A limited-angle CT reconstruction artifact suppression method based on multi-domain feature fusion network

Abstract

The present application belongs to the field of CT tomographic reconstruction technology and artificial intelligence, and discloses a limited-angle CT reconstruction artifact suppression method based on a multi-domain feature fusion network. In view of the problem that the reconstruction result of traditional CT scanning under limited-angle conditions is prone to artifacts and structural distortion, thereby affecting the image quality and defect detection accuracy, the present application constructs a multi-domain feature fusion artifact suppression network, takes the limited-angle reconstruction result as input, and realizes artifact suppression and detail recovery through the synergistic effect of the encoder part, the decoder part, the feature enhancement part and the feature conversion part. The present application can obtain high-quality tomographic images under limited-angle conditions, effectively reduces the scanning angle and time of industrial CT detection, improves the imaging clarity and reliability without increasing the radiation dose, is suitable for industrial detection of complex structure workpieces, and has important industrial application value.

Description

Technical Field

[0001] This invention belongs to the fields of CT tomographic reconstruction technology and artificial intelligence, and relates to a method for suppressing artifacts in finite-angle CT reconstruction based on a multi-domain feature fusion network. Background Technology

[0002] Computed Tomography (CT) is a three-dimensional imaging technique based on the principle of X-ray projection. By acquiring projection data from multiple angles and performing mathematical inversion, it can obtain the three-dimensional distribution of density or attenuation coefficients within an object. Compared with traditional two-dimensional X-ray imaging, CT can provide more accurate spatial structural information and is widely used in fields such as industrial non-destructive testing, materials analysis, and internal structure research. In industrial CT inspection, to improve scanning speed and reconstruction efficiency, finite-angle acquisition is often used instead of full-angle scanning. This finite-angle acquisition leads to the loss of projection information, resulting in artifacts and structural distortions in the reconstructed images, severely affecting image quality and subsequent defect identification and dimensional analysis.

[0003] Traditional finite-angle CT reconstruction typically relies on iterative optimization methods to mitigate reconstruction artifacts caused by missing projection data. Compared to direct reconstruction, iterative optimization algorithms effectively suppress artifacts and structural distortions by introducing prior constraints or regularization models during the reconstruction process. For example, patent CN109840927A proposes a finite-angle CT reconstruction method based on anisotropic total variational constraints, achieving artifact suppression through sparse constraints; patent CN112288762A discloses a discrete iterative reconstruction method for finite-angle CT, controlling reconstruction stability through regularization terms. These methods improve CT reconstruction quality under finite-angle conditions to some extent, but they generally suffer from problems such as numerous iterations, long reconstruction times, and less than ideal reconstruction results, failing to meet the dual demands of high precision and real-time performance in industrial inspection.

[0004] With the development of artificial intelligence technology, deep learning has been applied to the suppression of artifacts in CT reconstruction. These methods, using structures such as Convolutional Neural Networks (CNNs), U-Nets, and Generative Adversarial Networks (GANs), learn the nonlinear mapping relationship between tomographic images with artifacts and real images from a large amount of training data, achieving automatic artifact identification and removal. Patent CN115690255A discloses a CT image artifact removal method based on convolutional neural networks, achieving artifact suppression and detail enhancement for lithium battery CT data; Patent CN110211194A proposes a method for removing sparse angle CT imaging artifacts based on deep learning, effectively suppressing artifacts by constructing a neural network to learn the mapping between full-angle and sparse-angle reconstruction results. Although existing deep learning-based reconstruction methods have made some progress in artifact suppression, most only process in a single or dual domain in the spatial, wavelet, and frequency domains, still suffering from artifact residue, insufficient detail recovery, and easy loss of texture. For some complex and textured workpieces, it is difficult to meet the high-precision detection requirements.

[0005] Unlike existing methods, this invention proposes an artifact suppression method that combines spatial, wavelet, and frequency domain features for multi-domain feature fusion learning. This method constructs an end-to-end multi-scale neural network, using the blurred result obtained after finite-angle projection through the traditional reconstruction algorithm FDK (Feldkamp-Davis-Kress) as input, and the clear result of full-angle scan reconstruction as ground-value supervised learning. The final output is an artifact-free, high-fidelity result with both artifact suppression and structural refinement. Specifically, the network of this invention adopts an encoder-decoder structure: the encoder uses a multi-domain feature fusion module and a spatial-channel attention mechanism to progressively extract structural and artifact features from the input image; the decoder also uses a multi-domain feature fusion module to progressively restore clean images and high-frequency details. Simultaneously, a feature transformation part is introduced between corresponding layers of the encoder and decoder for cross-layer feature fusion and enhancement to improve artifact suppression. By fusing multi-domain features from the spatial, wavelet, and frequency domains, this invention can further improve the artifact removal effect of finite-angle CT reconstruction while maintaining a more complete structure, providing a more efficient and scalable artifact suppression scheme for finite-angle CT reconstruction of complex structural samples. Summary of the Invention

[0006] This invention provides a limited-angle CT reconstruction artifact suppression method based on a multi-domain feature fusion network of spatial domain, wavelet domain, and frequency domain. It reconstructs a clear three-dimensional tomographic structure using a projected view within a limited angular range of the tested sample, thereby reducing the scanning angle while maintaining detection accuracy and improving the efficiency and applicability of industrial CT inspection. The overall architecture of the proposed network model is as follows: Figure 1As shown, through multi-domain fusion feature processing, effective removal of finite-angle artifacts is achieved.

[0007] The technical solution of the present invention:

[0008] A method for suppressing artifacts in finite-angle CT reconstruction based on a multi-domain feature fusion network, comprising the following steps:

[0009] Step 1: Construct a "tomographic-tomographic" CT dataset;

[0010] 360° projection data of the workpiece is acquired through full-angle acquisition, and a 3D tomographic image of the workpiece is generated using the FDK reconstruction algorithm, which serves as the ground truth for the multi-domain feature fusion artifact suppression network. Simultaneously, projection data of the workpiece is acquired using finite-angle acquisition, such as 90° or 120°, and a blurred tomographic image of the workpiece containing artifacts is generated using the FDK reconstruction algorithm, which serves as the input to the multi-domain feature fusion artifact suppression network. Based on the 3D tomographic image of the workpiece and the blurred tomographic image containing artifacts, a "tomographic-tomographic" CT dataset is constructed for training and validating the multi-domain feature fusion artifact suppression network.

[0011] Step 2: Construct a multi-domain feature fusion artifact suppression network;

[0012] The multi-domain feature fusion artifact suppression network transforms blurry tomographic images containing artifacts into clear CT images with artifact removal. It includes an encoder, feature enhancement, decoder, and feature transformation components. The network architecture is as follows: Figure 1 As shown;

[0013] (1) Encoder section: includes a channel expansion module, three multi-domain feature fusion modules with the same structure and three spatial-channel attention modules with the same structure. It performs feature extraction and progressive downsampling on the input blurred tomographic image containing artifacts, and converts the blurred tomographic image containing artifacts into a compact representation with a higher number of channels. The blurred tomographic image containing artifacts is downsampled 1, 2 and 3 times respectively and then input to the three spatial-channel attention modules from top to bottom. At the same time, after the blurred tomographic image containing artifacts passes through the channel expansion module, the output feature is used as the input of the first multi-domain feature fusion module. The output of the first multi-domain feature fusion module is downsampled and fused with the first spatial-channel attention module as the input of the second multi-domain feature fusion module. The output of the second multi-domain feature fusion module is downsampled and fused with the output of the second spatial-channel attention module as the input of the third multi-domain feature fusion module. The output of the third multi-domain feature fusion module is downsampled and fused with the output of the third spatial-channel attention module as the final output of the encoder section and sent to the feature enhancement section.

[0014] The channel expansion module mainly consists of convolutional layers. The number of input channels of the convolutional layer is 3 and the number of output channels is 24. The number of channels of the input image is increased by convolutional layers, thereby enhancing the feature representation capability of the multi-domain feature fusion artifact suppression network.

[0015] The multi-domain feature fusion module is used to perform parallel modeling of features in the spatial, wavelet, and frequency domains to obtain multi-domain joint features with complementary expressive capabilities. The specific structure is as follows: Figure 2 As shown, the specific implementation process of the multi-domain feature fusion module is as follows: The input features first pass through a normalization layer, and then the output results are simultaneously used as inputs to the parallel spatial domain branch, wavelet domain branch, and frequency domain branch. Feature fusion is performed on the outputs of the spatial domain branch, wavelet domain branch, and frequency domain branch. Among them, the spatial domain branch adopts a visual state space model block based on the Mamba structure to perform sequential modeling of the input features, forming the spatial features output by the spatial domain branch. The wavelet domain branch includes a visual state space module based on the Mamba structure, a convolutional feature fusion module, and a cross-attention module. First, the input features are decomposed into a low-frequency component LL and three high-frequency components LH, HL, and HH using discrete wavelet transform. Then, the low-frequency component LL is processed by a Mamba-based... The visual state space module modeling of the structure yields global structural features q1. Simultaneously, high-frequency components LH, HL, and HH are fused into a fused feature q2 via a convolutional feature fusion module. Subsequently, a cross-attention module adaptively fuses the global structural features q1 and the fused feature q2. The adaptively fused features are then processed by convolution and inverse wavelet transform to restore them to the spatial domain, forming the wavelet domain features output by the wavelet domain branch. The frequency domain branch first maps the input features to the frequency domain using Fourier transform, performing feature decomposition and modeling in amplitude and phase. Then, position encoding is implemented using depthwise convolution, and frequency domain features at different scales are extracted using dilated convolutions with different dilation rates. The dilated convolution features are further processed by convolution and then restored to the spatial domain via inverse Fourier transform, forming the frequency domain features output by the frequency domain branch. Finally, the spatial, wavelet, and frequency domain features output by the spatial, wavelet, and frequency domain branches are fused to generate the output of the multi-domain feature fusion module.

[0016] The spatial-channel attention module uses blurred tomographic images containing artifacts after 1st, 2nd, and 3rd downsampling as input features. Based on the structural distribution and artifact characteristics in the input features, it generates channel and spatial attention weights and outputs prior enhancement features at different scales. These prior enhancement features are then fused with the output of the multi-domain feature fusion module. The spatial-channel attention module employs a convolutional block attention mechanism, including channel attention branches and spatial attention branches. The channel attention branch generates channel weights through global statistical analysis of the input features, which are used to dynamically adjust the weights of different channels. The spatial attention branch generates a spatial weight map through convolution based on the output feature map of the channel attention branch.

[0017] (2) Feature enhancement part: It consists of a multi-domain feature fusion module, located between the encoder part and the decoder part, and is used to process the features extracted by the encoder part in depth in the global scope; the feature enhancement part takes the output features of the encoder part as input and obtains the enhanced feature map through the multi-domain feature fusion mechanism.

[0018] (3) Decoder section: It consists of three multi-domain feature fusion modules and convolutional layers from top to bottom. The multi-domain feature fusion modules have the same structure as the multi-domain feature fusion modules in the encoder section. The input of the decoder section is the output features of the feature transformation section. After upsampling at each level, the features are passed through the three multi-domain feature fusion modules in sequence to supplement and reconstruct the features at different levels, and restore the spatial scale and structural information of the feature map. Finally, the number of channels is compressed from 24 to 3 through the convolutional layer to generate a clear CT image after artifact removal.

[0019] (4) Feature transformation part: includes three layers of concatenation and convolution operations. From top to bottom, the input of the first layer of concatenation and convolution operation is the output feature of the first multi-domain feature fusion module of the encoder part and the output feature of the second multi-domain feature fusion module of the decoder part. The input of the second layer of concatenation and convolution operation is the output feature of the second multi-domain feature fusion module of the encoder part and the output feature of the third multi-domain feature fusion module of the decoder part. The input of the third layer of concatenation and convolution operation is the output feature of the third multi-domain feature fusion module of the encoder part and the output feature of the feature enhancement part.

[0020] Each concatenation and convolution operation concatenates the two inputs and then performs channel compression and spatial modeling on the concatenated features through convolution. Finally, the output of the first concatenation and convolution operation is input into the first multi-domain feature fusion module of the decoder for further processing, the output of the second concatenation and convolution operation is input into the second multi-domain feature fusion module of the decoder for further processing, and the output of the third concatenation and convolution operation is input into the third multi-domain feature fusion module of the decoder for further processing.

[0021] Step 3: Train a multi-domain feature fusion artifact suppression network;

[0022] In the "tomographic-tomographic" CT dataset established in Step 1, the 3D tomographic image serves as the supervised ground truth for the multi-domain feature fusion artifact suppression network, while the blurred tomographic image containing artifacts serves as the input to the network. The input blurred tomographic image containing artifacts is named x, and the corresponding 3D tomographic image is named y. The loss function used consists of MSE loss and L1 loss. and The weight parameters for the loss are shown in equation (1):

[0023] (1)

[0024] After training for a certain number of algebras, the network converges, and the multi-domain feature fusion artifact suppression network is trained. Once training is complete, the tomographic data of the workpiece obtained in step 1 is input into the multi-domain feature fusion artifact suppression network, and a clear CT image after artifact removal is output.

[0025] The beneficial effects of this invention are:

[0026] (1) This invention proposes a finite-angle CT reconstruction artifact suppression method based on a multi-domain feature fusion network. Unlike existing single-domain or dual-domain methods, this invention achieves accurate identification and effective suppression of artifacts and detail loss in finite-angle CT reconstruction images by jointly modeling features in the spatial domain, wavelet domain, and frequency domain. This method significantly improves the clarity, structural fidelity, and visual quality of the reconstructed images while ensuring the integrity of structural details and the stability of texture information.

[0027] (2) The multi-domain feature fusion artifact suppression network constructed in this invention can be integrated into an industrial CT detection system to achieve high-quality CT reconstruction through limited angle scanning, reduce detection time and equipment movement complexity, and has high engineering practical value and promotion potential. Attached Figure Description

[0028] Figure 1 This is a diagram of the overall architecture of the multi-domain feature fusion artifact suppression network of the present invention.

[0029] Figure 2 This is an architecture diagram of the multi-domain feature fusion module of the present invention. Detailed Implementation

[0030] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0031] First, obtain the finite-angle projection data and full-angle projection data of the workpiece to be inspected, and perform FDK reconstruction respectively; the "tomographic-tomographic" CT dataset constructed according to step 1 is used to train and validate the multi-domain feature fusion artifact suppression network.

[0032] Then, follow step 2 and combine Figure 2A multi-domain feature fusion artifact suppression network is constructed as follows: First, an encoder is built to extract features from the input 1024×1024 blurred tomographic image containing artifacts and perform progressive downsampling, converting the image into a compact representation with a higher number of channels. The blurred tomographic image containing artifacts is then downsampled 1, 2, and 3 times, and input from top to bottom to three spatial-channel attention modules. Simultaneously, the blurred tomographic image containing artifacts is expanded to 24 channels by a channel expansion module, and the output features are used as input to the first multi-domain feature fusion module. The output of the first multi-domain feature fusion module is downsampled and fused with the first spatial-channel attention module, serving as input to the second multi-domain feature fusion module. The output of the second multi-domain feature fusion module is downsampled and fused with the output of the second spatial-channel attention module, serving as input to the third multi-domain feature fusion module. The output of the third multi-domain feature fusion module is downsampled and fused with the output of the third spatial-channel attention module. At this point, the feature map size is reduced to 128×128, which is then used as the final output of the encoder and fed into the feature enhancement part. Then, a feature enhancement section is set between the encoder and decoder, consisting of multiple multi-domain feature fusion modules. The input is the output features of the encoder, and the output is fed into the feature transformation section, used to integrate and enhance the encoder's output features at the network's bottom layer. Next, the decoder section is constructed, consisting of three multi-domain feature fusion modules from top to bottom and a convolutional layer. The decoder's input is the output features of the feature transformation section. Through progressive upsampling and then sequentially passing through the three multi-domain feature fusion modules, features at different levels are supplemented and reconstructed, gradually restoring the feature map resolution to 1024×1024. Finally, the feature transformation section is constructed, achieving feature transfer and fusion between the encoder and decoder at the same layer through concatenation and convolution operations. Finally, the convolutional layer of the decoder further restores and converts the number of channels to 3, outputting a high-quality CT image.

[0033] Finally, train the multi-domain feature fusion artifact suppression network according to step 3. Specifically: select multiple sets of artifact-containing blurred tomographic images and 3D tomographic images of workpieces as a training set, and select several other sets of artifact-containing blurred tomographic images and 3D tomographic images as a validation set. Training and validation can be completed following the procedure in step 3.

[0034] After completing the training and validation of the network, input the blurred tomographic image containing artifacts into the network according to step 3 to obtain a high-quality restored image.

Claims

1. A method for suppressing artifacts in finite-angle CT reconstruction based on a multi-domain feature fusion network, characterized in that, The steps are as follows: Step 1: Construct a "tomographic-tomographic" CT dataset; The workpiece's 360° projection data is acquired by full-angle acquisition, and a 3D tomographic image of the workpiece is generated using the FDK reconstruction algorithm, which serves as the ground truth for the multi-domain feature fusion artifact suppression network. At the same time, the workpiece's projection data is acquired by finite-angle acquisition, and a blurred tomographic image of the workpiece containing artifacts is generated using the FDK reconstruction algorithm, which serves as the input to the multi-domain feature fusion artifact suppression network. A "tomographic-tomographic" CT dataset was constructed based on the 3D tomographic images of the workpiece and the blurred tomographic images containing artifacts, which was used to train and validate a multi-domain feature fusion artifact suppression network. Step 2: Construct a multi-domain feature fusion artifact suppression network; The multi-domain feature fusion artifact suppression network transforms blurry tomographic images containing artifacts into clear CT images after artifact removal. It includes an encoder part, a feature enhancement part, a decoder part, and a feature transformation part. Step 3: Train a multi-domain feature fusion artifact suppression network; In the "tomographic-tomographic" CT dataset established in Step 1, the 3D tomographic image serves as the supervised ground truth for the multi-domain feature fusion artifact suppression network, while the blurred tomographic image containing artifacts serves as the input to the network. The input blurred tomographic image containing artifacts is named x, and the corresponding 3D tomographic image is named y. The loss function used consists of MSE loss and L1 loss. and The weight parameters for the loss are shown in equation (1): (1) ； After training for a certain number of algebras, the network converges, and the multi-domain feature fusion artifact suppression network is trained. After training is completed, the tomographic data of the workpiece obtained in step 1 is input into the multi-domain feature fusion artifact suppression network, and the clear CT image after artifact removal is output. The encoder part includes a channel expansion module, three multi-domain feature fusion modules with the same structure, and three spatial-channel attention modules with the same structure. The multi-domain feature fusion module is used to model features in parallel in the spatial domain, wavelet domain, and frequency domain to obtain multi-domain joint features with complementary expressive capabilities. The specific implementation process of the multi-domain feature fusion module is as follows: the input features first pass through a normalization layer, and then the output results are simultaneously used as inputs to the parallel spatial domain branch, wavelet domain branch, and frequency domain branch. Feature fusion is performed on the outputs of the spatial domain branch, wavelet domain branch, and frequency domain branch.

2. The finite-angle CT reconstruction artifact suppression method based on multi-domain feature fusion network according to claim 1, characterized in that, The encoder section includes a channel expansion module, three structurally identical multi-domain feature fusion modules, and three structurally identical spatial-channel attention modules. It performs feature extraction and progressive downsampling on the input blurred tomographic image containing artifacts, converting the blurred tomographic image containing artifacts into a compact representation with a higher number of channels. The blurred tomographic image containing artifacts is downsampled 1, 2, and 3 times respectively and then input from top to bottom to the three spatial-channel attention modules. Meanwhile, the blurred tomographic image containing artifacts, after passing through the channel expansion module, has its output features used as input to the first multi-domain feature fusion module. The output of the first multi-domain feature fusion module is downsampled and fused with the first spatial-channel attention module, serving as input to the second multi-domain feature fusion module. The output of the second multi-domain feature fusion module is downsampled and fused with the output of the second spatial-channel attention module, serving as input to the third multi-domain feature fusion module. The output of the third multi-domain feature fusion module is downsampled and fused with the output of the third spatial-channel attention module, serving as the final output of the encoder and fed into the feature enhancement section.

3. The finite-angle CT reconstruction artifact suppression method based on multi-domain feature fusion network according to claim 2, characterized in that, The channel expansion module mainly consists of convolutional layers, with 3 input channels and 24 output channels. In the multi-domain feature fusion module, the spatial branch adopts a visual state space model block based on the Mamba structure to perform serialization modeling of the input features, forming the spatial features output by the spatial branch; The wavelet domain branch includes a visual state space module based on the Mamba structure, a convolutional feature fusion module, and a cross-attention module. First, the input features are decomposed into a low-frequency component LL and three high-frequency components LH, HL, and HH using discrete wavelet transform. Then, the low-frequency component LL is modeled by the visual state space module based on the Mamba structure to obtain the global structural feature q1. At the same time, the high-frequency components LH, HL, and HH are generated into a fused feature q2 after being processed by the convolutional feature fusion module. Subsequently, the global structural feature q1 and the fused feature q2 are adaptively fused through the cross-attention module. The adaptively fused features are then processed by convolution and inverse wavelet transform to restore them to the spatial domain, forming the wavelet domain features output by the wavelet domain branch. The frequency domain branch first maps the input features to the frequency domain using Fourier transform, and then performs feature decomposition and modeling on amplitude and phase respectively. Subsequently, position encoding is implemented using depthwise convolution, and frequency domain features of different scales are extracted by combining dilated convolutions with different dilation rates. The features after dilated convolution are further processed by convolution and then restored to the spatial domain by inverse Fourier transform, forming the frequency domain features output by the frequency domain branch. The spatial domain features, wavelet domain features and frequency domain features output by the spatial domain branch, wavelet domain branch and frequency domain branch are fused to generate the output of the multi-domain feature fusion module. The spatial-channel attention module uses blurred tomographic images containing artifacts after 1, 2, and 3 downsampling as input features. Based on the structural distribution and artifact characteristics in the input features, it generates channel and spatial attention weights and outputs prior enhancement features at different scales. The prior enhancement features are then fused with the output of the multi-domain feature fusion module. The spatial-channel attention module adopts a convolutional block attention mechanism, which includes a channel attention branch and a spatial attention branch: the channel attention branch generates channel weights by performing global statistics on the input features, which is used to dynamically adjust the weights of different channels; the spatial attention branch generates a spatial weight map based on the output feature map of the channel attention branch through convolution.

4. The finite-angle CT reconstruction artifact suppression method based on multi-domain feature fusion network according to claim 3, characterized in that, Feature enhancement section: Composed of a multi-domain feature fusion module, located between the encoder and decoder sections, it is used to perform in-depth processing of features extracted by the encoder section on a global scale; The feature enhancement part takes the output features of the encoder part as input and obtains the enhanced feature map through a multi-domain feature fusion mechanism.

5. The finite-angle CT reconstruction artifact suppression method based on multi-domain feature fusion network according to claim 4, characterized in that, The decoder section consists of three multi-domain feature fusion modules and a convolutional layer, arranged from top to bottom. The multi-domain feature fusion modules have the same structure as those in the encoder section. The input to the decoder section is the output features of the feature transformation section. After being upsampled step by step, the features are then passed through the three multi-domain feature fusion modules in sequence to supplement and reconstruct the features at different levels, restoring the spatial scale and structural information of the feature map. Finally, the number of channels is compressed from 24 to 3 through the convolutional layer to generate a clear CT image after artifact removal.

6. The finite-angle CT reconstruction artifact suppression method based on multi-domain feature fusion network according to claim 5, characterized in that, Feature transformation section: includes three layers of concatenation and convolution operations. From top to bottom, the input of the first layer of concatenation and convolution operation is the output features of the first multi-domain feature fusion module of the encoder and the output features of the second multi-domain feature fusion module of the decoder. The input of the second layer of concatenation and convolution operation is the output features of the second multi-domain feature fusion module of the encoder and the output features of the third multi-domain feature fusion module of the decoder. The input of the third layer of concatenation and convolution operation is the output features of the third multi-domain feature fusion module of the encoder and the output features of the feature enhancement section. Each concatenation and convolution operation concatenates the two inputs and then performs channel compression and spatial modeling on the concatenated features through convolution. Finally, the output of the first concatenation and convolution operation is input into the first multi-domain feature fusion module of the decoder for further processing, the output of the second concatenation and convolution operation is input into the second multi-domain feature fusion module of the decoder for further processing, and the output of the third concatenation and convolution operation is input into the third multi-domain feature fusion module of the decoder for further processing.

Images