Angiography image optimization method

Abstract

The invention provides an angiography image optimization method, and relates to the field of image enhancement. The invention provides an angiography image data processing flow. Comprising the steps of collecting angiographic image data, constructing a first-stage feature coding module, constructing a weight supplement module, constructing a second-stage feature coding module, constructing a third-stage feature enhancement module, constructing a multi-stage image data enhancement module, and constructing an angiographic image data processing model. Training an angiography image data processing model and processing angiography image data; meanwhile, an angiography image data processing model is provided, a first-stage feature coding module can effectively extract high-level features, a weight supplementing module dynamically adjusts feature weights, a second-stage feature coding module further compresses and reconstructs the features, and a third-stage feature enhancing module enhances interactivity and global representation ability of the features. And the multi-stage image data enhancement module finally generates high-quality enhanced angiography image data.

Description

Technical Field

[0001] The invention belongs to the field of image enhancement, and in particular relates to a method for optimizing a vascular imaging image. Background Art

[0002] Vascular imaging technology is an extremely important part of medical imaging and is widely used in the diagnosis and treatment of cardiovascular and cerebrovascular diseases, tumors, diabetes and other diseases. However, due to the limitations of imaging equipment and environment, vascular imaging image data often have problems such as low contrast and severe noise interference, which brings great challenges to subsequent target detection and diagnosis. Image enhancement technology came into being in this context. Through a series of algorithms and processing methods, the quality of image data is improved, thereby assisting doctors in making more accurate diagnoses.

[0003] In vascular imaging image data, improving contrast and clarity is the primary goal of image enhancement. In low-contrast image data, small vascular structures are often difficult to identify and may even be drowned out by background noise. Image enhancement technology can significantly improve the contrast of the image, making small blood vessels and their branches more clearly visible. This processing not only allows doctors to more easily identify and analyze vascular structures and their diseased areas during the diagnosis process, but also provides better quality input data for subsequent automatic target detection algorithms.

[0004] The latest ViT architecture usually combines the multi-scale feature fusion method. By fusing features of different scales, it can more comprehensively capture the multi-scale information in the image data. Multi-scale feature fusion helps to improve the detail retention and global information integration of the image data, making the image data enhancement effect more significant. In this way, the ViT architecture can better handle various complex situations in the image data and improve the overall quality of the image data. Summary of the invention

[0005] The present invention provides a vascular imaging image optimization method, aiming to propose a vascular imaging image data processing model, wherein a first-stage feature encoding module applies a multi-head self-attention mechanism and a feedforward neural network, which can effectively extract high-level features of fuzzy vascular imaging image data, a weight supplementation module dynamically adjusts feature weights by introducing Gaussian noise and uniform noise to balance the distribution of samples of different categories, a second-stage feature encoding module can further compress and reconstruct features through the encoder, sampling and decoder structure of VAE, and capture the potential representation of image data, a third-stage feature enhancement module enhances the interactivity and global representation ability of features through a hybrid self-attention mechanism, and a multi-stage image data enhancement module finally generates high-quality enhanced vascular imaging image data through layer-by-layer feature extraction and enhancement.

[0006] The present invention aims to propose a vascular imaging image data processing model and provide a vascular imaging image optimization method, comprising the following steps: S1, vascular imaging image data collection, collecting multiple clear vascular imaging images to form a vascular imaging image data set; S2, build a one-stage feature encoding module, including an adapted Transformer module; S3, construct a weight supplement module, including weight supplement operations; S4, construct a two-stage feature encoding module, including a weight supplement module; S5, build a three-stage feature enhancement module, including hybrid self-attention; S6, constructing a multi-stage image data enhancement module, including a first-stage feature encoding module, a second-stage feature encoding module, a third-stage feature enhancement module and image generation convolution; S7, constructing a vascular imaging image data processing model, including an input, a multi-stage image data enhancement module and an output; S8. Training of vascular imaging image data processing model and processing of vascular imaging image data. The vascular imaging image data set is used to train the vascular imaging image data processing model. After the training is completed, the vascular imaging image data to be processed is obtained and input into the vascular imaging image data processing model to obtain enhanced vascular imaging image data.

[0007] Preferably, in step S1, for the vascular imaging image dataset, firstly, a plurality of clear vascular imaging images are collected, including obtaining a plurality of clear vascular imaging images from professional journals and conference proceedings, and obtaining a plurality of clear vascular imaging image data from a scientific image library, and blurring the plurality of clear vascular imaging images using Gaussian blur, defocus blur, motion blur and noise addition to obtain a plurality of blurred vascular imaging images, and forming a plurality of training pairs corresponding to the plurality of clear vascular imaging images, thereby forming a vascular imaging image dataset.

[0008] Preferably, in step S2, for the first-stage feature encoding module, a single blurred vascular imaging image data I is input. f , I f ∈R h×w×3 , h, w and 3 are single blurred vascular imaging image data I f The height, width and channels of a single blurred vascular imaging image data I f Input into multiple series-connected adaptive Transformer modules to obtain feature F1, F1∈R H×W×C, H, W and C are the height, width and channel of feature F1. A single adapted Transformer module consists of a Transformer module and an adapter layer. The Transformer module includes a multi-head self-attention mechanism and a feedforward neural network FFN. The adapter layer contains a dimensionality reduction linear layer, a nonlinear activation function, a dimensionality increase linear layer and a residual connection. The adapter layer is inserted into the middle position of the multi-head self-attention mechanism and the feedforward neural network FFN in a single Transformer module to form a single adapted Transformer module.

[0009] Preferably, in step S2, for the one-stage feature encoding module, an adapted Transformer module is integrated, including a multi-head self-attention mechanism and a feedforward neural network FFN, and an adapter layer is introduced to enhance feature expression through dimensionality reduction linear layers, nonlinear activation functions, dimensionality increase linear layers and residual connections. These adapter layers are inserted between the multi-head self-attention mechanism and FFN of the Transformer module, so that the feature encoding module can more effectively extract and represent high-level features of fuzzy vascular imaging image data.

[0010] Preferably, in step S3, for the weight supplementation module, the basic feature F is input WS1 and optimize the feature F WS2 , F WS1 ∈R H×W×C , H, W and C are the basic features F WS1 The height, width and channel of F WS2 ∈R H×W×C , optimize feature F WS2 and the basic feature F WS1 The dimensions are consistent, and then the supplementary optimization feature F is obtained through weight supplementation operation. WS2+ , F WS2+ ∈R H×W×C , the weight supplement operation is specifically F WS2+ =F WS2 (gt==i)+=α·(μ·Noise1+(1-μ)·Noise2), gt represents the true mask, gt==i is used to determine whether the true category of each pixel is equal to category i, where α is the noise intensity control parameter, Noise1 is the Gaussian noise term, Noise2 is the uniform noise term, μ represents the parameter controlling the ratio of Gaussian noise term and uniform noise term, μ takes a value from 0 to 1, · represents the multiplication operation, N represents Gaussian noise, β is a balance term to prevent the denominator from being too small, γ is the confidence adjustment coefficient, Confidence(i) is the sample confidence of category i, var(i) is the inverse variance of the category sample frequency, U represents uniform noise, and then the feature F is obtained WSO , FWSO ∈R H×W×C , F WSO =(FFN(SA(DW(F WS2 +)))+F WS2 )⊙F WS1 +F WS1 , where DW stands for deep convolution, SA stands for self-attention operation, FFN stands for feedforward neural network, + stands for element-by-element addition, ⊙ stands for element-by-element product, and feature F WSO is the output of the weight supplement module, feature F WSO 、Basic feature F WS1 and optimize the feature F WS2 The dimensions remain consistent.

[0011] Preferably, in step S3, for the weight supplementation module, Gaussian noise and uniform noise are combined, the distribution of samples of different categories is balanced by dynamically adjusting the feature weights, and the noise intensity is adjusted using the inverse variance of the category sample frequency and the sample confidence to enhance the feature representation of the sparse categories. Through the noise injection mechanism, a richer and more robust feature expression is provided, thereby improving the performance and accuracy of the vascular imaging image data processing model when processing unbalanced data sets.

[0012] Preferably, in step S4, for the second-stage feature encoding module, the feature F1 output by the first-stage feature encoding module is obtained, F1∈R H×W×C , H, W and C are the height, width and channel of feature F1, then we get feature F 1S , F 1S ∈R H×W×C , F 1S =DW(SA(F1)), SA represents the self-attention operation, DW represents the deep convolution, and then obtains the feature F 1V , F 1V ∈R H×W×C , F 1V =VAE(F1), VAE stands for variational autoencoder, which includes encoder, sampling and decoder, and then obtains the feature F 1M , F 1M ∈R H×W×C , F 1M =F 1S ⊙F 1V ⊙ represents the element-by-element product, and feature F1 is used as the basic feature F WS1 , the feature F 1M As the optimization feature F WS2 , the basic feature F WS1 and optimize the feature F WS2 Input to the weight supplement module to obtain feature F WSO , the feature F WSOAs feature F2, F2∈R H×W×C , feature F2, feature F 1V and feature F 1M is the output of the two-stage feature encoding module.

[0013] Preferably, in step S4, for the two-stage feature encoding module, a variational autoencoder VAE structure is combined, and features are further compressed and reconstructed through encoders, sampling and decoders, and self-attention operations and deep convolutions are combined to capture the potential representation of image data through VAE, enhance the diversity and representation ability of features, and dynamically adjust feature weights using a weight supplement module to optimize feature representation, so that the vascular imaging image data processing model is more robust and accurate when processing samples of different categories.

[0014] Preferably, in step S5, for the three-stage feature enhancement module, the features F2 and F3 output by the two-stage feature encoding module are obtained. 1V and feature F 1M , first obtain the feature F 3V , F 3V ∈R H×W×C , F 3V =VAE(F 1V +F2), VAE stands for variational autoencoder, + stands for element-by-element addition, and then feature F is obtained by hybrid self-attention 3M , F 3M ∈R H×W×C , the mixed self-attention is calculated as F 3M =FFN(Conv(F 1M ⊙SoftMax(QK T ))V), where Conv represents 3×3 convolution, FFN represents feedforward neural network, ⊙ represents element-wise product, SoftMax represents SoftMax activation function, and Q = F2W Q , K = F2W K , V=F2W V , Q stands for query, K stands for key, V stands for value, W Q is the weight matrix for generating query Q, W K is the weight matrix for generating key K, W V is the weight matrix that generates the value V, and then obtains the enhanced feature F3, F3∈R H×W×C , F3=F 3M ⊙F 3V , ⊙ represents the element-wise product, and the enhanced feature F3 is the output of the three-stage feature enhancement module.

[0015] Preferably, in step S5, for the three-stage feature enhancement module, a hybrid self-attention mechanism is introduced, combined with a variational autoencoder VAE structure, the interactivity and global representation capabilities of the features are enhanced through the relationship between queries, keys and values, VAE is used to further compress and reconstruct the features, and the hybrid self-attention mechanism is used to perform global interaction and enhancement of the features, ultimately generating a more robust and diverse feature representation, significantly improving the performance of the vascular imaging image data processing model in processing complex image data tasks.

[0016] Preferably, in step S6, for the multi-stage image data enhancement module, including a first-stage feature encoding module, a second-stage feature encoding module, a third-stage feature enhancement module and an image generation convolution, firstly input a single blurred vascular imaging image data I f To the first-stage feature encoding module, we get the feature F1, I f ∈R h×w×3 , h, w and 3 are single blurred vascular imaging image data I f The height, width and channel of F1∈R H×W×C , G, W and C are the height, width and channel of feature F1, and then input feature F1 into the two-stage feature encoding module to obtain features F2 and F 1V and feature F 1M , F2∈R H×W×C , F 1V ∈R H×W×C , F 1M ∈R H×W×C , then enter feature F2, feature F 1V and feature F 1M To the three-stage feature enhancement module, we get the enhanced feature F3, F3∈R H×W×C Finally, the enhanced feature F3 is input into the image generation convolution to obtain a single enhanced vascular imaging image data I + , I + ∈R h×w×3 , the image generation convolution is implemented by 3×3 convolution.

[0017] Preferably, in step S6, for the multi-stage image data enhancement module, a first-stage feature encoding, a second-stage feature encoding and a third-stage feature enhancement module are integrated, and high-quality enhanced vascular imaging image data is finally generated through layer-by-layer feature extraction and enhancement. The module combines the adaptive Transformer, the self-attention mechanism, the variational autoencoder VAE and the weight supplementation operation, so that the features are fully extracted and optimized at each stage, thereby improving the details and clarity of the vascular imaging images, and ensuring that the finally generated vascular imaging image data has higher quality and interpretability.

[0018] Compared with the prior art, the present invention has the following technical effects: The technical solution provided by the present invention proposes a vascular imaging image data processing model, in which a first-stage feature encoding module applies a multi-head self-attention mechanism and a feedforward neural network to effectively extract high-level features of fuzzy vascular imaging image data. The weight supplement module dynamically adjusts feature weights by introducing Gaussian noise and uniform noise to balance the distribution of samples of different categories. The second-stage feature encoding module can further compress and reconstruct features through the encoder, sampling and decoder structure of VAE to capture the potential representation of vascular imaging image data. The third-stage feature enhancement module enhances the interactivity and global representation ability of features through a hybrid self-attention mechanism. The multi-stage image data enhancement module extracts and enhances features layer by layer to ultimately generate high-quality enhanced vascular imaging image data. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of vascular imaging image data processing provided by the present invention.

[0020] Figure 2 It is a structural diagram of the multi-stage image data enhancement module provided by the present invention.

[0021] Figure 3 This is a single blurred blood vessel imaging image provided by the present invention.

[0022] Figure 4 This is a single enhanced vascular imaging image provided by the present invention. DETAILED DESCRIPTION

[0023] The present invention aims to propose a vascular imaging image optimization method and a vascular imaging image data processing model, wherein a first-stage feature encoding module applies a multi-head self-attention mechanism and a feedforward neural network to effectively extract high-level features of fuzzy vascular imaging image data, a weight supplementation module dynamically adjusts feature weights by introducing Gaussian noise and uniform noise to balance the distribution of samples of different categories, a second-stage feature encoding module can further compress and reconstruct features through the encoder, sampling and decoder structure of VAE to capture the potential representation of vascular imaging image data, a third-stage feature enhancement module enhances the interactivity and global representation ability of features through a hybrid self-attention mechanism, and a multi-stage image data enhancement module extracts and enhances features layer by layer to ultimately generate high-quality enhanced vascular imaging image data.

[0024] See also Figure 1 As shown, a method for optimizing vascular imaging images in an embodiment of the present application is as follows: S1, vascular imaging image data collection, collect 200 clear vascular imaging images to form a vascular imaging image data set; S2, build a one-stage feature encoding module, including an adapted Transformer module; S3, construct a weight supplement module, including weight supplement operations; S4, construct a two-stage feature encoding module, including a weight supplement module; S5, build a three-stage feature enhancement module, including hybrid self-attention; S6, constructing a multi-stage image data enhancement module, including a first-stage feature encoding module, a second-stage feature encoding module, a third-stage feature enhancement module and image generation convolution; S7, constructing a vascular imaging image data processing model, including an input, a multi-stage image data enhancement module and an output; S8. Training of vascular imaging image data processing model and processing of vascular imaging image data. The vascular imaging image data set is used to train the vascular imaging image data processing model. After the training is completed, the vascular imaging image data to be processed is obtained and input into the vascular imaging image data processing model to obtain enhanced vascular imaging image data.

[0025] Furthermore, in step S1, for the vascular imaging image dataset, first, 200 clear vascular imaging images are collected, including 100 clear vascular imaging images obtained from professional journals and conference proceedings, and 100 clear vascular imaging images obtained from a scientific image library, and Gaussian blur, defocus blur, motion blur, and noise addition are used to blur the multiple clear vascular imaging images to obtain multiple blurred vascular imaging images, and 200 training pairs are formed corresponding to the multiple clear vascular imaging images, thereby forming a vascular imaging image dataset.

[0026] Further, in step S2, for the first-stage feature encoding module, a single fuzzy vascular imaging image data I is input f , I f ∈R h×w×3 , h, w and 3 are single blurred vascular imaging image data I f The height, width and channels of a single blurred vascular imaging image data I f Input into multiple series-connected adaptive Transformer modules to obtain feature F1, F1∈R H×W×C , H, W and C are the height, width and channel of feature F1. A single adapted Transformer module consists of a Transformer module and an adapter layer. The Transformer module includes a multi-head self-attention mechanism and a feedforward neural network FFN. The adapter layer contains a dimensionality reduction linear layer, a nonlinear activation function, a dimensionality increase linear layer and a residual connection. The adapter layer is inserted into the middle position of the multi-head self-attention mechanism and the feedforward neural network FFN in a single Transformer module to form a single adapted Transformer module.

[0027] Further, in step S3, for the weight supplementation module, the basic feature F is input WS1 and optimize the feature F WS2 , F WS1 ∈R H×W×C , H, W and C are the basic features F WS1 The height, width and channel of F WS2 ∈R H×W×C , optimize feature F WS2 and the basic feature F WS1 The dimensions are consistent, and then the supplementary optimization feature F is obtained through weight supplementation operation. WS2+ , F WS2+ ∈R H×W×C , the weight supplement operation is specifically F WS2+ =F WS2 (gt==i)+βα·(μ·Noise1+(1-μ)·Noise2), gt represents the true mask, gt==i is used to determine whether the true category of each pixel is equal to category i, where α is the noise intensity control parameter, Noise1 is the Gaussian noise term, Noise2 is the uniform noise term, μ represents the parameter controlling the ratio of Gaussian noise term and uniform noise term, μ takes a value from 0 to 1, · represents the multiplication operation, N represents Gaussian noise, β is a balance term to prevent the denominator from being too small, γ is the confidence adjustment coefficient, Confidence(i) is the sample confidence of category i, var(i) is the inverse variance of the category sample frequency, U represents uniform noise, and then the feature F is obtained WSO , F WSO ∈R H×W×C , F WSO =(FFN(SA(DW(F WS2+ )))+F WS2 )⊙F WS1 +F WS1 , where DW stands for deep convolution, SA stands for self-attention operation, FFN stands for feedforward neural network, + stands for element-by-element addition, ⊙ stands for element-by-element product, and feature F WSO is the output of the weight supplement module, feature F WSO 、Basic feature F WS1 and optimize the feature F WS2 The dimensions remain consistent.

[0028] Further, in step S4, for the second-stage feature encoding module, the feature F1 output by the first-stage feature encoding module is obtained, F1∈R H×W×C , H, W and C are the height, width and channel of feature F1, then we get feature F 1S , F 1S ∈RH×W×C , F 1S =DW(SA(F1)), SA represents the self-attention operation, DW represents the deep convolution, and then obtains the feature F 1V , F 1V ∈R H×W×C , F 1V =VAE(F1), VAE stands for variational autoencoder, which includes encoder, sampling and decoder, and then obtains the feature F 1M , F 1M ∈R H×W×C , F 1M =F 1S ⊙F 1V ⊙ represents the element-by-element product, and feature F1 is used as the basic feature F WS1 , the feature F 1M As the optimization feature F WS2 , the basic feature F WS1 and optimize the feature F WS2 Input to the weight supplement module to obtain feature F WSO , the feature F WSO As feature F2, F2∈R H×W×C , feature F2, feature F 1V and feature F 1M is the output of the two-stage feature encoding module.

[0029] Further, in step S5, for the three-stage feature enhancement module, the features F2 and F2 output by the two-stage feature encoding module are obtained. 1V and feature F 1M , first obtain the feature F 3V , F 3V ∈R H×W×C , F 3V =VAE(F 1V +F2), VAE stands for variational autoencoder, + stands for element-by-element addition, and then feature F is obtained by hybrid self-attention 3M , F 3M ∈R H×W×C , the mixed self-attention is calculated as F 3M =FFN(Conv(F 1M ⊙SoftMax(QK T ))V), where Conv represents 3×3 convolution, FFN represents feedforward neural network, ⊙ represents element-wise product, SoftMax represents SoftMax activation function, and Q = F2W Q , K = F2W K , V=f2W V , Q stands for query, K stands for key, V stands for value, W Q is the weight matrix for generating query Q, WK is the weight matrix for generating key K, W V is the weight matrix that generates the value V, and then obtains the enhanced feature F3, F3∈R H×W×C , F3=F 3M ⊙F 3V , ⊙ represents the element-wise product, and the enhanced feature F3 is the output of the three-stage feature enhancement module.

[0030] Further, in step S6, for the multi-stage image data enhancement module, its structure is as follows Figure 2 As shown, it includes a first-stage feature encoding module, a second-stage feature encoding module, a third-stage feature enhancement module and an image generation convolution. First, a single blurred vascular imaging image data I is input. f To the first-stage feature encoding module, we get the feature F1, I f ∈R 2000×2000×3 , 2000, 2000 and 3 are single fuzzy vascular imaging image data I f The height, width and channel of F1∈R 512×512×64 , 512, 512 and 64 are the height, width and channel of feature F1, and then feature F1 is input to the two-stage feature encoding module to obtain features F2 and F 1V and feature F 1M , F2∈R 512×512×64 , F 1V ∈R 512×512×64 , F 1M ∈R 512×512×64 , then input feature F2, feature f 1V and feature F 1M To the three-stage feature enhancement module, we get the enhanced feature F3, F3∈R 512×512×64 , and finally enhance the features

[0031] Further, in step S7, for the vascular imaging image data processing model, written in Python language, improved based on the ViT algorithm architecture, using the Pytorch framework, all were trained from scratch for 300 times using the AdamW optimizer, with a learning rate of 0.001 and a total training batch size of 900.

[0032] Furthermore, taking a single blurred vascular imaging image with a height and width of 2000px as an example, its source is a larger vascular imaging image with a higher resolution, and it is enlarged to view more detailed structures and lesion areas. This enlargement will cause the vascular imaging image to be blurred and may cover up important details. Therefore, it is necessary to perform image enhancement through the vascular imaging image data processing model. At the same time, the vascular imaging image data processing model has lower computational overhead than the traditional ViT architecture and is more suitable for use in real scenarios.

[0033] Further, in step S8, the vascular imaging image data processing effect of the vascular imaging image data processing model is as follows: Figure 3 and Figure 4 As shown, Figure 3 A single blurry vascular imaging image is shown. Figure 4 It shows that a single blurred vascular imaging image is input into the vascular imaging image data processing model to obtain a single clear vascular imaging image.

[0034] The above are only preferred embodiments of the present invention. It should be pointed out that a person skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for optimizing a vascular imaging image, characterized in that: The following steps are involved: S1, vascular imaging image data collection, collecting multiple clear vascular imaging images to form a vascular imaging image data set; S2, build a one-stage feature encoding module, including an adapted Transformer module; S3, construct a weight supplement module, including weight supplement operations; S4, construct a two-stage feature encoding module, including a weight supplement module; S5, build a three-stage feature enhancement module, including hybrid self-attention; S6, constructing a multi-stage image data enhancement module, including a first-stage feature encoding module, a second-stage feature encoding module, a third-stage feature enhancement module and image generation convolution; S7, constructing a vascular imaging image data processing model, including an input, a multi-stage image data enhancement module, and an output; S8. Training of vascular imaging image data processing model and processing of vascular imaging image data. The vascular imaging image data set is used to train the vascular imaging image data processing model. After the training is completed, the vascular imaging image data to be processed is obtained and input into the vascular imaging image data processing model to obtain enhanced vascular imaging image data.

2. A method for optimizing vascular imaging images according to claim 1, characterized in that: In step S1, for the vascular imaging image dataset, firstly, a plurality of clear vascular imaging images are collected, including obtaining a plurality of clear vascular imaging images from professional journals and conference proceedings, and obtaining a plurality of clear vascular imaging image data from a scientific image library, and the plurality of clear vascular imaging images are blurred using Gaussian blur, defocus blur, motion blur and noise addition to obtain a plurality of blurred vascular imaging images, and a plurality of training pairs are formed corresponding to the plurality of clear vascular imaging images, thereby forming a vascular imaging image dataset.

3. The method for optimizing vascular imaging images according to claim 1, characterized in that: In step S2, for the first-stage feature encoding module, a single fuzzy vascular imaging image data I is input. f , I f ∈R h×w×3 , h, w and 3 are single blurred vascular imaging image data I f The height, width and channels of a single blurred vascular imaging image data I f Input into multiple series-connected adaptive Transformer modules to obtain feature F1, F1∈R H×W×C , H, W and C are the height, width and channel of feature F1. A single adapted Transformer module consists of a Transformer module and an adapter layer. The Transformer module includes a multi-head self-attention mechanism and a feedforward neural network FFN. The adapter layer contains a dimensionality reduction linear layer, a nonlinear activation function, a dimensionality increase linear layer and a residual connection. The adapter layer is inserted into the middle position of the multi-head self-attention mechanism and the feedforward neural network FFN in a single Transformer module to form a single adapted Transformer module.

4. The method for optimizing vascular imaging images according to claim 1, characterized in that: In step S3, for the weight supplementation module, the basic feature F is input WS1 and optimize the feature F WS2 , F WS1 ∈R H×W×C , H, W and C are the basic features F WS1 The height, width and channel of F WS2 ∈R H×W×C , optimize feature F WS2 and the basic feature F WS1 The dimensions are consistent, and then the supplementary optimization feature F is obtained through weight supplementation operation. WS2+ , F WS2+ ∈R H×W×C , the weight supplement operation is specifically F WS2+ =F WS2 (gt==i)+=α·(μ·Noise1+(1-μ)·Noise2), gt represents the true mask, gt==i is used to determine whether the true category of each pixel is equal to category i, where α is the noise intensity control parameter, Noise1 is the Gaussian noise term, Noise2 is the uniform noise term, μ represents the parameter controlling the ratio of Gaussian noise term and uniform noise term, μ takes a value from 0 to 1, · represents the multiplication operation, N represents Gaussian noise, β is a balance term to prevent the denominator from being too small, γ is the confidence adjustment coefficient, Confidence(i) is the sample confidence of category i, var(i) is the inverse variance of the category sample frequency, U represents uniform noise, and then the feature F is obtained WSO , F WSO ∈R H×W×C , F WSO =(FFN(SA(DW(F WS2 +)))+F WS2 )⊙F WS1 +F WS1 , where DW stands for deep convolution, SA stands for self-attention operation, FFN stands for feedforward neural network, + stands for element-by-element addition, ⊙ stands for element-by-element product, and feature F WSO is the output of the weight supplement module, feature F WSO 、Basic feature F WS1 and optimize the feature F WS2 The dimensions remain consistent.

5. The method for optimizing vascular imaging images according to claim 1, characterized in that: In step S4, for the second-stage feature encoding module, the feature F1 output by the first-stage feature encoding module is obtained, F1∈R H×W×C , H, W and C are the height, width and channel of feature F1, then we get feature F 1S , F 1S ∈R H×W×C , F 1S =DW(SA(F1)), SA represents the self-attention operation, DW represents the deep convolution, and then obtains the feature F 1V , F 1V ∈R H×W×C , F 1V =VAE(F1), VAE stands for variational autoencoder, which includes encoder, sampling and decoder, and then obtains the feature F 1M , F 1M ∈R H×W×C , F 1M =F 1S ⊙F 1V ⊙ represents the element-by-element product, and feature F1 is used as the basic feature F WS1 , the feature F 1M As the optimization feature F WS2 , the basic feature F WS1 and optimize the feature F WS2 Input to the weight supplement module to obtain feature F WSO , the feature F WSO As feature F2, F2∈R H×W×C , feature F2, feature F 1V and feature F 1M is the output of the two-stage feature encoding module.

6. The method for optimizing vascular imaging images according to claim 1, characterized in that: In step S5, for the three-stage feature enhancement module, the features F2 and F2 output by the two-stage feature encoding module are obtained. 1V and feature F 1M , first obtain the feature F 3V , F 3V ∈R H×W×C , F 3V =VAE(F 1V +F2), VAE stands for variational autoencoder, + stands for element-by-element addition, and then feature F is obtained by hybrid self-attention 3M , F 3M ∈R H×W×C , the mixed self-attention is calculated as F 3M =FFN(Conv(F 1M ⊙SoftMax(QK T ))V), where Conv represents 3×3 convolution, FFN represents feedforward neural network, ⊙ represents element-wise product, SoftMax represents SoftMax activation function, and Q = F2W Q , K = F2W K , V=F2W V , Q stands for query, K stands for key, V stands for value, W Q is the weight matrix for generating query Q, W K is the weight matrix for generating key K, W V is the weight matrix that generates the value V, and then obtains the enhanced feature F3, F3∈R H×W×C , F3=F 3M ⊙F 3V , ⊙ represents the element-wise product, and the enhanced feature F3 is the output of the three-stage feature enhancement module.

7. The method for optimizing vascular imaging images according to claim 1, characterized in that: In step S6, for the multi-stage image data enhancement module, including the first-stage feature encoding module, the second-stage feature encoding module, the third-stage feature enhancement module and the image generation convolution, firstly input a single blurred vascular imaging image data I f To the first-stage feature encoding module, we get the feature F1, I f ∈R h×w×3 , h, w and 3 are single blurred vascular imaging image data I f The height, width and channel of F1∈R H ×W×C , H, W and C are the height, width and channel of feature F1, and then feature F1 is input to the two-stage feature encoding module to obtain features F2 and F 1V and feature F 1M , F2∈R H×W×C , F 1V ∈R H×W×C , F 1M ∈R H×W×C , then enter feature F2, feature F 1V and feature F 1M To the three-stage feature enhancement module, we get the enhanced feature F3, F3∈R H×W×C Finally, the enhanced feature F3 is input into the image generation convolution to obtain a single enhanced vascular imaging image data I + , I + ∈R h×w×3 , the image generation convolution is implemented by 3×3 convolution.

Images