Vascular surgery image intelligent optimization method

Abstract

The invention discloses a vascular surgery image intelligent optimization method, and relates to the field of medical image intelligent processing. According to the technical scheme, the method comprises the following steps: collecting original image data according to a vascular surgical image, extracting image features through a deep convolutional neural network CNN, constructing an angiography lesion sensitivity map, generating a sensitivity initial weight map by using the sensitivity map, and constructing a vascular structure topological map based on the initial weight map; and carrying out sensitivity propagation and correction by adopting a graph neural network to obtain a corrected sensitivity weight graph, and carrying out local adaptive image enhancement processing according to the corrected sensitivity weight graph to obtain an intelligently optimized image. The problem that a lesion area is not prone to protruding in angiography is solved, and the diagnosis accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent medical image processing, and more specifically, to an intelligent optimization method for vascular surgical images. Background Art

[0002] Vascular surgical imaging technology is an extremely important part of the modern medical imaging field. Through means such as CT angiography (CTA), magnetic resonance angiography (MRA), and digital subtraction angiography (DSA), non-invasive diagnosis of vascular lesions can be achieved. However, existing image processing methods usually adopt overall or uniform image enhancement means, and such processing methods are difficult to accurately highlight the characteristics of the lesion area. Especially in the case of subtle lesions or when the contrast between the lesion and the surrounding tissues is not obvious, it is easy to lead to missed diagnosis or misdiagnosis. Although there are some methods in the prior art that attempt to solve the above problems, such as initially identifying the lesion area through traditional image filtering, threshold segmentation, or artificial intelligence algorithms, these methods usually do not fully consider the topological characteristics of the vascular structure itself, lack quantitative analysis of regional sensitivity, and fail to achieve local optimization according to lesion sensitivity. Therefore, they often still have the disadvantages of insufficient enhancement in sensitive areas and excessive interference in non-sensitive areas. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide an intelligent optimization method for vascular surgical images to solve the problems mentioned in the background art.

[0004] To achieve the above object, the present invention adopts the following technical solutions: An intelligent optimization method for vascular surgical images, comprising the following steps: Collect original image data according to vascular surgical images, and automatically extract vascular image features through a convolutional neural network to obtain the sensitivity of each image region to different types of lesions; Construct a vascular imaging lesion sensitivity map to quantify and record the sensitivity differences of each region in the image to vascular lesion diagnosis; Generate an initial sensitivity weight map using the sensitivity map, and construct a vascular structure topology map based on the initial weight map; Use a graph neural network to perform sensitivity propagation and correction with topological constraints on the initial sensitivity weight map to obtain a corrected sensitivity weight map; Determine a local enhancement factor according to the corrected sensitivity weight map, and use the local enhancement factor to perform adaptive image enhancement processing on the vascular lesion area to obtain an intelligently optimized vascular image.

[0005] Optionally, the construction method of the vascular imaging lesion sensitivity map specifically includes: Obtain multiple historical vascular image samples, perform manual marking on the lesion areas, and determine the lesion category and location corresponding to each sample; Train the vascular image samples through a convolutional neural network, using the lesion areas and non-lesion areas as supervised learning labels to generate a sensitivity index at the pixel or local region level, and obtain the sensitivity weight values with spatial distribution; Form a sensitivity map according to the numerical values of the sensitivity index for subsequent image optimization.

[0006] Optionally, the processing steps from the initial weight map of the map sensitivity to the optimized weight map specifically include: After initially extracting the sensitivity map using a convolutional neural network, select the nodes representing the vascular topological structure in the image, and use the initial sensitivity weight values of the nodes as the input of the graph node features; Establish a vascular topological structure network, and use a graph neural network to propagate and correct the initial sensitivity weights according to the connection relationships between the nodes; Through the mutual update of the node sensitivity weights, finally form a corrected sensitivity map with vascular topological structure constraints.

[0007] Optionally, the graph neural network specifically adopts a graph convolutional structure with an attention mechanism, and the update formula for the sensitivity weights is: ; where, represents a specific node whose sensitivity weight is currently being updated, represents the set of adjacent nodes of node , represents any one of the neighbor nodes in; represents the th layer of the graph neural network; is the sensitivity weight value of node at the th layer, is the attention weight coefficient, and respectively represent the weight matrix and bias term of the th layer, represents the non-linear activation function.

[0008] Optionally, among them, the local enhancement factor is specifically determined by the following formula: ; where, is the sensitivity weight at the image coordinate position , is the coordinate position The local enhancement factor at and is an adjustable parameter.

[0009] Optionally, the convolutional neural network model includes a feature extraction layer, a lesion type classification sub-network, and a sensitivity weight generation sub-network, and the model parameters are optimized through a supervised training method.

[0010] Optionally, the nodes of the vascular topology network are represented as topological nodes constructed with vascular bifurcation points and lesion key points in the image, and the edges are represented as the vascular image connection paths between adjacent nodes.

[0011] Optionally, the propagation method of the node sensitivity weight includes realizing the adaptive fusion of neighborhood node weights through an attention mechanism, and the attention weight is jointly determined by the sensitivity similarity and the topological connection relationship during each layer of propagation.

[0012] Optionally, the adaptive image enhancement process includes: using a conditional generative adversarial network guided by a local enhancement factor to enhance the image.

[0013] Optionally, a closed-loop iterative structure is formed among the generation of the sensitivity map, the correction of the sensitivity weight, and the image enhancement process. Through the actual diagnosis results of clinical feedback, the generation model parameters of the sensitivity index are dynamically adjusted to continuously optimize the overall image intelligent optimization method.

[0014] The advantages of the present invention over the prior art are as follows: by using a deep convolutional neural network CNN to automatically extract vascular image features, the precise quantification of the sensitivity of the image area to lesion diagnosis is realized for the first time, and an accurate vascular imaging lesion sensitivity map is constructed, thus effectively solving the problems that the lesion area is difficult to highlight and the diagnostic accuracy is insufficient in the prior art. The present invention utilizes the initial sensitivity weight map and constructs a graph neural network based on the vascular structure topology to realize the effective propagation and precise correction of the sensitivity weight, making the sensitivity judgment more in line with the real pathological situation. According to the corrected sensitivity weight map, the present invention implements a local adaptive image enhancement strategy, making the lesion areas with high sensitivity clearly appear, and performing appropriate noise reduction processing on the areas with low sensitivity, greatly improving the accuracy of clinical diagnosis. At the same time, the present invention also forms a closed-loop iterative optimization mechanism through clinical feedback, further improving the accuracy and stability of the intelligent optimization process. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is the overall flowchart of the method of the present invention; Figure 2 is the training flowchart of the convolutional neural network of the present invention; Figure 3 is a flowchart of optimizing the initial weights of the graph neural network of the present invention; Figure 4 is a schematic diagram of two image enhancement methods of the present invention. Specific Embodiments

[0016] The specific embodiments of the present invention will be described below with reference to the accompanying drawings.

[0017] As Figure 1 shown, an intelligent optimization method for vascular surgical images of the present invention includes the following specific steps: Collect original image data according to vascular surgical images, and automatically extract vascular image features through a convolutional neural network to obtain the sensitivity of each image region to different types of lesions; Construct a vascular imaging lesion sensitivity map to quantify and record the sensitivity differences of each region in the image to vascular lesion diagnosis; Generate an initial sensitivity weight map using the sensitivity map, and construct a vascular structure topology map based on the initial weight map; Use the graph neural network to perform sensitivity propagation and correction with topological constraints on the initial sensitivity weight map to obtain a corrected sensitivity weight map; Determine the local enhancement factor according to the corrected sensitivity weight map, and use the local enhancement factor to perform adaptive image enhancement processing on the vascular lesion area to obtain an intelligently optimized vascular image.

[0018] As shown in Figure 2, the training process of the convolutional neural network CNN is specifically as follows: First, a large number of historical vascular image samples (CTA, MRA or DSA) need to be collected from clinical actual diagnosis cases, covering different types and severities of lesion samples as well as healthy vascular images. Next, these image samples are manually labeled by professional doctors to clearly mark the location and scope of the lesion area and indicate the type of the lesion, such as clearly marking plaques, thrombi, aneurysms, etc. in the coronary artery or carotid artery. To improve the generalization ability of the CNN, data enhancement processing is usually also performed on the image samples, such as rotation, cropping, scale change, and noise addition, to increase data diversity.

[0019] During the training process, the input of the CNN is vascular image data, which can be three-dimensional images or two-dimensional slice data. The images are usually imported in a standard format (such as DICOM format). The supervision labels that can be provided for the classification sub-network are lesion types, such as plaque as category A, thrombus as category B, aneurysm as category C, and healthy area as category D, etc.; the supervision labels provided for the sensitivity weight sub-network are the numerically sensitive areas manually marked, for example, the lesion area is marked as high sensitivity (such as between 0.8 and 1.0), and the non-lesion area is marked as low sensitivity (such as between 0.0 and 0.2). During the training process: In the first step, the image is input into the CNN, and the image depth features are obtained through the feature extraction layer. In the second step, the depth features are sent into the classification sub-network to output the prediction result of the lesion type, and at the same time sent into the sensitivity weight generation sub-network to predict the sensitivity value; for the lesion classification sub-network, the cross-entropy loss function can be adopted; for the sensitivity weight sub-network, the mean square error loss function can be adopted; the two loss functions can be weighted and combined to form a total loss function for training.

[0020] In the third step, the gradient descent method is used to optimize the CNN network parameters: Through the gradient descent method (such as the Adam optimizer), the gradient is calculated according to the total loss function, and the parameters in the CNN network (such as convolution kernel weights, bias terms, etc.) are continuously updated to reduce the loss and improve the classification and sensitivity prediction accuracy.

[0021] For example, a patient undergoes a coronary CTA scan. After obtaining the coronary artery image, it is directly input into the trained CNN network. The CNN outputs an initial sensitivity weight map. For example, in a bifurcation area of the coronary artery, a sensitivity of 0.9 is given, indicating a high suspicion of plaque lesions.

[0022] After that, through the correction of the graph neural network, the sensitivities of the surrounding areas are considered together. After propagation and correction, the true sensitivity (such as 0.95) of the 0.9 area is made more clear to confirm that this area is a true lesion area, and more significant enhancement measures are implemented accordingly during image processing.

[0023] The reason for correcting through the graph neural network is that the convolutional neural network does not consider the topological relationship of the vascular structure (that is, the true spatial connection relationship between blood vessels) when calculating these sensitivity weights. Therefore, the sensitivity weights of some areas may not be accurate enough. For a simple example: Suppose we get a very high sensitivity weight for a region, but the sensitivity weights of all the surrounding connected regions are very low. Then this isolated high-sensitivity region may be problematic or less reliable; on the contrary, if a region has a high sensitivity weight and the other vascular regions connected to it also have a high sensitivity, then the high sensitivity of this region is more credible and is more likely to be a truly diseased region.

[0024] Therefore, it is hoped that there is a way to consider the situation of the surrounding areas to correct the sensitivity of each area. This is the role of the graph neural network GNN. As shown in Figure 3 specifically: After obtaining the initial sensitivity weight map, to further improve its accuracy and practicality, a vascular structure topology map is constructed. The topology map uses important structural points in the vascular image (such as vascular bifurcation points, key lesion points) as nodes, and the actual connection relationships between these nodes as edges, clarifying the spatial topological relationship of the vascular structure. Then, using the initial sensitivity weight values at the positions corresponding to these nodes as node features, they are input into the graph neural network GNN to achieve the propagation and correction of sensitivity.

[0025] Specifically, the graph neural network adopts a graph convolution structure with an attention mechanism to further improve the learning ability of the network and the degree of attention to key nodes. The update mechanism of the sensitivity weight is as follows: ; Among them, represents a specific node whose sensitivity weight is currently being updated, represents the set of adjacent nodes of node , represents any one of the neighbor nodes in represents the -th layer of the graph neural network; is the sensitivity weight value of node at the -th layer, is the attention weight coefficient, and its value range is 0 to 1; and respectively represent the weight matrix and bias term of the -th layer, represents a non-linear activation function (such as the ReLU function).

[0026] The meaning of the above formula is: the sensitivity weight value of each node after the -th (or the -th layer) update is obtained by weighted synthesis of the previous-round sensitivity weight values of all neighbor nodes around node , and then through linear transformation and non-linear activation processing. That is to say, in the graph neural network, each node is not updated independently, and each of its updates needs to comprehensively consider the states of its directly connected neighbor nodes.

[0027] For example, in the vascular network, node i represents a section of blood vessel currently being concerned, and its neighbor node j represents other adjacent blood vessel sections.

[0028] Initially, each node has sensitivity weights given by the CNN. However, these initial weights may have errors or noise. Therefore, we need to let each node "listen" to the opinions of its neighbor nodes before making a decision: If the sensitivity weights of neighbor nodes are all high, it means that the surrounding area is generally judged as a sensitive area, and then the current node also tends to increase its own sensitivity; if the sensitivities of neighbor nodes are all low, it indicates that the surrounding area may be normal, so even if the current node has a high sensitivity, it should appropriately decrease it to avoid isolated misjudgments.

[0029] Attention coefficient The role is to tell each node how to reasonably refer to the opinions of neighbors. Those more similar and reliable neighbor opinions will receive more attention and have greater influence.

[0030] Next, the information of all neighbors will be uniformly processed by a linear transformation matrix and then adjusted by a bias term Finally, after a non - linear transformation through an activation function the corrected sensitivity weights are obtained.

[0031] After multiple rounds of iteration of this process, the sensitivity weights of each node will eventually tend to a more stable and accurate value, thereby obtaining a clearer and more precise sensitivity distribution map.

[0032] When specifically implementing image enhancement, the enhancement factor is determined by the following formula: ; where is the sensitivity weight at the image coordinate position , and are adjustable parameters. The higher the sensitivity weight, the greater the local enhancement amplitude of the image. The value range is usually: 0.5 - 5, and the recommended value range is generally: 1 - 3.

[0033] The value range is usually: 1 - 5; the recommended value range is: 2 - 4.

[0034] The role of is to control the non - linear influence of the sensitivity weight on the enhancement factor. When

[0035] = 1, it is linear enhancement, and when it is greater than 1, it shows exponential enhancement, which is more conducive to highlighting the lesion characteristics of high - sensitive areas. When the sensitivity weight at a certain position is higher (indicating a higher likelihood of a lesion), the enhancement factor is larger, meaning that a greater degree of image quality enhancement is required (such as significantly improving contrast, sharpness, and sharpening details); When the sensitivity weight at a certain position is lower, the enhancement factor is closer to 1, indicating that the image requires little additional enhancement and may even be appropriately denoised.

[0036] That is to say, the sensitivity weight is the basis for judgment, while the enhancement factor is the control parameter directly used for image enhancement.

[0037] As shown in Figure 4, we can choose one of two ways to perform image enhancement: Method 1: Directly perform traditional enhancement through the enhancement factor (such as classic image sharpening, contrast adjustment, etc.) This method is more traditional and intuitive. The enhancement factor is directly used to control the intensity of image enhancement, such as multiplying the local contrast or the size of the sharpening filter kernel.

[0038] Method 2 (more advanced and intelligent): Use the enhancement factor as a conditional input to guide the conditional generative adversarial network (cGAN) to perform automatic enhancement.

[0039] This method is more intelligent and has better results. In this case, the enhancement factor is similar to providing an accurate "enhancement intensity guidance map" to the cGAN network, rather than directly telling the network the sensitivity magnitude. The network intelligently generates the optimal enhancement effect based on this enhancement factor guidance map.

[0040] For example, in a coronary CTA image of a patient, there is a possible plaque area: The predicted sensitivity weight by CNN + GNN is 0.9, which means the likelihood of a lesion in this area is very high; Next, calculate the enhancement factor according to the above formula. For example, select the parameters = 2, = 3, then: ; The enhancement factor of 2.458 indicates that the contrast or sharpness of this area needs to be significantly enhanced, about 2.458 times that of the normal area; after inputting this enhancement factor into the conditional GAN, the conditional GAN clearly knows that this area needs to be enhanced by a factor of 2.458, so it intelligently performs higher-contrast enhancement, more obvious edge sharpening, and highlighting of lesion details in this area; In contrast, for another area with a sensitivity weight of 0.1, the enhancement factor is: ; This indicates that little enhancement is required, and the conditional GAN basically keeps this area unchanged or slightly reduces noise.

[0041] In this way, the final image shows a clear and prominent lesion area, a natural and smooth normal area, and significantly improved diagnostic accuracy.

[0042] In the above method, the generation of the sensitivity map, the GNN correction of the sensitivity weight, and the image enhancement processing together form a closed-loop iterative structure. During specific implementation, by regularly collecting actual feedback data of clinical diagnoses, the relevant parameters of the CNN model and the GNN model are dynamically adjusted. For example, when the clinical diagnosis feedback shows that the recognition accuracy of a certain lesion area is low, this method can automatically adjust the training parameters of the model and the weight distribution of the attention mechanism to continuously optimize the sensitivity map and the final intelligent image enhancement effect, and improve the accuracy and stability of the clinical diagnosis effect.

[0043] The above specific implementation manners can be applied to different vascular surgery imaging scenarios, such as coronary CTA examinations, carotid ultrasound contrast examinations, etc., and can accurately assist doctors in making lesion diagnoses and subsequent treatment decisions.

[0044] The above is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered within the protection scope of the present invention.

Claims

1. An intelligent optimization method for vascular surgery imaging, characterized in that It includes the following steps: Collect the original image data according to the vascular surgery images, and automatically extract the vascular image features through a convolutional neural network to obtain the sensitivity of each image region to different types of lesions; Construct a vascular imaging lesion sensitivity map to quantify and record the sensitivity differences of each region in the image for vascular lesion diagnosis; Generate an initial sensitivity weight map using the sensitivity map and construct a vascular structure topology map based on the initial weight map; Use a graph neural network to perform sensitivity propagation and correction with topological constraints on the initial sensitivity weight map to obtain a corrected sensitivity weight map; Determine the local enhancement factor according to the corrected sensitivity weight map, and use the local enhancement factor to perform adaptive image enhancement processing on the vascular lesion region to obtain an intelligently optimized vascular image.

2. The method according to claim 1, wherein Among them, The construction method of the vascular imaging lesion sensitivity map specifically includes: Obtain multiple historical vascular image samples, manually mark the lesion regions, and determine the lesion category and location corresponding to each sample; Train the vascular image samples through a convolutional neural network, using the lesion region and the non-lesion region as supervised learning labels to generate sensitivity indices at the pixel or local region level, and obtain spatially distributed sensitivity weight values; Form a sensitivity map according to the numerical values of the sensitivity indices for subsequent image optimization.

3. The method according to claim 1, characterized in that Among them, The processing steps from the initial weight map of the map sensitivity to the optimized weight map specifically include: After preliminarily extracting the sensitivity map using a convolutional neural network, select the nodes representing the vascular topology structure in the image, and use the initial sensitivity weight values of the nodes as the input of the graph node features; Establish a vascular topology structure network, and use a graph neural network to propagate and correct the initial sensitivity weights according to the connection relationships between the nodes; Through the mutual update of the node sensitivity weights, finally form a corrected sensitivity map with vascular topology structure constraints.

4. The method according to claim 3, wherein Among them, The graph neural network specifically adopts a graph convolutional structure with an attention mechanism, and the update formula of the sensitivity weight is: ； Among them, represents a specific node that is currently updating the sensitivity weight, represents the node 's set of adjacent nodes, represents any one of the neighbor nodes in represents the -th layer of the graph neural network; For the node the layer sensitivity weight value, is the attention weight coefficient, and respectively represent the weight matrix and bias term of the layer, and represents the non-linear activation function.

5. The method according to claim 1, wherein Among them, The local enhancement factor is specifically determined by the following formula: ； wherein, is the sensitivity weight at the image coordinate position , is the local enhancement factor at the coordinate position , and are adjustable parameters.

6. The method according to claim 1, wherein The convolutional neural network model includes a feature extraction layer, a lesion type classification sub-network, and a sensitivity weight generation sub-network, and optimizes the model parameters through a supervised training method.

7. The method according to claim 3, wherein Among them, The nodes of the vascular topology structure network are represented as topological nodes constructed with vascular bifurcation points and lesion key points as the core in the image, and the edges are represented as the vascular image connection paths between adjacent nodes.

8. The method according to claim 3, wherein Among them, The propagation method of the node sensitivity weights includes realizing the adaptive fusion of the neighborhood node weights through the attention mechanism, and the attention weights are jointly determined by the sensitivity similarity and the topological connection relationship in each layer of the propagation process.

9. The method according to claim 1, characterized in that Among them, The adaptive image enhancement processing process includes: using a conditional generative adversarial network guided by the local enhancement factor to perform image enhancement.

10. The method according to claim 1, wherein Among them, A closed-loop iterative structure is formed among the generation of the sensitivity map, the correction of the sensitivity weights, and the image enhancement processing. Through the actual diagnosis results feedback clinically, the generation model parameters of the sensitivity indices are dynamically adjusted to continuously optimize the overall image intelligent optimization method.

Images