Sparse CT reconstruction method fusing channel weighting and residual dense network
By fusing channel weighting and residual dense networks into a sparse CT reconstruction method, the problems of stripe artifacts and insufficient structural modeling in sparse CT reconstruction are solved, and high-quality image reconstruction results are achieved.
Patent Information
- Application Number
- CN202511168213.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-20
- Publication Date
- 2025-12-05
AI Technical Summary
Existing sparse CT reconstruction methods are prone to streak artifacts when reducing radiation dose, affecting diagnostic accuracy, and lack modeling of long-distance dependencies on image structure, which limits the improvement of image quality.
A sparse CT reconstruction method combining channel weighting and residual dense networks is adopted. By constructing a residual dense network based on channel fusion, dense connections, residual networks and large kernel attention mechanisms are introduced. The feature weights are adaptively adjusted by combining the attention mechanism, and deep neural networks are used for encoding operations to extract structural features and perform deep modeling.
It improves the accuracy and quality of reconstructed images, completes missing or inaccurate projection information in sparse projection, and significantly enhances edge sharpness, tissue structure integrity, and noise control.
Smart Images
Figure CN121074178A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of medical image processing, in particular to a sparse CT reconstruction method fusing channel weighting and residual dense network. BACKGROUND
[0002] X-rays have radiation hazards to the human body, especially to sensitive groups such as children and pregnant women. In order to reduce the risk of radiation, medical professionals usually reduce the tube current (tube voltage) or reduce the sampling frequency to reduce the dose of X-rays. However, for sparse CT with reduced sampling frequency, if the traditional filtered back projection (FBP) method is used for image reconstruction, it will cause serious stripe artifacts, affecting the accuracy of diagnosis. In recent years, deep learning technology has made significant progress in the field of image processing and has been gradually applied to CT image reconstruction. However, most existing methods are limited to image domain processing artifacts, and fail to effectively utilize the physical information in the projection domain, and lack modeling of long-range dependencies of image structure, thereby limiting the further improvement of image quality. SUMMARY
[0003] In view of the above problems in the prior art, the present application provides a sparse CT reconstruction method fusing channel weighting and residual dense network.
[0004] In order to achieve the above-mentioned purposes, the technical scheme adopted by the present application is as follows: A sparse CT reconstruction method fusing channel weighting and residual dense network, comprising the following steps: S1, performing filtered back projection and re-projection operation on the original sparse sinogram to generate a full-size re-projected sinogram; S2, constructing a residual dense network based on channel fusion, which includes a convolution normalization activation layer, a down-sampling layer, a first residual dense channel fusion block, a second residual dense channel fusion block, a third residual dense channel fusion block, a fourth residual dense channel fusion block, a large kernel attention block, and a deconvolution layer; S3, inputting the full-size re-projected sinogram into the residual dense network based on channel fusion for training, introducing an improved joint loss function for optimization, learning the error between the recovered sinogram and the reference sinogram, and the error between the recovered sinogram filtered back projection reconstructed image and the reference sinogram filtered back projection reconstructed image, to generate a trained residual dense network based on channel fusion; S4, inputting the acquired sparse sinogram into the trained residual dense network based on channel fusion for projection recovery to generate a recovered sinogram; S5, performing filtered back projection on the recovered sinogram to generate a CT image without artifacts.
[0005] The present application has the following advantages: The application provides a sparse CT reconstruction method fusing channel weighting and residual dense network. BRIEF DESCRIPTION OF DRAWINGS
[0006] Figure 1 FIG. 1 is a flowchart of the sparse CT reconstruction method fusing channel weighting and residual dense network according to the application; Figure 2 FIG. 2 is a structure diagram of the residual dense network based on channel fusion in the embodiment; Figure 3 FIG. 3 is a structure diagram of the residual dense channel fusion block in the embodiment; Figure 4 FIG. 4 is a structure diagram of the dense sub-block in the embodiment; Figure 5 FIG. 5 is a structure diagram of the residual sub-block in the embodiment; Figure 6 FIG. 6 is a structure diagram of the channel fusion weighting module in the embodiment; Figure 7 FIG. 7 is a structure diagram of the large-core attention block in the embodiment. DETAILED DESCRIPTION
[0007] The specific embodiments of the application are described below to facilitate the understanding of the application by those skilled in the art, but it should be clear that the application is not limited to the scope of the specific embodiments. For those skilled in the art, it is obvious that various changes are within the spirit and scope of the application as defined in the appended claims, and all the applications utilizing the concept of the application are included in the protection.
[0008] As shown in FIG. 1, a sparse CT reconstruction method fusing channel weighting and residual dense network comprises the following steps S1-S5: Figure 1 S1, filtering back projection and re-projection operation are performed on the original sparse sinogram to generate a full-size re-projected sinogram.
[0009] In this embodiment, the original sparse sinogram is subjected to filter back projection operation to generate a CT image containing stripe artifacts. In order to utilize the structural information in the image domain to compensate for the missing projection data, the CT image containing stripe artifacts is subjected to full scan angle range reprojection operation to generate a reference projection same size sinogram, i.e. a full size reprojection sinogram. Therefore, the generated full size reprojection sinogram is input into the residual dense network based on channel fusion for training, so as to solve the problem of sparse CT image reconstruction and improve the quality of the reconstructed image.
[0010] S2, constructing a residual dense network based on channel fusion.
[0011] In this embodiment, the residual dense network based on channel fusion is a deep neural network introducing dense connection, residual network, channel fusion weighting and large kernel attention mechanism. The network extracts structural features through a multi-layer encoder and performs deep semantic modeling, and then combines the attention mechanism to adaptively adjust the feature fusion proportion, so as to output a high-quality sinogram close to the reference sinogram, i.e. a recovered sinogram. Therefore, the network is used for image reconstruction, which effectively compensates for the missing or inaccurate projection information in the original sparse projection.
[0012] The structure and connection relationship of the constructed residual dense network based on channel fusion are as follows: Figure 2As shown, the residual dense network based on channel fusion includes a first convolutional normalization activation layer, a second convolutional normalization activation layer, a first down-sampling layer, a first residual dense channel fusion block, a second down-sampling layer, a second residual dense channel fusion block, a third down-sampling layer, a third residual dense channel fusion block, a fourth down-sampling layer, a fourth residual dense channel fusion block, a first deconvolutional layer, a third convolutional normalization activation layer, a fourth convolutional normalization activation layer, a second deconvolutional layer, a fifth convolutional normalization activation layer, a sixth convolutional normalization activation layer, a third deconvolutional layer, a seventh convolutional normalization activation layer, an eighth convolutional normalization activation layer, a fourth deconvolutional layer, a ninth convolutional normalization activation layer, a tenth convolutional normalization activation layer, an eleventh convolutional normalization activation layer, a first large kernel attention block, a second large kernel attention block, a third large kernel attention block, and a fourth large kernel attention block; the first deconvolutional layer, the first large kernel attention block, and the third convolutional normalization activation layer are connected in a channel to input the feature map output by the first large kernel attention block and the feature map output by the first deconvolutional layer into the third convolutional normalization activation layer through channel connection; the second deconvolutional layer, the second large kernel attention block, and the fifth convolutional normalization activation layer are connected in a channel to input the feature map output by the second large kernel attention block and the feature map output by the third deconvolutional layer into the fifth convolutional normalization activation layer through channel connection; the third deconvolutional layer, the third large kernel attention block, and the seventh convolutional normalization activation function layer are connected in a channel to input the feature map output by the third large kernel attention block and the feature map output by the third deconvolutional layer into the seventh convolutional normalization activation layer through channel connection; and the fourth deconvolutional layer, the fourth large kernel attention block, and the ninth convolutional normalization activation layer are connected in a channel to input the feature map output by the fourth large kernel attention block and the feature map output by the fourth deconvolutional layer into the ninth convolutional normalization activation layer through channel connection.
[0013] As can be seen from Figure 2 It can also be seen from the above that the residual dense network based on channel fusion as a whole belongs to an encoder-decoder structure, which includes four encodings and four decodings, and in each layer of the encoder, the convolution of the conventional ordinary U-Net network is optimized into a residual dense channel fusion block to enhance the ability of extracting features and improve the quality of reconstructed images; meanwhile, in each layer of the decoder, instead of directly connecting the encoder output features and the up-sampled feature maps in a jump connection, the features of each stage of the encoder are extracted in a deeper semantic feature extraction through a large kernel attention block and then connected in a jump connection (residual structure), so as to enhance the structural modeling capability in the jump connection, introduce a non-local information perception mechanism, and improve the fusion quality of the jump connection and the main semantic; and in the encoding-decoding process, the error between the re-projection and the reference projection is learned, that is, the artifact is learned; therefore, the residual dense network based on channel fusion is used for image reconstruction, and the quality of the reconstructed image is improved.
[0014] The structure and connection relationship of the residual dense channel fusion block are as shown in Figure 3 The residual dense channel fusion block comprises a first 1*1 convolution kernel, a dense sub-block, a residual sub-block, and a channel fusion weighting module. The residual sub-block comprises a twelfth convolution normalization activation layer, a convolution normalization layer, an activation layer, and a second 1*1 convolution kernel. The channel fusion weighting module comprises an average pooling layer, a fully connected linear rectification activation layer, and a fully connected S-type activation layer. The residual dense channel fusion block fuses the dense connection structure, the residual connection structure, and the 1*1 convolution preprocessing, thereby enhancing the ability of the network to extract features. The dense connection structure is the dense sub-block, and the structure and connection relationship thereof are as shown in Figure 4 Since each layer of the dense connection is connected with all the previous layers, it can effectively utilize historical information for feature propagation and reuse, and can promote feature reuse and gradient flow. The residual connection structure is the residual sub-block, and the structure and connection relationship thereof are as shown in Figure 5 The residual connection structure is the residual sub-block, and the structure and connection relationship thereof are as shown in Figure 6 The structure and connection relationship of the channel fusion weighting module are as shown in Figure 6 The dense sub-block and the residual sub-block are fused in the channel. By introducing the channel weighting fusion mechanism, the fusion ratio between the dense path and the residual path is dynamically adjusted according to the semantic response of the input features, so that the network can also flexibly focus on more useful feature branches under different input conditions. In addition, the dense sub-block has interlayer connection in the feature layer, which is suitable for extracting complex textures, and the residual block maintains structural continuity, so the complementary path fusion of the two helps to improve the network's ability to capture small structures and textures and the ability to preserve edges, contours and other mid-low frequency structures.
[0015] In addition, the structure and connection relationship of the large kernel attention block are as shown in Figure 7 Unlike the traditional U-Net network, the residual dense network based on channel fusion constructed by the present application performs channel splicing on the features extracted by the large kernel attention and the upsampled feature maps in the decoding stage. The large kernel attention block adopts a skip connection structure, and the advantages achieved are as follows: 1) Enhance the structure modeling capability in the skip connection; since the encoder output contains a large amount of low-level texture or local redundant information, by introducing the large kernel attention, the spatial structure perception modeling of the encoded features can be performed in advance, the meaningful edges, tissue boundaries or texture regions are strengthened, and the background or useless features are suppressed, so that the decoder focuses on the structural regions, thereby improving the structural integrity of the sinusogram; 2) Introduce a non-local information perception mechanism; large kernel attention blocks simulate a wider spatial area response by combining multi-scale large convolution, which is more suitable for target perception and cross-region symmetry and correlation modeling under large-scale artifact interference; therefore, the use of large kernel attention jump connection can more accurately provide non-local and semantically rich guiding features to the decoder; 3) Improve the fusion quality of the jump connection and the main semantic; the direct splicing between the traditional encoder output and the up-sampling feature often causes fusion conflicts due to inconsistent semantic levels; the large kernel attention block can improve the abstract degree of the shallow encoder feature in advance, alleviate this semantic difference, and make the feature after splicing have better uniformity and fusibility for the decoding convolution.
[0016] S3, input the reprojected full-size sinogram into the residual dense network based on channel fusion for training, introduce an improved joint loss function for optimization, learn the error between the restored sinogram and the reference sinogram, and the error between the restored sinogram filtered back-projection reconstructed image and the reference sinogram filtered back-projection reconstructed image, to generate the trained residual dense network based on channel fusion.
[0017] Specifically, step S3 specifically includes S301-S318: S301, input the reprojected full-size sinogram into the first convolution normalization activation layer and the second convolution normalization activation layer in sequence for feature extraction, generate the P1 feature map through convolution, normalization and activation operations.
[0018] S302, input the P1 feature map into the fourth large kernel attention block for depth separable convolution operation, depth separable dilated convolution operation and point convolution operation, and then multiply pixel by pixel with the P1 feature map through the residual connection structure to obtain the feature map of the fourth large kernel attention block.
[0019] S303, input the P1 feature map into the first down-sampling layer for down-sampling to generate the first down-sampling feature map, and input it into the first residual dense channel fusion block for feature extraction, generate the P2 feature map through channel weighting and residual dense operation, specifically: S3031, input the first down-sampling feature map into the first 1x1 convolution kernel for convolution operation to generate the first convolution feature map.
[0020] S3032, input the first down-sampling feature map into the dense sub-block for feature extraction, generate the feature map of the dense sub-block through batch normalization, activation and convolution operations.
[0021] S3033, input the first down-sampling feature map into the twelfth convolutional normalization activation layer of the residual sub-block to perform convolution, normalization and activation operations, then input the convolutional normalization layer to perform convolution and normalization operations, generate the feature map of the convolutional normalization layer, and through the residual connection structure, add the feature map of the convolutional normalization layer and the first down-sampling feature map pixel by pixel, then input the activation layer and the second 1*1 convolution kernel in sequence to perform activation and convolution operations, and generate the feature map of the residual sub-block.
[0022] S3034, input the feature map of the dense sub-block and the feature map of the residual sub-block into the channel fusion weighting module to perform feature extraction, and generate the first feature map and the second feature map of the channel fusion weighting module through channel fusion weighting, specifically as follows: First, add the feature map of the dense sub-block and the feature map of the residual sub-block pixel by pixel, then input the average pooling layer, the fully connected linear activation layer and the fully connected S-type activation layer in sequence to perform average pooling, fully connected linear activation and fully connected S-type activation, and generate the first weight and the second weight.
[0023] Then, through the residual connection structure, multiply the feature map of the dense sub-block and the first weight pixel by pixel to generate the first feature map of the channel fusion weighting module.
[0024] At the same time, through the residual connection structure, multiply the feature map of the residual sub-block and the second weight pixel by pixel to generate the second feature map of the channel fusion weighting module.
[0025] S3035, add the first convolutional feature map, the first feature map of the channel fusion weighting module and the second feature map of the channel fusion weighting module pixel by pixel to generate the P2 feature map.
[0026] S304, input the P2 feature map into the third large kernel attention block to perform depth separable convolution operation, depth separable dilated convolution operation and point convolution operation, then multiply the P2 feature map pixel by pixel through the residual connection structure to obtain the feature map of the third large kernel attention block.
[0027] S305, input the P2 feature map into the second down-sampling layer to perform down-sampling and generate the second down-sampling feature map, and input the second down-sampling feature map into the second residual dense channel fusion block to perform feature extraction, and generate the P3 feature map through channel weighting and residual dense operation.
[0028] S306, input the P3 feature map into the second large kernel attention block to perform depth separable convolution operation, depth separable dilated convolution operation and point convolution operation, then multiply the P3 feature map pixel by pixel through the residual connection structure to obtain the feature map of the second large kernel attention block.
[0029] S307, input the P3 feature map into a third down-sampling layer for down-sampling to generate a third down-sampled feature map, and input the third down-sampled feature map into a third residual dense channel fusion block for feature extraction, and generate a P4 feature map through channel weighting and residual dense operation.
[0030] S308, after the P4 feature map is input into the first large kernel attention block for depth separable convolution operation, depth separable atrous convolution operation and point convolution operation, the P4 feature map is multiplied with the P4 feature map pixel by pixel through a residual connection structure to obtain a feature map of the first large kernel attention block.
[0031] S309, input the P4 feature map into a fourth down-sampling layer for down-sampling to generate a fourth down-sampled feature map, and input the fourth down-sampled feature map into a fourth residual dense channel fusion block for feature extraction, and generate a P5 feature map through channel weighting and residual dense operation.
[0032] S310, input the P5 feature map into a first deconvolution layer for up-sampling operation to generate a P6 feature map.
[0033] S311, after the feature map of the first large kernel attention block is spliced with the P6 feature map in the channel, the feature map is sequentially input into a third convolution normalization activation layer, a fourth convolution normalization activation layer and a second deconvolution layer for feature extraction, and a P7 feature map is generated through convolution, normalization, activation operation and up-sampling operation.
[0034] S312, after the feature map of the second large kernel attention block is spliced with the P7 feature map in the channel, the feature map is sequentially input into a fifth convolution normalization activation layer, a sixth convolution normalization activation layer and a third deconvolution layer for feature extraction, and a P8 feature map is generated through convolution, normalization, activation operation and up-sampling operation.
[0035] S313, after the feature map of the third large kernel attention block is spliced with the P8 feature map in the channel, the feature map is sequentially input into a seventh convolution normalization activation layer, an eighth convolution normalization activation layer and a fourth deconvolution layer for feature extraction, and a P9 feature map is generated through convolution, normalization, activation operation and up-sampling operation.
[0036] S314, after the feature map of the fourth large kernel attention block is spliced with the P9 feature map in the channel, the feature map is sequentially input into a ninth convolution normalization activation layer, a tenth convolution normalization activation layer and an eleventh convolution normalization activation layer for feature extraction, and a P10 feature map is generated through convolution, normalization and activation operation.
[0037] S315, pixel-by-pixel sum the P10 feature map and the reprojected full-size sinogram to obtain an output sinogram.
[0038] In this embodiment, the network finally adds the reprojected full-size sinogram and the network output (output sinogram) pixel by pixel to form a global residual connection, so that the network as a whole has a stable backpropagation path, and is not prone to gradient vanishing or degradation problems when deep stacking, which helps to build a deeper encoder structure.
[0039] S316, the projection value of the corresponding projection angle in the output sinogram is replaced with the sparse sinogram to obtain a recovered sinogram.
[0040] S317, an improved joint loss function is calculated and used as a hybrid loss value of the recovered sinogram and the reference sinogram.
[0041] In this embodiment, in the process of deep learning modeling in the sparse CT projection domain, in order to improve the accuracy and perceptual quality of image reconstruction, the present application designs a multi-item joint loss function, i.e. an improved joint loss function, to simultaneously constrain the reconstruction effect of the projection domain and the image domain; the loss function combines traditional mean square error, structural similarity, edge feature, frequency domain consistency and perceptual quality indicators, and has stronger stability and generalization ability. Its formula is as follows:
[0042] wherein, L denotes the improved joint loss function value, denotes the recovered sinogram, denotes the reference sinogram, denotes the mean square error loss of the recovered sinogram and the reference sinogram, used to constrain the overall accuracy of the projection data, denotes the reconstructed image of the recovered sinogram filtered back projection, denotes the reconstructed image of the reference sinogram filtered back projection, denotes the multi-scale structural similarity loss of the reconstructed image and the reconstructed image , used to ensure the structural guarantee degree of the reconstructed image, denotes the edge loss of the reconstructed image and the reconstructed image , which calculates the gradient difference between the reconstructed image and the reconstructed image to measure the restoration ability of image details, denotes the frequency domain loss of the reconstructed image and the reconstructed image , denotes the perceptual loss of the reconstructed image and the reconstructed image , which is the learning perceptual image block similarity based on a pre-trained network to simulate human eye perceptual consistency, , , , , represents the weight hyperparameters for mean squared error, multi-scale structural similarity loss, edge loss, frequency domain loss, and perceptual loss, respectively.
[0043] The formula for calculating the mean squared error loss is:
[0044] in, This indicates the number of training samples in the current batch. Indicates the first The predicted projection value of each sample. Indicates the first The true projected value of each sample.
[0045] The formula for calculating multi-scale structural similarity loss is:
[0046]
[0047]
[0048]
[0049] in, This represents the total number of scales; in this invention, the value is 5. Indicates the first Brightness at each scale, , They represent the first Contrast and structure at various scales Indicates the first Weights of brightness at each scale , They represent the first Contrast at each scale, structural weights, , Representing the reconstructed images Reconstructing images The mean, , Representing the reconstructed images Reconstructing images standard deviation Represents the reconstructed image With reconstructed images covariance, Represents the reconstructed image With reconstructed images brightness, Represents the reconstructed image With reconstructed images Contrast Represents the reconstructed image With reconstructed images The structure, , , These represent distinct minimal numbers that are protected from division by zero.
[0050] In this embodiment, typically , , , , , This represents the maximum value of the image.
[0051] The formula for calculating edge loss is:
[0052] in, Indicates absolute error. , Representing the reconstructed images exist , Gradient of direction, , Representing the reconstructed images exist direction, Gradient of direction.
[0053] The formula for calculating frequency domain loss is:
[0054] in, Represents the reconstructed image Fourier transform, Represents the reconstructed image Fourier transform, This represents the amplitude spectrum after the Fourier transform of the reconstructed image.
[0055] The formula for calculating perceived loss is:
[0056] in, , These represent the first and second generations of the pre-trained network. The height and width of the layer feature map Indicates the pre-trained network's... The weights of the layer feature channels, This indicates multiplication by channel. Represents the reconstructed image In the first The channel is normalized in the layer, The reconstructed image is represented by In the first The channel is normalized in the layer, The Euclidean distance is represented by
[0057] S318, updating the network parameters of the residual dense network based on channel fusion by using the hybrid loss value, to obtain a trained residual dense network based on channel fusion.
[0058] S4, inputting the collected sparse sinogram into the trained residual dense network based on channel fusion for projection recovery, to generate a recovered sinogram.
[0059] S5, performing filtered back-projection on the recovered sinogram, to generate a CT image free of artifacts.
[0060] The principles and implementation manners of the present application are described by specific embodiments in the present application, and the above embodiment descriptions are only used to help understand the method of the present application and its core idea; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manners and application ranges can be changed, and the above descriptions should not be understood as limitations of the present application.
[0061] Those skilled in the art will realize that the embodiments described herein are for the purpose of understanding the principles of the present application and should be understood as not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations according to the technical inspirations disclosed in the present application without departing from the spirit of the present application, and these modifications and combinations are still within the scope of protection of the present application.
Claims
1. A sparse CT reconstruction method that integrates channel-weighted and residual dense networks, characterized in that, Includes the following steps: S1. Perform filtering, backprojection, and reprojection operations on the original sparse sine graph to generate a full-size reprojected sine graph. S2. Construct a residual dense network based on channel fusion, which includes a convolutional normalized activation layer, a downsampling layer, a first residual dense channel fusion block, a second residual dense channel fusion block, a third residual dense channel fusion block, a fourth residual dense channel fusion block, a large kernel attention block, and a deconvolutional layer. S3. Input the full-size sine wave reprojected into the channel fusion-based residual dense network for training. Introduce an improved joint loss function for optimization. By learning the errors between the recovered sine wave and the reference sine wave, and between the recovered sine wave filtered back-projection reconstructed image and the reference sine wave filtered back-projection reconstructed image, a trained channel fusion-based residual dense network is generated. S4. Input the collected sparse sinusoidal image into the trained channel fusion-based residual dense network for projection reconstruction to generate the reconstructed sinusoidal image. S5. Filter and back-project the recovered sine wave to generate an artifact-free CT image.
2. The sparse CT reconstruction method based on fused channel weighted and residual dense networks according to claim 1, characterized in that, The convolutional normalized activation layer includes the first convolutional normalized activation layer, the second convolutional normalized activation layer, the third convolutional normalized activation layer, the fourth convolutional normalized activation layer, the fifth convolutional normalized activation layer, the sixth convolutional normalized activation layer, the seventh convolutional normalized activation layer, the eighth convolutional normalized activation layer, the ninth convolutional normalized activation layer, the tenth convolutional normalized activation layer, and the eleventh convolutional normalized activation layer.
3. The sparse CT reconstruction method based on fused channel weighted and residual dense networks according to claim 1, characterized in that, The downsampling layers include a first downsampling layer, a second downsampling layer, a third downsampling layer, and a fourth downsampling layer.
4. The sparse CT reconstruction method based on fused channel weighted and residual dense networks according to claim 1, characterized in that, Large attention blocks include the first large attention block, the second large attention block, the third large attention block, and the fourth large attention block.
5. The sparse CT reconstruction method based on fused channel weighted and residual dense networks according to claim 1, characterized in that, The deconvolutional layer includes a first deconvolutional layer, a second deconvolutional layer, a third deconvolutional layer, and a fourth deconvolutional layer.
6. The sparse CT reconstruction method based on fused channel weighted and residual dense networks according to claim 1, characterized in that, The first residual dense channel fusion block, the second residual dense channel fusion block, and the third residual dense channel fusion block all include a first 1×1 convolutional kernel, a dense sub-block, a residual sub-block, and a channel fusion weighting module; The residual sub-block includes a twelfth convolutional normalized activation layer, a convolutional normalized layer, an activation layer, and a second 1×1 convolutional kernel; The channel fusion weighting module includes an average pooling layer, a fully connected linear rectified activation layer, and a fully connected sigmoid activation layer.
7. The sparse CT reconstruction method based on fused channel weighted and residual dense networks according to any one of claims 1 to 6, characterized in that, Step S3 specifically includes: S301. Input the full-size sine map of the reprojection into the first convolutional normalization activation layer and the second convolutional normalization activation layer in sequence for feature extraction. Through convolution, normalization and activation operations, the P1 feature map is generated. S302. Input the P1 feature map into the fourth large kernel attention block and perform depthwise separable convolution, depthwise separable dilated convolution, and point convolution operations. Then, multiply the P1 feature map pixel by pixel through the residual connection structure to obtain the feature map of the fourth large kernel attention block. S303. Input the P1 feature map into the first downsampling layer for downsampling to generate the first downsampled feature map, and input it into the first residual dense channel fusion block for feature extraction. Through channel weighting and residual dense operation, the P2 feature map is generated. S304. Input the P2 feature map into the third large kernel attention block and perform depthwise separable convolution, depthwise separable dilated convolution, and point convolution operations. Then, multiply the P2 feature map pixel by pixel through the residual connection structure to obtain the feature map of the third large kernel attention block. S305. Input the P2 feature map into the second downsampling layer for downsampling to generate the second downsampled feature map, and input it into the second residual dense channel fusion block for feature extraction. Through channel weighting and residual dense operation, the P3 feature map is generated. S306. Input the P3 feature map into the second large kernel attention block and perform depthwise separable convolution, depthwise separable dilated convolution, and point convolution operations. Then, multiply the P3 feature map pixel by pixel through the residual connection structure to obtain the feature map of the second large kernel attention block. S307. Input the P3 feature map into the third downsampling layer for downsampling to generate the third downsampled feature map, and input it into the third residual dense channel fusion block for feature extraction. Through channel weighting and residual dense operation, the P4 feature map is generated. S308. Input the P4 feature map into the first large kernel attention block and perform depthwise separable convolution, depthwise separable dilated convolution, and point convolution operations. Then, multiply the P4 feature map pixel by pixel through the residual connection structure to obtain the feature map of the first large kernel attention block. S309. Input the P4 feature map into the fourth downsampling layer for downsampling to generate the fourth downsampled feature map, and input it into the fourth residual dense channel fusion block for feature extraction. Through channel weighting and residual dense operation, the P5 feature map is generated. S310. Input the P5 feature map into the first deconvolutional layer for upsampling to generate the P6 feature map. S311. After concatenating the feature map of the first large kernel attention block with the P6 feature map on the channel, it is then sequentially input into the third convolutional normalization activation layer, the fourth convolutional normalization activation layer, and the second deconvolutional layer for feature extraction. Through convolution, normalization, activation operations, and upsampling operations, the P7 feature map is generated. S312. After concatenating the feature map of the second largest kernel attention block with the P7 feature map on the channel, it is then sequentially input into the fifth convolutional normalization activation layer, the sixth convolutional normalization activation layer, and the third deconvolutional layer for feature extraction. Through convolution, normalization, activation operations, and upsampling operations, the P8 feature map is generated. S313. After concatenating the feature map of the third large kernel attention block with the P8 feature map on the channel, it is then input into the seventh convolutional normalization activation layer, the eighth convolutional normalization activation layer, and the fourth deconvolutional layer in sequence for feature extraction. Through convolution, normalization, activation operations and upsampling operations, the P9 feature map is generated. S314. After concatenating the feature map of the fourth large kernel attention block with the P9 feature map on the channel, input them sequentially into the ninth convolutional normalized activation layer, the tenth convolutional normalized activation layer, and the eleventh convolutional normalized activation layer for feature extraction. Through convolution, normalization, and activation operations, the P10 feature map is generated. S315. Sum the P10 feature map and the reprojected full-size sine map pixel by pixel to obtain the output sine map; S316. Replace the projection value of the corresponding projection angle in the output sine graph with the sparse sine graph to obtain the restored sine graph; S317. Calculate the improved joint loss function and use it as the mixed loss value of the recovered sine curve and the reference sine curve; S318. Update the network parameters of the channel fusion-based residual dense network using the hybrid loss value to obtain the trained channel fusion-based residual dense network.
8. The sparse CT reconstruction method based on fused channel weighted and residual dense networks according to claim 7, characterized in that, Step S303 specifically includes: S3031. Input the first downsampled feature map into the first 1×1 convolution kernel to perform a convolution operation to generate the first convolution feature map; S3032. Input the first downsampled feature map into the dense sub-block for feature extraction, and generate the feature map of the dense sub-block through batch normalization, activation and convolution operations. S3033. The first downsampled feature map is input into the twelfth convolutional normalized activation layer of the residual sub-block for convolution, normalization, and activation operations. Then, it is input into the convolutional normalization layer for convolution and normalization operations to generate the feature map of the convolutional normalization layer. Through the residual connection structure, the feature map of the convolutional normalization layer is added to the first downsampled feature map pixel by pixel. Then, it is input into the activation layer and the second 1×1 convolutional kernel in sequence for activation and convolution operations to generate the feature map of the residual sub-block. S3034. Input the feature maps of dense sub-blocks and residual sub-blocks into the channel fusion weighting module for feature extraction. Through channel fusion weighting, generate the first feature map and the second feature map of the channel fusion weighting module. S3035. The first convolutional feature map, the first feature map of the channel fusion weighting module, and the second feature map of the channel fusion weighting module are added pixel by pixel to generate the P2 feature map.
9. The sparse CT reconstruction method based on fused channel weighted and residual dense networks according to claim 8, characterized in that, Step S3034 specifically includes: First, the feature maps of dense sub-blocks and residual sub-blocks are added pixel by pixel, and then sequentially input into an average pooling layer, a fully connected linear activation layer, and a fully connected sigmoid activation layer for average pooling, fully connected linear activation, and fully connected sigmoid activation to generate the first weight and the second weight. Then, through the residual connection structure, the feature map of the dense sub-block is multiplied pixel by pixel with the first weight to generate the first feature map of the channel fusion weighting module; Meanwhile, through the residual connection structure, the feature map of the residual sub-block is multiplied pixel by pixel with the second weight to generate the second feature map of the channel fusion weighting module.
10. The sparse CT reconstruction method based on fused channel weighted and residual dense networks according to claim 9, characterized in that, The formula for calculating the improved joint loss function is as follows: in, This represents the improved joint loss function value. This indicates the restoration of the sine curve. Indicates the reference sine curve. This represents the mean square error loss between the recovered sine curve and the reference sine curve. This represents the reconstructed image obtained by restoring the sinusoidal filter backprojection. This represents the reconstructed image obtained by backprojection using a reference sine wave filter. Represents the reconstructed image With reconstructed images Multi-scale structural similarity loss, Represents the reconstructed image With reconstructed images Edge loss, Represents the reconstructed image With reconstructed images Frequency domain loss, Represents the reconstructed image With reconstructed images Perceptual loss, , , , , Let represent the weight hyperparameters for mean squared error, multi-scale structural similarity loss, edge loss, frequency domain loss, and perceptual loss, respectively. This indicates the number of samples in the current training batch. Indicates the first The predicted projection value of each sample. Indicates the first The true projected value of each sample Indicates the total number of scales. Indicates the first Brightness at each scale, , They represent the first Contrast and structure at various scales Indicates the first Weights of brightness at each scale , They represent the first Contrast at each scale, structural weights, , Representing the reconstructed images Reconstructing images The mean, , Representing the reconstructed images Reconstructing images standard deviation Represents the reconstructed image With reconstructed images covariance, Represents the reconstructed image With reconstructed images brightness, Represents the reconstructed image With reconstructed images Contrast Represents the reconstructed image With reconstructed images The structure, , , These represent distinct minimal numbers that are prevented from being divided by zero. Indicates absolute error. , Representing the reconstructed images exist , Gradient of direction, , Representing the reconstructed images exist , Gradient of direction, Represents the reconstructed image Fourier transform, Represents the reconstructed image Fourier transform, This represents the amplitude spectrum after the Fourier transform of the reconstructed image. , These represent the first and second generations of the pre-trained network. The height and width of the layer feature map Indicates the pre-trained network's... The weights of the layer feature channels, This indicates multiplication by channel. Represents the reconstructed image In the pre-trained network The layer performs a normalization operation on the channel. Represents the reconstructed image In the pre-trained network The layer performs a normalization operation on the channel. This indicates Euclidean distance.