract the features of each component separately, and then the extracted features are spliced ​​and fused with the other component corresponding to each component, and then enhanced by a U-shaped network structure of SFFB to obtain the enhanced high-frequency components. Figure 3 .

[0181] When the number of high-frequency components within a group is 3, HFEM is used to extract the features of each component separately, and then the extracted features are spliced ​​and fused with the other two components corresponding to each component, and then enhanced by a U-shaped network structure of SFFB to obtain the enhanced high-frequency components. Figure 4 .

[0182] Finally, MSE loss is used to guide GHFEN to learn, and the expression is as follows:

[0183]

[0184] N represents the number of images, M represents the number of high-frequency components in each group, high i,j Represents the high-frequency component extracted from the normal dose, Represents the learning results of GHFEN.

[0185] Furthermore, S5 includes the following steps:

[0186] S5.1: merging the enhanced high-frequency component and the enhanced low-frequency component;

[0187] S5.2: Perform an inverse GTF transform on the combined components so that the components are transformed back into the image domain;

[0188] S5.3: Evaluate the denoising effect of the model based on the calculated quantitative indicators PSNR, SSIM and qualitative human eye visual effects.

[0189] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make other equivalent modifications or substitutions without violating the spirit of the invention, and these equivalent modifications or substitutions are included in the scope defined by the claims.

Claims

1. A low-dose CT denoising method based on a tight-frame wavelet residual diffusion model, characterized in that: The method comprises the following steps: S1: Collect low-dose CT images and conventional-dose CT images to construct CT datasets and convert them into Numpy matrix data; S2: Perform GTF (Geometric Tight Framelet) transformation on the image data to obtain 1 low-frequency component and 8 high-frequency components. The high-frequency components are grouped, and the number of high-frequency components in each group is 2, 3, and 3; S3: Create a tight frame diffusion model to denoise the low-frequency components; S4: Create a group high-frequency enhancement network to restore high-frequency components; S5: Finally, the enhanced and denoised components are subjected to inverse GTF transformation to obtain a denoised image.

2. The low-dose CT denoising method based on a tight-frame wavelet residual diffusion model according to claim 1, characterized in that: S1 includes the following steps: S1.1: Acquire DICOM format CT raw data containing normal dose and low dose; S1.2: Label the normal-dose CT data with the "target" label and the low-dose CT data with the "input" label in the CT raw data. S1.3: Convert the labeled normal-dose CT and low-dose CT in DICOM format into NPY matrix format data.

3. The low-dose CT denoising method based on a tight-frame wavelet residual diffusion model according to claim 1, characterized in that: S2 includes the following steps: S2.1: Construct GTF transform and its inverse transform filter; S2.2: Perform GTF transformation on the Numpy matrix to obtain one low-frequency component and eight high-frequency components; S2.3: Group the transformed components. The high-frequency components are divided into three groups, with the numbers of components in each group being 2, 3, and 3 respectively. S2.1 includes the following steps: Construct the basic matrices W0, W1 and W2, which will serve as the key starting modules for generating a series of complex matrix structures. The three basic matrices are as follows: The size of the matrix is ​​consistent with the slice size of the image; Current definition Where i, j = 0, 1, 2, Represents the Kronecker product, which is combined to generate the GTF transform filter matrix group: D0=W 0,0 , D5=W 1,2 D7=W 2,1 , Where D0 is a low-pass filter, and the remaining components are high-pass filters obtained at different angles, and both satisfy: D L T D H =E D L Represents a low-pass filter, D H represents a high-pass filter, and E represents the identity matrix; Among them, S2.2 includes the following steps: Given a single-channel preprocessed Numpy image data matrix X, after GTF transformation, one low-frequency component and eight high-frequency components are obtained. The description of GTF transformation is as follows: [X L ,X H ]=Conv([D L ,D H ],X) X L is the low-frequency component, X H It is the high frequency component. The inverse transform of GTF transform is now defined as IGTF, which is described as follows: X′=Conv(Transposed[D L ,D H ],[X L ,X H ]) X′ is the Numpy matrix of the image output after IGTF, Conv is the convolution operation, and Transposed is the transposition operation; Among them, S2.3 includes the following steps: According to the connection between high-frequency components, the transformation components with the same first digit of the basic matrix subscript number obtained by Kronecker product are grouped into one group. The grouping is as follows: D1 and D2 are grouped together; D3, D4 and D5 are grouped together; D6, D7 and D8 are grouped together.

4. The low-dose CT denoising method based on a tight-frame wavelet residual diffusion model according to claim 1, characterized in that: S3 includes the following steps: S3.1: Introduce a tight frame diffusion model to perform noise diffusion and residual diffusion on low-frequency components; S3.2: Construct residual prediction network and noise prediction network; S3.3: reverse sampling the degraded image after diffusion to gradually remove the residual and noise in the components; Among them, S3.1 includes the following steps: Order I t-1 is the low-frequency component of the previous state, I t is the low-frequency component in the current state, T is the total time step of diffusion, and the diffusion process is as follows: In the above formula, I res is the low-frequency component of the low-dose image after GTF transformation I in The difference between the low-frequency component I0 obtained after GTF transformation of the normal dose image is I res =I in -I0,α t and β t Control residual diffusion and noise diffusion, t∈{1,…,T}, ∈ t-1 ,…,∈~N(0,1), when t=T, Among them, S3.2 includes the following steps: Use U-Net network to build residual prediction network and noise prediction network ∈ θ (I t ,t,I in ), use the L1 loss function to guide the learning of the network, the loss functions of the two networks are as follows: L ∈ (θ):=E[λ ∈ |∈-∈ θ (I t ,t,I in )| 2 ] The residual loss is L res (θ), the noise loss is L ∈ (θ), λ res ,λ ε ∈{0,1}, when λ res =1 and λ ε =1, both residual and noise are predicted; S3.3 includes the following steps: I at each time step t It has been obtained from the diffusion process. Difference between them and the trained prediction network can get the reverse sampling process as follows: Finally, the restored low-frequency component and the low-frequency component of the normal dose after GTF transformation are subjected to MSE loss: is the low-frequency component obtained by the model, low is the low-frequency component obtained by GTF transformation of normal dose, L low is the MSE loss of the two, and N is the number of images.

5. The low-dose CT denoising method based on a tight-frame wavelet residual diffusion model according to claim 1, characterized in that: S4 includes the following steps: S4.1: Construct a feature extraction module for high-frequency components; S4.2: Construct a spatial frequency enhancement module for high-frequency components; S4.3: Construct a group-specific high-frequency enhancement network to enhance each high-frequency group component; Among them, S4.1 includes the following steps: A High-frequency Feature Extraction Module (HFEM) is designed. This network module has four branches. The first three branches are 1×1, 3×3, and 5×5 multi-scale convolutions for extracting features of high-frequency components. The last branch is a learnable Sobel operator (LSO) for extracting edge information of high-frequency components. The features extracted by the shallow network are then flattened and normalized. It is described as follows: Relu(·) represents the Relu activation function, Conv n×n (·) represents the convolution kernel of size n, represents the splicing operation, LSO(·) represents the edge detection algorithm, Flatten(·) represents the linear representation of flattening the two-dimensional space into one dimension, LayerNorm(·) represents layer normalization, and f g Represents the local features extracted from multiple scales, f e Represents the edge features extracted by the learnable Sobel operator, f m It is a combination of local features and edge features, and finally f m Flatten and normalize to get the output f of the shallow network o ; Among them, S4.2 includes the following steps: Construct a spatial-frequency enhancement module Spatial-Frequency Enhance Block (SFEB), and use the State Spatial Model (SSM) and convolution operation to extract the features of the component spatial domain. It is described as follows: f ssm1 =SiLU(x) SiLU(·) represents the SiLU activation function, LayerNorm(·) represents layer normalization, SSM(·) represents the spatial state model, DWConv(·) represents the 1×1 separable depth convolution, It is a matrix point-by-point multiplication operation; The phase spectrum and amplitude obtained by two-dimensional Fourier transform of the high-frequency component are convolved to obtain the frequency domain features, which are described as follows: f input =Conv 1×1 (x) Conv n×n Represents n-dimensional convolution operation, FFT represents two-dimensional Fourier transform, Amp and Phase represent amplitude and phase spectra, and IFFT represents two-dimensional inverse Fourier transform; Among them, S4.3 includes the following steps: A Grouped High-Frequency Enhancement Network (GHFEN) is designed to extract the features of each high-frequency component, fuse them with other high-frequency components, and enhance the components. Its structure is as follows: When the number of high-frequency components within a group is 2, HFEM is used to extract the features of each component separately, and then the extracted features are spliced ​​and fused with the other component corresponding to each component. Subsequently, they are enhanced through a U-shaped network structure of SFFB to obtain the enhanced high-frequency components. The detailed structure is shown in Figure 3. When the number of high-frequency components within a group is 3, HFEM is used to extract the features of each component separately. The extracted features are then concatenated and fused with the other two components corresponding to each component. Subsequently, the features are enhanced through a U-shaped network structure of SFFB to obtain the enhanced high-frequency components. The detailed structure is shown in Figure 4. Finally, MSE loss is used to guide GHFEN to learn, and the expression is as follows: N represents the number of images, M represents the number of high-frequency components in each group, high i,j Represents the high-frequency component extracted from the normal dose, Represents the learning results of GHFEN.

6. The low-dose CT denoising method based on a tight-frame wavelet residual diffusion model according to claim 1, characterized in that: S5 includes the following steps: S5.1: merging the enhanced high-frequency component and the enhanced low-frequency component; S5.2: Perform an inverse GTF transform on the combined components so that the components are transformed back into the image domain; S5.3: Evaluate the denoising effect of the model based on the calculated quantitative indicators PSNR, SSIM and qualitative human eye visual effects.

7. A low-dose CT denoising device, applying the low-dose CT denoising method of the tight-frame wavelet residual diffusion model according to claim 1, characterized by: Preprocessing module: used to collect data, label the data and convert it into Numpy matrix; GTF transformation module: used to perform wavelet transform on the image. The low-frequency component represents the image contour, and the high-frequency component represents the image edge and details. Tight frame diffusion model module: used to enhance low-frequency components; HFEM: used to extract high-frequency component features; SFFB: used to enhance the spatial and frequency domains of high-frequency components; GHFEN: used to perform fusion enhancement between high-frequency components; Image reconstruction module: used to perform inverse GTF transformation on the enhanced low-frequency components and high-frequency components to generate a reconstructed CT image.

Images