A sparse CT projection reconstruction method based on dense large kernel attention

Abstract

The invention relates to the technical field of image processing, and discloses a sparse CT projection reconstruction method based on dense large-core attention, comprising the following steps: interpolating a sparse sinusoidal graph by a cubic spline interpolation method to obtain an initial artifact sinusoidal graph; introducing a dense large-core attention block including a dense sub-block, a large-core attention sub-block and a convolutional layer to construct a dense large-core attention encoding and decoding network model; inputting the initial artifact sinusoidal graph and a non-sparse sinusoidal graph into the dense large-core attention encoding and decoding network model for training, learning the error between the artifact sinusoidal graph and the non-sparse sinusoidal graph, and obtaining a trained dense large-core attention encoding and decoding network model; reacquiring the sparse sinusoidal graph and inputting it into the trained dense large-core attention encoding and decoding network model, obtaining a repaired sinusoidal graph and reconstructing it by a filtered back-projection method, and obtaining a high-quality CT image; the method reduces the error between the output result and the label value, and improves the model performance and prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a sparse CT projection reconstruction method based on dense large kernel attention. Background Art

[0002] X-ray CT can obtain the internal information of an object without destroying the internal information of the object, so X-ray CT has been widely used in clinical, industrial and other fields. With the advancement of CT technology, the issue of radiation dose has attracted more and more attention. Reducing CT dose is to meet the principle of reasonable minimum dose. Sparse CT is one of the methods to reduce CT dose. Since sparse projection reconstruction is an ill-posed problem, when sparse projection reconstruction of CT images is performed by analytical method, there are serious streak artifacts. At the same time, the methods for solving sparse projection reconstruction in the prior art are divided into three categories: sinusoidal projection compensation method, iterative method, and deep learning-based method. Among them, the sinusoidal projection compensation method has the problem that the effect of restoring the projection is difficult to grasp, and even the error of the sinusoidal data causes new artifacts to appear in the reconstructed image; the iterative method has the problem that it needs to repeatedly project and back-project during the iteration process, resulting in time-consuming reconstruction, and the hyperparameters have a greater impact on the convergence and reconstruction effect of the iteration; in addition, the existing sparse projection reconstruction method based on deep learning ignores the correlation between projections when using the neural network model for prediction, causing part of the image information to be lost, resulting in low accuracy. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention provides a sparse CT projection reconstruction method based on dense large kernel attention, which is used to solve the problems of low image quality and large error in the existing sparse projection reconstruction method.

[0004] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0005] A sparse CT projection reconstruction method based on dense large kernel attention includes the following steps:

[0006] S1, using cubic spline interpolation method to interpolate the sparse sinusoidal graph to obtain the initial artifact sinusoidal graph;

[0007] S2. Introduce a dense large-core attention block containing dense sub-blocks, large-core attention sub-blocks and convolutional layers to build a dense large-core attention encoding and decoding network model;

[0008] S3, inputting the initial artifact sinusoidal graph and the non-sparse sinusoidal graph into the constructed dense large-core attention encoding and decoding network model for training, learning the error between the artifact sinusoidal graph and the non-sparse sinusoidal graph, and obtaining the trained dense large-core attention encoding and decoding network model;

[0009] S4, re-obtain the sparse sinusoidal graph and input it into the trained dense large-core attention encoder-decoder network model to obtain the repaired sinusoidal graph;

[0010] S5. Reconstruct the repaired sinusoidal image using a filtered back-projection method to obtain a high-quality CT image.

[0011] Furthermore, step S1 specifically includes:

[0012] The sparse sinusoidal graph is interpolated using the cubic spline interpolation method to obtain the initial artifact sinusoidal graph, namely:

[0013] g=Af+η

[0014] Where g represents the initial artifact sinusoidal graph, A represents the system matrix, f represents the sparse sinusoidal graph, and η represents the noise.

[0015] Furthermore, the dense large-core attention encoder-decoder network model includes an encoder and a decoder connected in sequence;

[0016] The encoder includes a first convolution activation normalization layer, a second convolution activation normalization layer, a first downsampling convolution layer, a first dense large core attention block, a second downsampling convolution layer, a second dense large core attention block, a third downsampling convolution layer, a third dense large core attention block, a fourth downsampling convolution layer, and a fourth dense large core attention block, which are connected in sequence;

[0017] The decoder includes a first deconvolution layer, a third convolution activation normalization layer, a second deconvolution layer, a fourth convolution activation normalization layer, a third deconvolution layer, a fifth convolution activation normalization layer, a fourth deconvolution layer, and a sixth convolution activation normalization layer, which are connected in sequence.

[0018] Further, the first dense large core attention block, the second dense large core attention block, the third dense large core attention block and the fourth dense large core attention block each include a four-layer dense sub-block, a large core attention sub-block and a 1×1 convolution layer;

[0019] The dense sub-block consists of four batch-normalized convolutional layers connected sequentially, and each batch-normalized convolutional layer has skip connections with non-adjacent batch-normalized convolutional layers.

[0020] Furthermore, step S3 specifically includes:

[0021] S31, using the initial artifact sinusoidal graph as raw data, using the non-sparse sinusoidal graph as label data, inputting the labeled raw data into the encoder of the constructed dense large core attention encoding and decoding network model for encoding operation, and obtaining a first feature map, an output feature map of the first dense large core attention block, an output feature map of the second dense large core attention block, an output feature map of the third dense large core attention block, and an output feature map of the fourth dense large core attention block respectively;

[0022] S32, inputting the first feature map, the output feature map of the first dense large core attention block, the output feature map of the second dense large core attention block, the output feature map of the third dense large core attention block, and the output feature map of the fourth dense large core attention block into the decoder for decoding operation to obtain the optimal output feature map;

[0023] S33. The optimal output feature map and the initial artifact sinusoidal map are summed pixel by pixel to obtain the output sinusoidal map, and the mixed loss value of the output sinusoidal map and the non-sparse sinusoidal map is calculated. The network parameters of the dense large-core attention encoder-decoder network model are updated using the mixed loss value to obtain a trained dense large-core attention encoder-decoder network model.

[0024] Furthermore, step S31 specifically includes:

[0025] S311, inputting the labeled original data into the first convolution activation normalization layer and the second convolution activation normalization layer of the constructed dense large core attention encoding and decoding network model in sequence to perform convolution activation normalization operations to obtain a first feature map;

[0026] S312, after inputting the first feature map into the first downsampling convolution layer for downsampling operation, the downsampled first feature map is input into four batch normalization activation convolution layers of the dense sub-block of the first dense large core attention block for batch normalization activation convolution operation to obtain the first feature map of the dense sub-block;

[0027] Each batch normalization activated convolution layer receives the feature maps of all previous batch normalization activated convolution layers as input, and outputs the features to all subsequent batch normalization activated convolution layers after the batch normalization activated convolution operation;

[0028] At the same time, the downsampled first feature map is input into the large core attention sub-block of the first dense large core attention block for depthwise separable convolution operation, depthwise separable hole convolution operation, and point convolution operation to obtain the first feature map of the large core attention sub-block; and the first feature map of the large core attention block is input into the 1×1 convolution layer for convolution operation to obtain the first feature map of the 1×1 convolution layer;

[0029] The first feature map of the dense sub-block is summed element by element with the first feature map of the 1×1 convolutional layer to obtain the output feature map of the first dense large-core attention block;

[0030] S313, inputting the output feature map of the first dense large core attention block into the second downsampling convolution layer for downsampling operation, and then inputting into the second dense large core attention block to obtain the output feature map of the second dense large core attention block;

[0031] S314, inputting the output feature map of the second dense large core attention block into the third downsampling convolution layer for downsampling operation, and then inputting into the third dense large core attention block to obtain the output feature map of the third dense large core attention block;

[0032] S315. Input the output feature map of the third dense large core attention block into the fourth downsampling convolution layer for downsampling operation and then input it into the fourth dense large core attention block to obtain the output feature map of the fourth dense large core attention block.

[0033] Further, step S32 specifically includes:

[0034] S321, inputting the output feature map of the fourth dense large core attention block into the first deconvolution layer for upsampling, and then splicing it with the output feature map of the third dense large core attention block through a channel to obtain a spliced ​​first output feature map;

[0035] S322, inputting the spliced ​​first output feature map into the third convolution activation normalization layer and the second deconvolution layer in sequence to perform convolution activation normalization operation and upsampling operation, and splicing it with the output feature map of the second dense large core attention block through a channel to obtain a spliced ​​second output feature map;

[0036] S323, inputting the spliced ​​second output feature map into the fourth convolution activation normalization layer and the third deconvolution layer in sequence to perform convolution activation normalization operation and upsampling operation, and splicing it with the output feature map of the first dense large core attention block through a channel to obtain a spliced ​​third output feature map;

[0037] S324, inputting the spliced ​​third output feature map into the fifth convolution activation normalization layer and the fourth deconvolution layer in sequence to perform convolution activation normalization operation and upsampling operation, and splicing it with the first feature map through a channel to obtain a spliced ​​fourth output feature map;

[0038] S325. Input the concatenated fourth output feature map into the sixth convolution activation normalization layer for convolution activation normalization operation to obtain an optimal output feature map.

[0039] Furthermore, the calculation formula of the mixed loss value of the output sinusoidal graph and the non-sparse sinusoidal graph in step S33 is:

[0040]

[0041] Where L represents the mixed loss value, Y represents the non-sparse sinusoidal graph, and f(X) ′ represents the output sinusoidal graph, represents the mean square error, α represents the weighting factor of the structural similarity error, and L MS-SSIM (Y,f(X) ′ ) represents the multi-scale structural similarity error between the non-sparse sinusoidal graph and the output sinusoidal graph.

[0042] The present invention has the following beneficial effects:

[0043] 1. The present invention proposes a sparse CT projection reconstruction method based on dense large-core attention. By introducing a dense large-core attention block including a dense connection block, a large-core attention block and a convolutional layer, a dense large-core attention codec network model is constructed. The dense sub-block can reduce the number of parameters of the entire model through parameter sharing and prevent gradient disappearance. The large-core attention sub-block can adaptively extract the features of each layer and pay attention to the global and local features of the artifact sinusoidal graph of the input model, thereby improving the generalization ability of the model, that is, improving the performance and prediction accuracy of the dense large-core attention codec network model.

[0044] 2. Local connections are added to the encoder and decoder of the dense large-core attention encoding and decoding network model, so that the gradient can flow directly to farther layers, thereby alleviating the gradient vanishing problem and accelerating the rapid convergence of the training process;

[0045] 3. Adding skip connections between batch normalized activated convolutional layers of dense sub-blocks of the decoder avoids the problem of gradient vanishing or gradient exploding, thereby fully reusing and preserving features, improving model performance and utilization of image features;

[0046] 4. During the model training process, the mixed loss value of the output sinusoidal graph and the non-sparse sinusoidal graph is calculated, and the network parameters of the dense large-core attention encoding and decoding network model are updated using the mixed loss value, which retains the correlation of the projection data of the sparse sinusoidal graph, reduces the error between the output result and the label value, and improves the model prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A flowchart of a sparse CT projection reconstruction method based on dense large kernel attention proposed in the present invention;

[0048] Figure 2 This is a schematic diagram of the structure of the dense large-core attention encoding and decoding network model;

[0049] Figure 3 Schematic diagram of the structure of the dense large core attention block;

[0050] Figure 4 It is a schematic diagram of the structure of dense sub-blocks;

[0051] Figure 5 Schematic diagram of the structure of the large core attention sub-block. DETAILED DESCRIPTION

[0052] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0053] like Figure 1 As shown, a sparse CT projection reconstruction method based on dense large kernel attention includes the following steps S1-S5:

[0054] S1. Use cubic spline interpolation to interpolate the sparse sinusoidal graph to obtain the initial artifact sinusoidal graph.

[0055] In this embodiment, the sparse sinusoidal graph is interpolated using the cubic spline interpolation method to obtain an initial artifact sinusoidal graph of the same size as the non-sparse sinusoidal graph (learning label graph), so that the initial artifact sinusoidal graph can be used in subsequent steps to train the dense large-core attention encoding and decoding network model.

[0056] Specifically, step S1 includes:

[0057] The sparse sinusoidal graph is interpolated using the cubic spline interpolation method to obtain the initial artifact sinusoidal graph, namely:

[0058] g=Af+η

[0059] Where g represents the initial artifact sinusoidal graph, A represents the system matrix, f represents the sparse sinusoidal graph, and η represents the noise.

[0060] S2. Introduce a dense large-core attention block consisting of a dense sub-block, a large-core attention sub-block and a convolutional layer to construct a dense large-core attention encoding and decoding network model.

[0061] like Figure 2 As shown, Figure 2 This is a schematic diagram of the structure of the dense large-core attention encoding and decoding network model; Figure 2The medium dense large core attention encoding and decoding network model includes an encoder and a decoder, and performs downsampling in the encoder. After each downsampling, a feature extraction of a dense large core attention block is performed, that is, the low-level features of the input image are extracted in the low-level dense large core attention block of the encoder, and higher-level features are extracted in the high-level dense large core attention block; while in the decoder, an upsampling operation is performed, and the output features of the corresponding modules in the encoder are connected, thereby enhancing the representation ability and performance of the model; the specific structures of the encoder and the decoder are as follows:

[0062] Specifically, the dense large-core attention encoder-decoder network model includes an encoder and a decoder connected in sequence;

[0063] The encoder includes a first convolution activation normalization layer, a second convolution activation normalization layer, a first downsampling convolution layer, a first dense large core attention block, a second downsampling convolution layer, a second dense large core attention block, a third downsampling convolution layer, a third dense large core attention block, a fourth downsampling convolution layer, and a fourth dense large core attention block, which are connected in sequence.

[0064] The decoder includes a first deconvolution layer, a third convolution activation normalization layer, a second deconvolution layer, a fourth convolution activation normalization layer, a third deconvolution layer, a fifth convolution activation normalization layer, a fourth deconvolution layer, and a sixth convolution activation normalization layer, which are connected in sequence.

[0065] In this embodiment, from Figure 2 It can also be seen that the second convolution activation normalization layer, the first dense large core attention block, the second dense large core attention block, the third dense large core attention block, and the fourth dense large core attention block in the encoder have a local connection relationship with the decoder. Their functions are: (1) helping to restore the details of the sinusoidal graph; (2) enhancing the model's understanding of the sinusoidal graph; (3) reusing the feature map of the encoding stage in the decoding stage, thereby reducing the amount of computation and improving the efficiency of the model; (4) helping to improve the generalization ability of the model.

[0066] like Figure 3 As shown, Figure 3 This is a schematic diagram of the structure of a dense large core attention block. Figure 3 It can be seen that the dense large core attention block includes a dense sub-block, a large core attention sub-block and a 1×1 convolution layer. Among them, the 1×1 convolution layer is used for convolution operation.

[0067] The structure of dense sub-blocks is as follows Figure 4 As shown, Figure 4The dense sub-block includes four batch-normalized activated convolutional layers connected in sequence, and each batch-normalized activated convolutional layer is jump-connected to non-adjacent batch-normalized activated convolutional layers; each batch-normalized activated convolutional layer of the dense sub-block receives the feature maps of all previous layers as its input, and passes its output to all subsequent layers. This jump connection method enables the information to be fully reused and retained during the transmission process, thereby avoiding the problem of gradient disappearance. At the same time, each layer is connected to all previous layers, and historical information can be effectively used for feature propagation and reuse, thereby improving the utilization of image features and the performance of the model.

[0068] The structure of the large core attention sub-block is as follows Figure 5 As shown, from Figure 5 It can be seen that the large core attention sub-block uses the large core attention mechanism to perform depthwise separable convolution, depthwise separable hole convolution, and point convolution in sequence; among them, the depthwise separable convolution is to use an independent convolution kernel to convolve each channel of the input, which can extract spatial features in the input while keeping the number of parameters small; the depthwise separable hole convolution is a variant of the depthwise separable convolution, that is, the hole rate is set in the ordinary convolution kernel. By using the depthwise separable hole convolution, the receptive field can be increased while maintaining the number of parameters, and the long-distance spatial relationship can be better captured; the point convolution is to perform a convolution operation on the input in the channel direction to achieve cross-channel information fusion; after the depthwise separable convolution operation and the depthwise separable hole convolution operation, the point convolution can ensure that the information between different channels is exchanged and integrated, thereby enhancing the expression ability of the model; therefore, the above operation enables the large core attention sub-block to cover a larger area of ​​the image through the large core convolution, thereby capturing the relationship between pixels at a longer distance, and better understanding the contextual relationship between different areas in the image, and extracting features of different scales. At the same time, it can also extract more complex features through nonlinear activation, increase the model's understanding of image content, reduce the impact of noise to a certain extent, and extract more robust features. At the same time, the large-core attention mechanism introduced can adaptively adjust its focus area according to the input data, so that the model can flexibly process different image contents.

[0069] Therefore, combining dense sub-blocks, large core attention sub-blocks, and 1×1 convolutional layers can ensure feature reuse and capture a wider range of contextual information, providing richer feature representations. The long-distance dependency of large core attention is very useful for understanding complex image structures, and the densely connected multi-level feature reuse further enhances this. The combination of the three can achieve more effective feature integration and improve the model’s understanding of image details and global structures.

[0070] Specifically, the first dense large core attention block, the second dense large core attention block, the third dense large core attention block and the fourth dense large core attention block each include a four-layer dense sub-block, a large core attention sub-block and a 1×1 convolution layer; the dense sub-block includes four batch normalization activated convolution layers connected in sequence, and each batch normalization activated convolution layer is jump-connected to non-adjacent batch normalization activated convolution layers.

[0071] S3. The dense large-core attention encoding-decoding network model constructed by inputting the initial artifact sinusoidal graph and the non-sparse sinusoidal graph is trained to learn the error between the artifact sinusoidal graph and the non-sparse sinusoidal graph, and obtain the trained dense large-core attention encoding-decoding network model.

[0072] Specifically, step S3 specifically includes S31-S33:

[0073] S31. Use the initial artifact sinusoidal graph as the original data and the non-sparse sinusoidal graph as the label data, input the labeled original data into the encoder of the constructed dense large core attention encoding and decoding network model for encoding operation, and obtain the first feature map, the output feature map of the first dense large core attention block, the output feature map of the second dense large core attention block, the output feature map of the third dense large core attention block, and the output feature map of the fourth dense large core attention block respectively.

[0074] Specifically, step S31 specifically includes S311-S315:

[0075] S311. Input the labeled original data into the first convolution activation normalization layer and the second convolution activation normalization layer of the constructed dense large-core attention encoding and decoding network model in sequence to perform convolution activation normalization operations to obtain a first feature map.

[0076] S312. After the first feature map is input into the first downsampling convolution layer for downsampling operation, the downsampled first feature map is input into the four batch normalization activation convolution layers of the dense sub-block of the first dense large core attention block for batch normalization activation convolution operation to obtain the first feature map of the dense sub-block.

[0077] Each batch normalized activated convolutional layer receives the feature maps of all previous batch normalized activated convolutional layers as input, and outputs features to all subsequent batch normalized activated convolutional layers after the batch normalized activated convolution operation.

[0078] At the same time, the downsampled first feature map is input into the large core attention block of the first dense large core attention block for depthwise separable convolution operation, depthwise separable hole convolution operation, and point convolution operation to obtain the first feature map of the large core attention block; and the first feature map of the large core attention block is input into the 1×1 convolution layer for convolution operation to obtain the first feature map of the 1×1 convolution layer.

[0079] The first feature map of the dense sub-block is summed element-wise with the first feature map of the 1×1 convolutional layer to obtain the output feature map of the first dense large-core attention block.

[0080] S313, input the output feature map of the first dense large core attention block into the second downsampling convolution layer for downsampling operation, and then input it into the second dense large core attention block to obtain the output feature map of the second dense large core attention block.

[0081] S314. Input the output feature map of the second dense large core attention block into the third downsampling convolution layer for downsampling operation and then input it into the third dense large core attention block to obtain the output feature map of the third dense large core attention block.

[0082] S315. Input the output feature map of the third dense large core attention block into the fourth downsampling convolution layer for downsampling operation and then input it into the fourth dense large core attention block to obtain the output feature map of the fourth dense large core attention block.

[0083] S32. Input the first feature map, the output feature map of the first dense large core attention block, the output feature map of the second dense large core attention block, the output feature map of the third dense large core attention block, and the output feature map of the fourth dense large core attention block into the decoder for decoding operation to obtain the optimal output feature map.

[0084] Specifically, step S32 specifically includes S321-S325:

[0085] S321. Input the output feature map of the fourth dense large core attention block into the first deconvolution layer for upsampling, and then splice it with the output feature map of the third dense large core attention block through a channel to obtain a spliced ​​first output feature map.

[0086] S322. The spliced ​​first output feature map is sequentially input into the third convolution activation normalization layer and the second deconvolution layer for convolution activation normalization operation and upsampling operation, and is spliced ​​with the output feature map of the second dense large core attention block through a channel to obtain a spliced ​​second output feature map.

[0087] S323. Input the spliced ​​second output feature map into the fourth convolution activation normalization layer and the third deconvolution layer in sequence for convolution activation normalization operation and upsampling operation, and splice it with the output feature map of the first dense large core attention block through a channel to obtain a spliced ​​third output feature map.

[0088] S324. Input the spliced ​​third output feature map into the fifth convolution activation normalization layer and the fourth deconvolution layer in sequence to perform convolution activation normalization operation and upsampling operation, and splice it with the first feature map through a channel to obtain a spliced ​​fourth output feature map.

[0089] S325. Input the concatenated fourth output feature map into the sixth convolution activation normalization layer for convolution activation normalization operation to obtain an optimal output feature map.

[0090] S33. The optimal output feature map and the initial artifact sinusoidal map are summed pixel by pixel to obtain the output sinusoidal map, and the mixed loss value of the output sinusoidal map and the non-sparse sinusoidal map is calculated. The network parameters of the dense large-core attention encoder-decoder network model are updated using the mixed loss value to obtain a trained dense large-core attention encoder-decoder network model.

[0091] Specifically, the calculation formula for the mixed loss value of the output sinusoidal graph and the non-sparse sinusoidal graph in step S33 is:

[0092]

[0093] Where L represents the mixed loss value, Y represents the non-sparse sinusoidal graph, and f(X) ′ represents the output sinusoidal graph, represents the mean square error, α represents the weighting factor of the structural similarity error, and L MS-SSIM (Y,f(X) ′ ) represents the multi-scale structural similarity error between the non-sparse sinusoidal graph and the output sinusoidal graph.

[0094] In this embodiment, α is set to 0.1 in the experiment.

[0095] S4. Re-acquire the sparse sinusoidal graph and input it into the trained dense large-core attention encoder-decoder network model to obtain the repaired sinusoidal graph.

[0096] In this embodiment, the sparse sinusoidal graph is input into the trained dense large-core attention encoding and decoding network model, so that the projection of the sparse sinusoidal graph can be repaired to obtain a repaired sinusoidal graph.

[0097] S5. Reconstruct the repaired sinusoidal image using a filtered back-projection method to obtain a high-quality CT image.

[0098] In summary, the present invention proposes a sparse CT projection reconstruction method based on dense large core attention, which constructs a dense large core attention codec network model by introducing a dense large core attention block including a dense connection block, a large core attention block and a convolutional layer, wherein the dense sub-block can reduce the parameter amount of the entire model through parameter sharing and prevent the gradient from disappearing, and the large core attention sub-block can adaptively extract the features of each layer and pay attention to the global and local features of the artifact sinusoidal graph of the input model, thereby improving the generalization ability of the model, that is, improving the performance and prediction accuracy of the dense large core attention codec network model; at the same time, in the encoder and decoder of the dense large core attention codec network model, Local connections are also added so that the gradient can flow directly to farther layers, thereby alleviating the gradient vanishing problem and accelerating the rapid convergence of model training; secondly, jump connections are added between the batch normalization activated convolutional layers of the dense sub-blocks of the decoder to avoid the gradient vanishing or gradient exploding problems, thereby fully reusing and retaining features and improving model performance and feature utilization; finally, during the model training process, the mixed loss value of the output sinusoidal graph and the non-sparse sinusoidal graph is calculated, and the network parameters of the dense large-core attention encoder-decoder network model are updated using the mixed loss value, thereby retaining the correlation of the projection data of the sparse sinusoidal graph, reducing the error between the output result and the label value, and improving the model prediction accuracy.

[0099] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0100] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A sparse CT projection reconstruction method based on dense large kernel attention, characterized in that: The following steps are involved: S1, using cubic spline interpolation method to interpolate the sparse sinusoidal graph to obtain the initial artifact sinusoidal graph; S2. Introduce a dense large-core attention block containing dense sub-blocks, large-core attention sub-blocks and convolutional layers to build a dense large-core attention encoding and decoding network model; Among them, the dense large-core attention encoding and decoding network model includes an encoder and a decoder connected in sequence; The encoder includes a first convolution activation normalization layer, a second convolution activation normalization layer, a first downsampling convolution layer, a first dense large core attention block, a second downsampling convolution layer, a second dense large core attention block, a third downsampling convolution layer, a third dense large core attention block, a fourth downsampling convolution layer, and a fourth dense large core attention block, which are connected in sequence; The decoder includes a first deconvolution layer, a third convolution activation normalization layer, a second deconvolution layer, a fourth convolution activation normalization layer, a third deconvolution layer, a fifth convolution activation normalization layer, a fourth deconvolution layer, and a sixth convolution activation normalization layer connected in sequence; S3, inputting the initial artifact sinusoidal graph and the non-sparse sinusoidal graph into the constructed dense large-core attention encoding and decoding network model for training, learning the error between the artifact sinusoidal graph and the non-sparse sinusoidal graph, and obtaining the trained dense large-core attention encoding and decoding network model; Wherein, step S3 specifically includes: S31, using the initial artifact sinusoidal graph as raw data, using the non-sparse sinusoidal graph as label data, inputting the labeled raw data into the encoder of the constructed dense large core attention encoding and decoding network model for encoding operation, and obtaining a first feature map, an output feature map of the first dense large core attention block, an output feature map of the second dense large core attention block, an output feature map of the third dense large core attention block, and an output feature map of the fourth dense large core attention block respectively; S32, inputting the first feature map, the output feature map of the first dense large core attention block, the output feature map of the second dense large core attention block, the output feature map of the third dense large core attention block, and the output feature map of the fourth dense large core attention block into the decoder for decoding operation to obtain the optimal output feature map; S33, summing the optimal output feature map and the initial artifact sinusoidal map pixel by pixel to obtain the output sinusoidal map, and calculating the mixed loss value of the output sinusoidal map and the non-sparse sinusoidal map, and using the mixed loss value to update the network parameters of the dense large-core attention codec network model to obtain a trained dense large-core attention codec network model; S4, re-obtain the sparse sinusoidal graph and input it into the trained dense large-core attention encoder-decoder network model to obtain the repaired sinusoidal graph; S5. Reconstruct the repaired sinusoidal image using a filtered back-projection method to obtain a high-quality CT image.

2. The sparse CT projection reconstruction method based on dense large kernel attention according to claim 1, characterized in that: Step S1 specifically includes: The sparse sinusoidal graph is interpolated using the cubic spline interpolation method to obtain the initial artifact sinusoidal graph, namely: g=Af+η Where g represents the initial artifact sinusoidal graph, A represents the system matrix, f represents the sparse sinusoidal graph, and η represents the noise.

3. The sparse CT projection reconstruction method based on dense large kernel attention according to claim 2, characterized in that: The first dense large core attention block, the second dense large core attention block, the third dense large core attention block, and the fourth dense large core attention block each include a four-layer dense sub-block, a large core attention sub-block, and a 1×1 convolutional layer; The dense sub-block consists of four batch-normalized convolutional layers connected sequentially, and each batch-normalized convolutional layer has skip connections with non-adjacent batch-normalized convolutional layers.

4. The sparse CT projection reconstruction method based on dense large kernel attention according to claim 3, characterized in that: Step S31 specifically includes: S311, inputting the labeled original data into the first convolution activation normalization layer and the second convolution activation normalization layer of the constructed dense large core attention encoding and decoding network model in sequence to perform convolution activation normalization operations to obtain a first feature map; S312, after inputting the first feature map into the first downsampling convolution layer for downsampling operation, the downsampled first feature map is input into four batch normalization activation convolution layers of the dense sub-block of the first dense large core attention block for batch normalization activation convolution operation to obtain the first feature map of the dense sub-block; Each batch normalization activated convolution layer receives the feature maps of all previous batch normalization activated convolution layers as input, and outputs the features to all subsequent batch normalization activated convolution layers after the batch normalization activated convolution operation; At the same time, the downsampled first feature map is input into the large core attention sub-block of the first dense large core attention block for depthwise separable convolution operation, depthwise separable hole convolution operation, and point convolution operation to obtain the first feature map of the large core attention sub-block; and the first feature map of the large core attention block is input into the 1×1 convolution layer for convolution operation to obtain the first feature map of the 1×1 convolution layer; The first feature map of the dense sub-block is summed element by element with the first feature map of the 1×1 convolutional layer to obtain the output feature map of the first dense large-core attention block; S313, inputting the output feature map of the first dense large core attention block into the second downsampling convolution layer for downsampling operation, and then inputting into the second dense large core attention block to obtain the output feature map of the second dense large core attention block; S314, inputting the output feature map of the second dense large core attention block into the third downsampling convolution layer for downsampling operation, and then inputting into the third dense large core attention block to obtain the output feature map of the third dense large core attention block; S315. Input the output feature map of the third dense large core attention block into the fourth downsampling convolution layer for downsampling operation and then input it into the fourth dense large core attention block to obtain the output feature map of the fourth dense large core attention block.

5. The sparse CT projection reconstruction method based on dense large kernel attention according to claim 4, characterized in that: Step S32 specifically includes: S321, inputting the output feature map of the fourth dense large core attention block into the first deconvolution layer for upsampling, and then splicing it with the output feature map of the third dense large core attention block through a channel to obtain a spliced ​​first output feature map; S322, inputting the spliced ​​first output feature map into the third convolution activation normalization layer and the second deconvolution layer in sequence to perform convolution activation normalization operation and upsampling operation, and splicing it with the output feature map of the second dense large core attention block through a channel to obtain a spliced ​​second output feature map; S323, inputting the spliced ​​second output feature map into the fourth convolution activation normalization layer and the third deconvolution layer in sequence to perform convolution activation normalization operation and upsampling operation, and splicing it with the output feature map of the first dense large core attention block through a channel to obtain a spliced ​​third output feature map; S324, inputting the spliced ​​third output feature map into the fifth convolution activation normalization layer and the fourth deconvolution layer in sequence to perform convolution activation normalization operation and upsampling operation, and splicing it with the first feature map through a channel to obtain a spliced ​​fourth output feature map; S325. Input the concatenated fourth output feature map into the sixth convolution activation normalization layer for convolution activation normalization operation to obtain an optimal output feature map.

6. The sparse CT projection reconstruction method based on dense large kernel attention according to claim 5, characterized in that: The calculation formula of the mixed loss value of the output sinusoidal graph and the non-sparse sinusoidal graph in step S33 is: Where L represents the mixed loss value, Y represents the non-sparse sinusoidal graph, f(X)′ represents the output sinusoidal graph, represents the mean square error, α represents the weighting factor of the structural similarity error, and L MS-SSIM (Y,f(X)′) represents the multi-scale structural similarity error between the non-sparse sinusoidal graph and the output sinusoidal graph.

Images