A method, system, device and medium for generating CTA images based on CT images

Abstract

The present invention relates to the technical field of two-dimensional image generation, and specifically to a method, system, device and medium for generating a CTA image based on a CT image. In order to solve the technical problem of low synthesis quality of CTA images in the prior art, the present invention constructs a bidirectional synthesis path for forming a CTA synthesis image based on a CT image and generating a CT image based on a CTA image, as well as a bidirectional circulation path for forming a CT simulation image based on a CT image and generating a CTA simulation image based on a CTA image. By controlling the sum of adversarial losses, the sum of negative sample contrast losses and the sum of circulation losses in the bidirectional synthesis path, the morphological difference between the CT image and the CTA image is reduced, so that the CT image can capture the vascular structure and texture details in the CTA image during the synthesis process, thereby improving the synthesis quality. Finally, the path result of generating a CTA simulation image from a CT image in a trained path is extracted to obtain a high-quality CTA synthesis image.

Description

Technical Field

[0001] The present invention relates to the technical field of two-dimensional image generation, and in particular to a method, system, device and medium for generating a CTA image based on a CT image. Background Art

[0002] CTA images, or CT Angiography images, are used to analyze vascular lesions. This type of image can clearly show the morphology of blood vessels, the degree of lumen stenosis, the presence or absence of vascular malformations, aneurysms and other lesions, and is of great value for the diagnosis, evaluation and formulation of treatment plans for vascular diseases. However, when obtaining CTA images of patients, it is difficult to obtain accurate CTA images due to limitations such as heart rate and respiratory control, contrast agent allergies, and renal insufficiency, resulting in high costs and unstable acquisition of CTA images.

[0003] In order to solve the problems of high cost and unstable acquisition of CTA images, existing technologies often synthesize CTA images by processing the original CT scan data with mathematical algorithms, such as filtered back projection algorithm, iterative reconstruction algorithm, etc. However, these methods are sensitive to noise and are prone to image noise in low-dose scanning or large-sized patients, which affects image quality. Moreover, traditional image reconstruction algorithms are relatively weak in the ability to display vascular details and detect lesions, and are prone to miss some lesions, resulting in low synthesis quality of CTA images. Summary of the invention

[0004] The object of the present invention is to provide a method, system, device and medium for generating a CTA image based on a CT image.

[0005] The technical solution of the present invention is as follows:

[0006] A method for generating a CTA image based on a CT image comprises the following operations:

[0007] S1, the standard CT image is processed by the first generator to obtain the first CTA simulation image; the first CTA simulation image is processed by the second generator to obtain the first CT simulation image; the standard CTA image is processed by the third generator to obtain the second CT simulation image; the second CT simulation image is processed by the fourth generator to obtain the second CTA simulation image;

[0008] The first CTA simulation image and the standard CTA image are processed by the first discriminator to obtain a first CTA simulation feature and a standard CTA feature; the second CT simulation image and the standard CT image are processed by the second discriminator to obtain a second CT simulation feature and a standard CT feature;

[0009] When the sum of the cycle loss of the standard CT image and the first CT simulation image and the cycle loss of the standard CTA image and the second CTA simulation image is less than the cycle loss threshold, and the sum of the adversarial loss of the first CTA simulation feature and the standard CTA feature and the adversarial loss of the second CT simulation feature and the standard CT feature is less than the adversarial loss threshold, and the sum of the negative sample contrast loss of the first CTA simulation feature and the standard CTA feature and the negative sample contrast loss of the second CT simulation feature and the standard CT feature is less than the contrast loss threshold, output the first generator, the second generator, the third generator, the fourth generator, the first discriminator and the second discriminator to obtain a trained synthesis model;

[0010] The processing process in the first generator, or / and the second generator, or / and the third generator, or / and the fourth generator is: the input is subjected to several feature convolution processes with the number of channels increasing successively to obtain feature channel enhancement features; the feature channel enhancement features are subjected to several feature residual processes to obtain fusion features; the fusion features are subjected to several nearest neighbor upsampling processes and feature convolution processes with the number of channels decreasing successively to obtain output;

[0011] S2. The CT image to be processed is processed by the trained synthesis model, and the output of the first generator is extracted to obtain the target CTA synthesis image.

[0012] The processing process of the first discriminator and / or the second discriminator in S1 is: the input is processed by several feature convolutions with increasing number of channels to obtain convolution features; the convolution features are processed by global average pooling and global maximum pooling respectively to obtain average pooling features and maximum pooling features; the average pooling features and the maximum pooling features are processed by full connection respectively, and then multiplied element by element with the convolution features, and the output is obtained by superposition processing.

[0013] The feature residual processing in S1 is specifically as follows: the feature channel enhanced feature is subjected to convolution, instance normalization and nonlinear processing to obtain an initial feature map; the initial feature is subjected to instance normalization and layer normalization after convolution processing to obtain an instance standard feature map and a layer normalized feature map; based on the pixel variance and pixel mean of the instance standard feature map and the initial feature, a first enhanced feature is obtained; based on the pixel variance and pixel mean of the layer normalized feature map and the initial feature, a second enhanced feature is obtained; the first enhanced feature and the second enhanced feature are subjected to feature balancing processing to obtain a balanced feature; the balanced feature is superimposed with the feature channel enhanced feature after nonlinear processing to obtain the current round of fused features; and so on, after the last feature residual operation is performed, the fused feature is obtained.

[0014] After the feature residual processing operation in S1, it also includes multi-scale feature extraction processing on the fused features to obtain multi-scale fused features for performing several nearest neighbor upsampling processes; the multi-scale feature extraction processing operation is specifically as follows: the fused features are superimposed after being processed by convolutions of different scales to obtain multi-scale convolution features; the fused features are processed by global average pooling, convolution and upsampling to obtain pooled sampling features; the multi-scale convolution features and the pooled sampling features are superimposed and convoluted to obtain multi-scale fused features.

[0015] The CT image or / and CTA image, before being processed by the generator or / and discriminator, includes cropping the CT image or / and CTA image so that the CT value of the image is within the corresponding numerical range, and then normalizing, cropping and normalizing the image to obtain a qualified CT image or / and a qualified CTA image, and executing the operations in the generator or / and discriminator.

[0016] The circulation loss in S1 is obtained by normalizing the standard CT image with the first CT simulation image, or the standard CTA image with the second CTA simulation image.

[0017] The negative sample contrast loss in S1 is obtained based on the cosine similarity between the first CTA simulation feature and the standard CTA feature, or the second CT simulation feature and the standard CT feature.

[0018] A system for generating a CTA image based on a CT image, used to implement the above-mentioned method for generating a CTA image based on a CT image, comprising:

[0019] A first generator, used for processing a standard CT image to obtain a first CTA simulation image;

[0020] A second generator is used to process the first CTA simulation image to obtain a first CT simulation image;

[0021] The third generator is used to process the standard CTA image to obtain a second CT simulation image;

[0022] A fourth generator, used for processing the second CT simulation image to obtain a second CTA simulation image;

[0023] A first discriminator is used to process the first CTA simulation image and the standard CTA image respectively to obtain a first CTA simulation feature and a standard CTA feature;

[0024] The second discriminator is used to process the second CT simulation image and the standard CT image respectively to obtain a second CT simulation feature and a standard CT feature;

[0025] Training the synthesis model, when the sum of the cycle loss of the standard CT image and the first CT simulation image and the cycle loss of the standard CTA image and the second CTA simulation image is less than the cycle loss threshold, and the sum of the adversarial loss of the first CTA simulation feature and the standard CTA feature and the adversarial loss of the second CT simulation feature and the standard CT feature is less than the adversarial loss threshold, and the sum of the negative sample contrast loss of the first CTA simulation feature and the standard CTA feature and the negative sample contrast loss of the second CT simulation feature and the standard CT feature is less than the contrast loss threshold, output the first generator, the second generator, the third generator, the fourth generator, the first discriminator and the second discriminator to obtain the training synthesis model; used to process the CT image to be processed through the training synthesis model, extract the output of the first generator, and obtain the target CTA synthesis image;

[0026] The processing process in the first generator, or / and the second generator, or / and the third generator, or / and the fourth generator is: the input is processed by several feature convolutions with the number of channels increasing successively to obtain feature channel enhanced features; the feature channel enhanced features are processed by several feature residuals to obtain fused features; the fused features are processed by several nearest neighbor upsampling and feature convolutions with the number of channels decreasing successively to obtain output.

[0027] A device for generating a CTA image based on a CT image comprises a processor and a memory, wherein the processor implements the above-mentioned method for generating a CTA image based on a CT image when executing a computer program stored in the memory.

[0028] A computer-readable storage medium is used to store a computer program, wherein when the computer program is executed by a processor, the method for generating a CTA image based on a CT image is implemented.

[0029] The beneficial effects of the present invention are:

[0030] The present invention provides a method for generating a CTA image based on a CT image, which constructs a bidirectional synthesis path for forming a CTA synthetic image based on the CT image and generating a CT image based on the CTA image, and reduces the morphological difference between the CT image and the CTA image by controlling the sum of adversarial losses and the sum of negative sample contrast losses in the bidirectional synthesis path, so that the CT image can capture the vascular structure and texture details in the CTA image during the synthesis process, thereby improving the synthesis quality; at the same time, a bidirectional circulation path for forming a CT simulation image based on the CT image and generating a CTA simulation image based on the CTA image is constructed, and by controlling the sum of circulation losses in the bidirectional circulation path, key information is prevented from being lost during the synthesis process, synthesis details are retained, and synthesis quality is further improved; finally, the path result for generating a CTA simulation image from a CT image in the trained path is extracted to obtain a high-quality CTA synthetic image. DETAILED DESCRIPTION

[0031] This embodiment provides a method for generating a CTA image based on a CT image, including the following operations:

[0032] S1, the standard CT image is processed by the first generator to obtain the first CTA simulation image; the first CTA simulation image is processed by the second generator to obtain the first CT simulation image; the standard CTA image is processed by the third generator to obtain the second CT simulation image; the second CT simulation image is processed by the fourth generator to obtain the second CTA simulation image;

[0033] The first CTA simulation image and the standard CTA image are processed by the first discriminator to obtain a first CTA simulation feature and a standard CTA feature; the second CT simulation image and the standard CT image are processed by the second discriminator to obtain a second CT simulation feature and a standard CT feature;

[0034] When the sum of the cycle loss of the standard CT image and the first CT simulation image and the cycle loss of the standard CTA image and the second CTA simulation image is less than the cycle loss threshold, and the sum of the adversarial loss of the first CTA simulation feature and the standard CTA feature and the adversarial loss of the second CT simulation feature and the standard CT feature is less than the adversarial loss threshold, and the sum of the negative sample contrast loss of the first CTA simulation feature and the standard CTA feature and the negative sample contrast loss of the second CT simulation feature and the standard CT feature is less than the contrast loss threshold, output the first generator, the second generator, the third generator, the fourth generator, the first discriminator and the second discriminator to obtain a trained synthesis model;

[0035] The processing process in the first generator, or / and the second generator, or / and the third generator, or / and the fourth generator is: the input is subjected to several feature convolution processes with the number of channels increasing successively to obtain feature channel enhancement features; the feature channel enhancement features are subjected to several feature residual processes to obtain fusion features; the fusion features are subjected to several nearest neighbor upsampling processes and feature convolution processes with the number of channels decreasing successively to obtain output;

[0036] S2. The CT image to be processed is processed by the trained synthesis model, and the output of the first generator is extracted to obtain the target CTA synthesis image.

[0037] S1. The standard CT image is processed by the first generator to obtain the first CTA simulation image; the first CTA simulation image is processed by the second generator to obtain the first CT simulation image; the standard CTA image is processed by the third generator to obtain the second CT simulation image; the second CT simulation image is processed by the fourth generator to obtain the second CTA simulation image; the first CTA simulation image and the standard CTA image are processed by the first discriminator to obtain the first CTA simulation feature and the standard CTA feature; the second CT simulation image and the standard CT image are processed by the second discriminator to obtain the second CT simulation feature and the standard CT feature; when the standard CT image and the first CT When the sum of the cycle loss of the simulation image and the cycle losses of the standard CTA image and the second CTA simulation image is less than the cycle loss threshold, and the sum of the adversarial loss of the first CTA simulation feature and the standard CTA feature and the adversarial loss of the second CT simulation feature and the standard CT feature is less than the adversarial loss threshold, and the sum of the negative sample contrast loss of the first CTA simulation feature and the standard CTA feature and the negative sample contrast loss of the second CT simulation feature and the standard CT feature is less than the contrast loss threshold, the first generator, the second generator, the third generator, the fourth generator, the first discriminator and the second discriminator are output to obtain a trained synthetic model.

[0038] First, an initial synthesis model formed by a first generator, a second generator, a third generator, a fourth generator, a first discriminator and a second discriminator is constructed.

[0039] Then, a number of standard CT images and a number of corresponding standard CTA images are obtained to form a training set, and an initial synthesis model is trained.

[0040] In order to capture more useful features and reduce interference information, the CT image or / and CTA image, before being processed by the generator or / and discriminator, includes cropping the CT image or / and CTA image so that the CT value of the image reaches the corresponding value range, and then normalizing, cropping and normalizing to obtain a qualified CT image or / and a qualified CTA image, and performing operations in the generator or / and discriminator. Specifically, the CT image or / and CTA image in Dicom format is cropped respectively, so that the Hu value (CT value) of the CT image is limited to [-500,900], and the Hu value of the CTA image is limited to [-500,2000], and then the two images are normalized to [0,1] respectively, and the two images are cropped to 512×512 respectively, and finally normalized to [-1,1] respectively, to obtain a qualified CT image or / and a qualified CTA image as the input of the model.

[0041] During the training process, the processing operations of the initial synthesis model are as follows: the standard CT image is processed by the first generator to obtain the first CTA simulation image; the first CTA simulation image is processed by the second generator to obtain the first CT simulation image; the standard CTA image is processed by the third generator to obtain the second CT simulation image; the second CT simulation image is processed by the fourth generator to obtain the second CTA simulation image; the first CTA simulation image and the standard CTA image are respectively processed by the first discriminator to obtain the first CTA simulation feature and the standard CTA feature; the second CT simulation image and the standard CT image are respectively processed by the second discriminator to obtain the second CT simulation feature and the standard CT feature.

[0042] Among them, the processing process in the first generator, or / and the second generator, or / and the third generator, or / and the fourth generator is as follows.

[0043] In the first step, the input is processed by several feature convolutions with increasing number of channels, which enhances the feature expression ability and resolution adaptability of the input and obtains the feature channel enhancement feature. The operations of several feature convolutions include: one convolution with a convolution kernel of 3, instance normalization, nonlinear processing (which can be achieved through the ReLU activation function), and two convolutions with a convolution kernel of 7, instance normalization, and nonlinear processing. The number of channels in the process of several feature convolutions is 1, 32, and 64 respectively.

[0044] In the second step, the feature channel enhancement features are processed several times with feature residuals to obtain fused features, which helps the model network to learn the differences between CT images and CTA images more easily, thereby better synthesizing CTA images.

[0045] Taking the first feature residual processing as an example, the specific operation is as follows: the feature channel enhanced feature is subjected to convolution (through convolution processing with a convolution kernel of 3, instance normalization and nonlinear processing (which can be achieved through the ReLU activation function)), instance normalization and nonlinear processing to obtain the initial feature map; the initial feature is subjected to instance normalization and layer normalization after convolution processing to extract image shape and texture features to obtain the instance standard feature map and layer normalized feature map; based on the pixel variance and pixel mean of the instance standard feature map, as well as the initial feature, it can effectively remove the style information of the CT image itself, which is convenient for synthesis in the direction of the CTA image style, and obtain the first enhanced feature; based on the pixel variance and pixel mean of the layer normalized feature map, as well as the initial feature, the difference between the CT image and the CTA image is ignored, which is convenient for synthesis in the direction of the CTA image style, and obtain the second enhanced feature; the first enhanced feature and the second enhanced feature are subjected to feature balancing processing to obtain the balanced feature; the balanced feature is subjected to nonlinear processing (which can be achieved through the ReLU activation function) and then superimposed with the feature channel enhanced feature to obtain the current round of fused features; and so on, after the last feature residual operation is performed, the fused feature is obtained.

[0046] The above operation of obtaining the first enhanced feature can be implemented by the following formula:

[0047] ,

[0048] X 1 is the first enhanced feature, x is the initial feature, σ 1 is the pixel variance of the instance standard feature map, μ 1 is the pixel mean of the instance standard feature map, α is the compensation amount, and the spatial dimension of the instance standard feature map is [1, 2].

[0049] The above operation of obtaining the second enhanced feature can be implemented by the following formula:

[0050] ,

[0051] X 2 is the second enhanced feature, x is the initial feature, σ 2 is the pixel variance of the layer normalized feature map, μ 2 is the pixel mean of the layer normalized feature map, α To compensate, the spatial dimension of the layer normalized feature map is [1, 2].

[0052] The feature balance process can be achieved by the following formula: B=γ·(θ·X 1 +(1-θ)·X 2 )+β , B For the balance feature, γ is the scaling parameter, θ is the weight adjustment parameter, β is the offset parameter.

[0053] In order to improve the quality of CTA image synthesis, after several feature residual processing operations, the fused features are also subjected to multi-scale feature extraction processing to obtain multi-scale fused features for subsequent several nearest neighbor upsampling processes. Specifically, the fused features are superimposed after convolution processing at different scales to obtain multi-scale convolution features; the fused features are subjected to global average pooling, convolution and upsampling processing to obtain pooled sampling features; the multi-scale convolution features and the pooled sampling features are superimposed and convoluted to obtain multi-scale fused features, which are used to enhance the generator's ability to capture multi-scale information of the image, optimize the transmission of contextual information, and improve the texture detail vascular structure generated by the image.

[0054] In the third step, the fused features or multi-scale fused features are subjected to several nearest neighbor upsampling and feature convolution processes with decreasing channel numbers to obtain output. The several nearest neighbor upsampling and feature convolution processes specifically include: 2 nearest neighbor upsampling processes and convolution processes with a convolution kernel of 3, instance normalization, nonlinear processing (which can be achieved through the ReLU activation function), and 1 convolution process with a convolution kernel of 7 and nonlinear processing (which can be achieved through the tanh function). The number of channels is 128, 64, 64, 32, and 1, respectively.

[0055] In addition, the processing process of the first discriminator and / or the second discriminator is as follows: the input is subjected to several feature convolution processes with increasing numbers of channels to obtain convolution features; the convolution features are subjected to global average pooling and global maximum pooling respectively to obtain average pooling features and maximum pooling features; the average pooling features and the maximum pooling features are subjected to full connection processing respectively, and then element-wise multiplied with the convolution features, and output is obtained through superposition processing, which is used to extract important features of key areas and improve the image generation quality.

[0056] The operations of several feature convolution processing include: one convolution processing with a convolution kernel of 4 and nonlinear processing (which can be achieved through the ReLU activation function), and three convolution processing with a convolution kernel of 4, instance normalization, and nonlinear processing. The number of channels in the process of several feature convolution processing is 1, 64, 128, 256, and 512 respectively.

[0057] Finally, if during the training process, when the sum of the cycle loss of the standard CT image and the first CT simulation image and the cycle loss of the standard CTA image and the second CTA simulation image is less than the cycle loss threshold, and the sum of the adversarial loss of the first CTA simulation feature and the standard CTA feature and the adversarial loss of the second CT simulation feature and the standard CT feature is less than the adversarial loss threshold, and the sum of the negative sample contrast loss of the first CTA simulation feature and the standard CTA feature and the negative sample contrast loss of the second CT simulation feature and the standard CT feature is less than the contrast loss threshold, the first generator, the second generator, the third generator, the fourth generator, the first discriminator and the second discriminator are output to obtain a trained synthetic model.

[0058] The above-mentioned cycle loss is obtained by normalizing the standard CT image with the first CT simulation image, or the standard CTA image with the second CTA simulation image; the adversarial loss is obtained based on the expected value of the first CTA simulation feature and the standard CTA feature, or the second CT simulation feature and the standard CT feature; the negative sample contrast loss is obtained based on the cosine similarity of the first CTA simulation feature and the standard CTA feature, or the second CT simulation feature and the standard CT feature.

[0059] The total cycle loss is obtained by the following formula:

[0060] ,

[0061] L 1 is the total circulation loss, G(G(C)) is the first CT simulation image, C is a standard CT image. G(G(CA)) This is the second CTA simulation diagram. CA This is a standard CTA chart.

[0062] The sum of adversarial losses is obtained by the following formula:

[0063] ，

[0064] L 2 To combat the total loss, D(CA) is a standard CTA feature. D(G(C)) is the first CTA simulation feature, D(C) is the standard CT feature, D(G(CA)) It is the second CT simulation feature.

[0065] The sum of negative sample contrast loss is obtained by the following formula:

[0066] ,

[0067] L 3 is the sum of the negative sample contrast losses, D(G(C)) is the first CTA simulation feature, D(CA) is a standard CTA feature. D (G(CA)) is the second CT simulation feature, D(C) is the standard CT feature, sim( ) is the cosine similarity function.

[0068] S2. The CT image to be processed is processed by the trained synthesis model, and the output of the first generator is extracted to obtain the target CTA synthesis image.

[0069] The CT image to be processed is placed into a training synthesis model with high synthesis quality, good stability and high efficiency for processing, and the output of the first generator is extracted to obtain a high-quality target CTA synthesis image, which can improve the accuracy when used for subsequent lesion analysis.

[0070] This embodiment further provides a system for generating a CTA image based on a CT image, which is used to implement the above-mentioned method for generating a CTA image based on a CT image, including:

[0071] A first generator, used for processing a standard CT image to obtain a first CTA simulation image;

[0072] A second generator is used to process the first CTA simulation image to obtain a first CT simulation image;

[0073] The third generator is used to process the standard CTA image to obtain a second CT simulation image;

[0074] A fourth generator, used for processing the second CT simulation image to obtain a second CTA simulation image;

[0075] A first discriminator is used to process the first CTA simulation image and the standard CTA image respectively to obtain a first CTA simulation feature and a standard CTA feature;

[0076] The second discriminator is used to process the second CT simulation image and the standard CT image respectively to obtain a second CT simulation feature and a standard CT feature;

[0077] Training the synthesis model, when the sum of the cycle loss of the standard CT image and the first CT simulation image and the cycle loss of the standard CTA image and the second CTA simulation image is less than the cycle loss threshold, and the sum of the adversarial loss of the first CTA simulation feature and the standard CTA feature and the adversarial loss of the second CT simulation feature and the standard CT feature is less than the adversarial loss threshold, and the sum of the negative sample contrast loss of the first CTA simulation feature and the standard CTA feature and the negative sample contrast loss of the second CT simulation feature and the standard CT feature is less than the contrast loss threshold, output the first generator, the second generator, the third generator, the fourth generator, the first discriminator and the second discriminator to obtain the training synthesis model; used to process the CT image to be processed through the training synthesis model, extract the output of the first generator, and obtain the target CTA synthesis image;

[0078] The processing process in the first generator, or / and the second generator, or / and the third generator, or / and the fourth generator is: the input is processed by several feature convolutions with the number of channels increasing successively to obtain feature channel enhanced features; the feature channel enhanced features are processed by several feature residuals to obtain fused features; the fused features are processed by several nearest neighbor upsampling and feature convolutions with the number of channels decreasing successively to obtain output.

[0079] This embodiment further provides a device for generating a CTA image based on a CT image, comprising a processor and a memory, wherein the processor implements the above-mentioned method for generating a CTA image based on a CT image when executing a computer program stored in the memory.

[0080] This embodiment further provides a computer-readable storage medium for storing a computer program, wherein when the computer program is executed by a processor, the method for generating a CTA image based on a CT image is implemented.

[0081] The present embodiment provides a method for generating a CTA image based on a CT image, which constructs a bidirectional synthesis path for forming a CTA synthetic image based on the CT image and for generating a CT image based on the CTA image, and reduces the morphological difference between the CT image and the CTA image by controlling the sum of adversarial losses and the sum of negative sample contrast losses in the bidirectional synthesis path, so that the CT image can capture the vascular structure and texture details in the CTA image during the synthesis process, thereby improving the synthesis quality; at the same time, a bidirectional circulation path for forming a CT simulation image based on the CT image and for generating a CTA simulation image based on the CTA image is constructed, and by controlling the sum of circulation losses in the bidirectional circulation path, key information is prevented from being lost during the synthesis process, synthesis details are retained, and synthesis quality is further improved; finally, the path result for generating a CTA simulation image from a CT image in the trained path is extracted to obtain a high-quality CTA synthetic image.

Claims

1. A method for generating a CTA image based on a CT image, characterized in that: The following operations are included: S1, the standard CT image is processed by the first generator to obtain a first CTA simulation image; the first CTA simulation image is processed by the second generator to obtain a first CT simulation image; The standard CTA image is processed by the third generator to obtain a second CT simulation image; The second CT simulation image is processed by a fourth generator to obtain a second CTA simulation image; The first CTA simulation image and the standard CTA image are processed by a first discriminator to obtain a first CTA simulation feature and a standard CTA feature respectively; The second CT simulation image and the standard CT image are processed by a second discriminator to obtain a second CT simulation feature and a standard CT feature respectively; When the sum of the cycle loss of the standard CT image and the first CT simulation image and the cycle loss of the standard CTA image and the second CTA simulation image is less than the cycle loss threshold, and the sum of the adversarial loss of the first CTA simulation feature and the standard CTA feature and the adversarial loss of the second CT simulation feature and the standard CT feature is less than the adversarial loss threshold, and the sum of the negative sample contrast loss of the first CTA simulation feature and the standard CTA feature and the negative sample contrast loss of the second CT simulation feature and the standard CT feature is less than the contrast loss threshold, output the first generator, the second generator, the third generator, the fourth generator, the first discriminator and the second discriminator to obtain a trained synthesis model; The sum of negative sample contrast loss is obtained by the following formula: ， L 3 is the sum of the negative sample contrast losses, D(G(C)) is the first CTA simulation feature, D(CA) is a standard CTA feature. D(G (CA) is the second CT simulation feature, D(C) It is the standard CT feature; The processing process in the first generator, or / and the second generator, or / and the third generator, or / and the fourth generator is: inputting a plurality of feature convolution processes with the number of channels increasing successively to obtain feature channel enhancement features; The feature channel enhancement features are processed several times by feature residual to obtain fusion features; The fused features are processed by several nearest neighbor upsampling and feature convolution with decreasing number of channels to obtain the output; The operation of feature residual processing is as follows: the feature channel enhancement feature is subjected to convolution, instance normalization and nonlinear processing to obtain an initial feature map; the initial feature is subjected to instance normalization and layer normalization after convolution processing to obtain an instance standard feature map and a layer normalized feature map; based on the pixel variance and pixel mean of the instance standard feature map and the initial feature, a first enhanced feature is obtained; Based on the pixel variance and pixel mean of the layer normalized feature map and the initial feature, a second enhanced feature is obtained; The first enhanced feature and the second enhanced feature are subjected to feature balancing processing to obtain a balanced feature; After nonlinear processing, the balanced features are superimposed with the feature channel enhancement features to obtain the current round of fusion features; and so on, after the last feature residual operation is performed, the fusion features are obtained; After the feature residual processing operation, the fused features are subjected to multi-scale feature extraction processing to obtain multi-scale fused features for performing several nearest neighbor upsampling processes; the multi-scale feature extraction processing operation is specifically as follows: the fused features are subjected to convolution processing of different scales and then superimposed to obtain multi-scale convolution features; The fused features are processed by global average pooling, convolution and upsampling to obtain pooled sampling features; the multi-scale convolution features and the pooled sampling features are superimposed and convolved to obtain multi-scale fused features; S2. The CT image to be processed is processed by the trained synthesis model, and the output of the first generator is extracted to obtain the target CTA synthesis image.

2. The method for generating a CTA image based on a CT image according to claim 1, characterized in that: In S1, the processing process of the first discriminator and / or the second discriminator is: The input is processed by several feature convolutions with increasing number of channels to obtain convolution features; The convolution features are processed by global average pooling and global maximum pooling respectively to obtain average pooling features and maximum pooling features; The average pooling features and the maximum pooling features are fully connected, multiplied element by element with the convolutional features, and then superimposed to obtain the output.

3. The method for generating a CTA image based on a CT image according to claim 1, characterized in that: The CT image or / and CTA image, before being processed by the generator or / and discriminator, includes cropping the CT image or / and CTA image so that the CT value of the image is within the corresponding numerical range, and then normalizing, cropping and normalizing the image to obtain a qualified CT image or / and a qualified CTA image, and executing the operations in the generator or / and discriminator.

4. The method for generating a CTA image based on a CT image according to claim 1, characterized in that: In S1, the circulation loss is obtained by performing paradigm processing on the standard CT image and the first CT simulation image, or the standard CTA image and the second CTA simulation image.

5. A system for generating a CTA image based on a CT image, used to implement the method for generating a CTA image based on a CT image according to claim 1, characterized in that: include: A first generator, used for processing a standard CT image to obtain a first CTA simulation image; A second generator is used to process the first CTA simulation image to obtain a first CT simulation image; The third generator is used to process the standard CTA image to obtain a second CT simulation image; A fourth generator, used for processing the second CT simulation image to obtain a second CTA simulation image; A first discriminator is used to process the first CTA simulation image and the standard CTA image respectively to obtain a first CTA simulation feature and a standard CTA feature; The second discriminator is used to process the second CT simulation image and the standard CT image respectively to obtain a second CT simulation feature and a standard CT feature; Training the synthesis model, when the sum of the cycle loss of the standard CT image and the first CT simulation image and the cycle loss of the standard CTA image and the second CTA simulation image is less than the cycle loss threshold, and the sum of the adversarial loss of the first CTA simulation feature and the standard CTA feature and the adversarial loss of the second CT simulation feature and the standard CT feature is less than the adversarial loss threshold, and the sum of the negative sample contrast loss of the first CTA simulation feature and the standard CTA feature and the negative sample contrast loss of the second CT simulation feature and the standard CT feature is less than the contrast loss threshold, output the first generator, the second generator, the third generator, the fourth generator, the first discriminator and the second discriminator to obtain the training synthesis model; used to process the CT image to be processed through the training synthesis model, extract the output of the first generator, and obtain the target CTA synthesis image; The processing process in the first generator, or / and the second generator, or / and the third generator, or / and the fourth generator is: inputting a plurality of feature convolution processes with the number of channels increasing successively to obtain feature channel enhancement features; The feature channel enhancement features are processed several times by feature residual to obtain fusion features; The fused features are processed by several nearest neighbor upsampling and feature convolution with the number of channels decreasing successively to obtain the output.

6. A device for generating a CTA image based on a CT image, characterized in that: The method comprises a processor and a memory, wherein when the processor executes the computer program stored in the memory, the method for generating a CTA image based on a CT image according to any one of claims 1 to 4 is implemented.

7. A computer-readable storage medium, characterized in that: Used to store a computer program, wherein when the computer program is executed by a processor, the method for generating a CTA image based on a CT image according to any one of claims 1 to 4 is implemented.

Images