A coronary CT image crime vessel detection system and method based on multi-scale feature extraction

Abstract

The present application relates to a kind of coronary CT image criminal vessel detection system and method based on multiscale feature extraction, belong to image processing technical field, including data preprocessing module, multiscale coronary stenosis identification module, criminal evaluation module, visualization module.Combined with multiscale coronary stenosis identification module and criminal evaluation module, multiscale image generation, feature extraction and fusion, can capture the full-scale structure from coronary trunk to small branch, through the accurate identification of coronary stenosis and its criminal evaluation realizes the coronary criminal vessel detection based on coronary CT image.Through identification and logistic regression evaluation, aims at improving the accuracy of coronary criminal vessel detection.

Description

Technical Field

[0001] This invention relates to a system and method for detecting criminal vessels in coronary CT images based on multi-scale feature extraction, belonging to the field of image processing technology. Background Technology

[0002] With the widespread application of coronary CT imaging technology, the detection of culprit vessels in the coronary arteries plays a crucial role in reducing the workload of physicians and determining the optimal stent placement location for clinical diagnosis and surgical planning. The detection of culprit vessels is based on the discovery of coronary artery stenosis; therefore, the accurate identification of coronary artery stenosis and its culprit assessment directly impact treatment plans or surgical decisions for coronary heart disease. However, existing methods for detecting culprit vessels in the coronary arteries face difficulties in identifying coronary artery stenosis.

[0003] Coronary arteries exhibit a complex tree-like structure, typically comprising small, tortuous, or intersecting vessel segments. Stenosis in these segments is often difficult to effectively identify using traditional image processing methods. Existing methods for detecting coronary artery stenosis largely rely on techniques such as edge detection and threshold segmentation. However, these methods are often ineffective when processing small vessel segments in coronary CT images, easily missing tiny lesion areas or producing artifacts, leading to missed diagnoses or misdiagnoses.

[0004] This invention proposes a multi-scale coronary artery stenosis identification module and a crime assessment module, enabling more accurate identification of coronary artery stenosis regions and crime assessment, avoiding the shortcomings of traditional methods that neglect global structure. This invention demonstrates higher sensitivity and accuracy in detecting coronary artery stenosis regions in coronary CT images, and higher diagnostic precision in crime assessment of coronary artery stenosis regions. This invention can accomplish the task of coronary artery crime vessel detection based on coronary CT images, improving the efficiency and accuracy of clinical diagnosis and surgical decision-making, reducing the workload of physicians, and has significant application prospects in clinical diagnosis and surgical decision-making. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a system and method for detecting criminal vessels in coronary CT images based on multi-scale feature extraction. This system combines a multi-scale coronary artery stenosis identification module and a criminal assessment module. Through accurate identification of coronary artery stenosis and its criminal assessment, it achieves the detection of criminal vessels in coronary arteries based on coronary CT images. By employing identification and logistic regression evaluation, it aims to improve the accuracy of coronary criminal vessel detection, assisting physicians in clinical diagnosis or surgical decisions, and has significant clinical application value.

[0006] The technical solution of the present invention is as follows:

[0007] A coronary CT image crime vessel detection system based on multi-scale feature extraction includes the following modules:

[0008] Data preprocessing module: used to remove noise and correct grayscale in coronary CT images, with the aim of removing interference, highlighting image details, and identifying coronary arteries;

[0009] Multi-scale coronary artery stenosis identification module: It uses deep learning methods for multi-scale image generation and multi-scale feature extraction to identify the stenotic area of ​​the coronary artery;

[0010] Crime assessment module: Using a logistic regression model, it determines whether pixels belonging to the coronary artery stenosis area belong to the coronary artery crime vessel area;

[0011] Visualization module: Displays the culprit vessels in the coronary arteries on coronary CT images and labels their probability of committing crimes.

[0012] Preferably, the data preprocessing module performs the following operations: noise removal and grayscale transformation, and coronary artery identification;

[0013] Noise Removal: Assume the original 2D images are stacked into a 3D image. I raw Gaussian filtering is used for I raw The image is smoothed by weighted averaging of each pixel and its neighborhood to obtain the denoised image. ;

[0014] Gray-scale correction: Coronary CT images acquired by different devices or at different times using the same device may exhibit gray-scale differences, which can affect subsequent analysis. Gray-scale correction aims to adjust the gray-scale distribution of the image to achieve optimal contrast and brightness. This invention uses histogram equalization to adjust the gray-scale distribution of the denoised image. Gray-level correction is performed to adjust the gray-level histogram distribution to a uniform distribution, thereby enhancing the image contrast and obtaining a denoised and gray-level corrected image. I * ;

[0015] Coronary artery identification: Using the K-means method on denoised and grayscale-corrected images I * The coronary arteries were isolated, resulting in a CT image containing only the coronary arteries. H is I The height, W is I The width, C is I The number of channels. The specific steps are as follows:

[0016] enter: I * ;

[0017] Process: To I * All pixels are clustered using K-means, with 2 clusters, one of which represents the coronary arteries;

[0018] Output: CT images containing only coronary arteries .

[0019] Preferably, the multi-scale coronary artery stenosis identification module performs the following operations: automatic identification of coronary artery stenosis using multi-scale local feature extraction and global morphological analysis. Accurate identification of coronary artery stenosis provides a reasonable and correct analysis area for subsequent crime assessment. The specific steps are as follows:

[0020] enter: ;

[0021] Process: Using deep learning methods for multi-scale image generation and multi-scale feature extraction, determine whether a pixel belongs to a coronary artery stenosis area;

[0022] Output: Coronary artery stenosis area R .

[0023] In a further optimized manner, the specific steps during the execution of the multi-scale coronary artery stenosis identification module include:

[0024] 1. Multi-scale image generation

[0025] Input coronary CT images I Multi-scale images are generated using pyramid image processing techniques. I s ( s =1 , 2 , … , S ), S The total number of scales, each scale representing an image at a different resolution; images at each scale... I s Independent input residual network (ResNet) is used to capture coronary artery features at different scales;

[0026] 2. Multi-scale feature extraction

[0027] Each scale I s The image will undergo feature extraction via a residual network; ResNet learns deeper features by introducing residual connections, while avoiding the vanishing gradient problem that may occur during training. The input to ResNet is... I s Its structure is as follows:

[0028] (1) Initial convolutional layer

[0029] Convolutional layer: kernel size is k 1×k 1. Step size is s 1;

[0030] Batch normalization layer: Perform normalization;

[0031] ReLU activation layer: introduces a nonlinear transformation;

[0032] Max pooling layer: pooling kernel size is p × p Step size is s p ;

[0033] (2) Residual block

[0034] 4 residual blocks:

[0035] Convolutional layers: k 2× k 2. k 3× k 3 and k 4× k A 4-fold convolutional kernel is used for feature extraction;

[0036] Batch normalization layer: Normalizes the output of each convolutional layer;

[0037] ReLU activation layer: introduces a nonlinear transformation;

[0038] (3) Skip connection: Perform residual connection, that is, add the input of one residual block to the output of another residual block;

[0039] (4) Global average pooling layer: generates global feature vectors F s .

[0040] 3. Multi-scale feature fusion

[0041] Will come from different scales s Feature maps output by ResNet F s The feature maps are fused to obtain the fused feature maps. F , F= concat ( F 1 , F 2 , … , F s ), F It contains multi-level information from different scales, which helps to more accurately identify coronary artery stenosis areas of different sizes.

[0042] 4. Prediction of fully connected layers and narrow regions

[0043] Feature map after fusion F Through the fully connected layer and sigmod An activation function maps high-dimensional features to the probability that each pixel is a narrow region:

[0044] p ( x , y ) = σ ( w 1 F 1 ( x , y ) + w 2 F 2 ( x , y ) +…+ w s F s ( x , y ) + b 1) (1)

[0045] in, w 1 ,w 2 , … ,w s These are the weight coefficients learned during training. b 1 is the bias. σ for sigmod Activation function x , y Represents the position coordinates in the image. F ( x , y ) is the location ( x , y The pixel value at ()

[0046] 5. Determination of coronary artery stenosis area

[0047] Based on the predicted probability value p ( x, y Set a threshold δ :

[0048] (2)

[0049] in f ( x, y )=1 indicates a pixel I ( x, y ) belongs to the area of ​​coronary artery stenosis; otherwise, pixels I ( x, y This does not belong to the coronary artery stenosis area; the coronary artery stenosis area can be represented as...R ={( x, y )}| f ( x, y ) = 1;

[0050] 6. Loss Function

[0051] The binary cross-entropy loss function is used to evaluate whether each pixel is correctly classified as a narrow region or a non-narrow region:

[0052] (3)

[0053] in yes I Medium pixels ( x, y The true label indicates whether the pixel is a region of coronary artery stenosis.

[0054] Preferably, the crime assessment module performs the following operations:

[0055] enter: Coronary artery stenosis area R ;

[0056] Process: Using a logistic regression model, determine whether pixels belonging to the coronary artery stenosis area are coronary culprit vessels;

[0057] Output: Coronary culprit vessel region C and the crime probability of each pixel. Q .

[0058] Furthermore, the execution process of the crime assessment module includes the following steps: First, select... R exist F The corresponding value is obtained. Using a logistic regression model, the output is... Pixels I ( x, y The probability of committing the crime is:

[0059] (4)

[0060] in m =1 indicates the narrow location I ( x, y This refers to the situation of coronary artery crime vessels; setting a crime threshold. θ ,when q ( x, y > θ At that time, the pixel was considered I ( x, y () is a culprit in the coronary artery. w T It is the weight vector of the logistic regression model. b2 is the bias term; from this, all coronary culprit vessel regions are obtained. C and the crime probability of each pixel. Q The binary cross-entropy loss function is used to evaluate whether each pixel is correctly classified as a coronary culprit or a non-coronary culprit.

[0061] (5)

[0062] in It is a pixel I ( x, y The true label indicates whether the pixel is a coronary culprit; the total loss of the system training method proposed in this invention. L = L 1+ λL 2, of which λ This is a smoothing factor.

[0063] Preferably, the visualization module performs the following operations for display:

[0064] enter: I raw , C , Q ;

[0065] Procedure: [The text abruptly shifts to a seemingly unrelated topic about coronary artery malfunction.] C In the original coronary CT images I raw The above is presented, and the probability of crime for each pixel in the coronary artery crime vessel region is labeled. Q ;

[0066] Output: Visualized diagnostic images.

[0067] A method for detecting criminal vessels in coronary CT images based on multi-scale feature extraction is proposed. Based on the aforementioned system for detecting criminal vessels in coronary CT images based on multi-scale feature extraction, each module in the system performs its own operation for detection.

[0068] The beneficial effects of this invention are as follows:

[0069] This invention enables the detection of criminal vessels without the need for invasive intravascular probing; multi-scale image generation, feature extraction, and fusion can capture the full-scale structure from the main coronary artery to small branches; this invention directly assesses the "criminality" of stenotic pixels through a logistic regression model, making it more intuitive and visually interpretable. Attached Figure Description

[0070] Figure 1 This is a schematic diagram of the technical route of the present invention;

[0071] Figure 2This is a schematic diagram of a visualized diagnostic image. Detailed Implementation

[0072] The present invention will be further described below with reference to the embodiments and accompanying drawings, but is not limited thereto.

[0073] Example 1:

[0074] A coronary CT image crime vessel detection system based on multi-scale feature extraction, such as Figure 1 As shown, it includes the following modules:

[0075] Data preprocessing module: used to remove noise and correct grayscale in coronary CT images, with the aim of removing interference, highlighting image details, and identifying coronary arteries.

[0076] Multi-scale coronary artery stenosis identification module: It uses deep learning methods for multi-scale image generation and multi-scale feature extraction to identify the stenotic area of ​​the coronary artery.

[0077] Crime assessment module: Using a logistic regression model, it determines whether pixels belonging to the coronary artery stenosis area belong to the coronary artery crime vessel area.

[0078] Visualization module: Displays the culprit vessels in the coronary arteries on coronary CT images and labels their probability of committing crimes.

[0079] The data preprocessing module performs the following operations: noise removal and grayscale transformation, and coronary artery identification.

[0080] Noise Removal: Assume the original 2D images are stacked into a 3D image. I raw Gaussian filtering is used for I raw The image is smoothed by weighted averaging of each pixel and its neighborhood to obtain the denoised image. .

[0081] Gray-scale correction: Coronary CT images acquired by different devices or at different times using the same device may exhibit gray-scale differences, which can affect subsequent analysis. Gray-scale correction aims to adjust the gray-scale distribution of the image to achieve optimal contrast and brightness. This invention uses histogram equalization to adjust the gray-scale distribution of the denoised image. Gray-level correction is performed to adjust the gray-level histogram distribution to a uniform distribution, thereby enhancing the image contrast and obtaining a denoised and gray-level corrected image. I * .

[0082] Coronary artery identification: Using the K-means method on denoised and grayscale-corrected images I * The coronary arteries were isolated, resulting in a CT image containing only the coronary arteries. H is I The height, W is I The width, C is I The number of channels. The specific steps are as follows:

[0083] enter: I * ;

[0084] Process: To I * All pixels are clustered using K-means, with 2 clusters, one of which represents the coronary arteries;

[0085] Output: CT images containing only coronary arteries .

[0086] The multi-scale coronary artery stenosis identification module performs the following operations: automatic identification of coronary artery stenosis using multi-scale local feature extraction and global morphological analysis. Accurate identification of coronary artery stenosis provides a reasonable and correct analysis area for subsequent crime assessment. The specific steps are as follows:

[0087] enter: ;

[0088] Process: Using deep learning methods for multi-scale image generation and multi-scale feature extraction, determine whether a pixel belongs to a coronary artery stenosis area;

[0089] Output: Coronary artery stenosis area R .

[0090] The specific steps involved in the execution of the multi-scale coronary artery stenosis identification module include:

[0091] 1. Multi-scale image generation

[0092] Input coronary CT images I Multi-scale images are generated using pyramid image processing techniques. I s ( s =1 , 2 , … , S ), S The total number of scales, each scale representing an image at a different resolution; images at each scale... I s Independent input residual network (ResNet) is used to capture coronary artery features at different scales.

[0093] 2. Multi-scale feature extraction

[0094] Each scale I sThe image will undergo feature extraction via a residual network; ResNet learns deeper features by introducing residual connections, while avoiding the vanishing gradient problem that may occur during training. The input to ResNet is... I s Its structure is as follows:

[0095] (1) Initial convolutional layer

[0096] Convolutional layer: kernel size is k 1× k 1. Step size is s 1;

[0097] Batch normalization layer: Perform normalization;

[0098] ReLU activation layer: introduces a nonlinear transformation;

[0099] Max pooling layer: pooling kernel size is p × p Step size is s p .

[0100] (2) Residual block

[0101] 4 residual blocks:

[0102] Convolutional layers: k 2× k 2. k 3× k 3 and k 4× k A 4-fold convolutional kernel is used for feature extraction;

[0103] Batch normalization layer: Normalizes the output of each convolutional layer;

[0104] ReLU activation layer: Introduces nonlinear transformation.

[0105] (3) Skip connection: Perform residual connection, that is, add the input of one residual block to the output of another residual block.

[0106] (4) Global average pooling layer: generates global feature vectors F s .

[0107] 3. Multi-scale feature fusion

[0108] Will come from different scales s Feature maps output by ResNet F s The feature maps are fused to obtain the fused feature maps. F ,F= concat ( F 1 , F 2 , … , F s ), F It contains multi-level information from different scales, which helps to more accurately identify coronary artery stenosis areas of different sizes.

[0109] 4. Prediction of fully connected layers and narrow regions

[0110] Feature map after fusion F Through the fully connected layer and sigmod An activation function maps high-dimensional features to the probability that each pixel is a narrow region:

[0111] p ( x , y ) = σ ( w 1 F 1 ( x , y ) + w 2 F 2 ( x , y ) +…+ w s F s ( x , y ) + b 1)(1)

[0112] in, w 1 ,w 2 , … ,w s These are the weight coefficients learned during training. b 1 is the bias. σ for sigmod Activation function x , y Represents the position coordinates in the image. F ( x , y ) is the location ( x , y The pixel value at ().

[0113] 5. Determination of coronary artery stenosis area

[0114] Based on the predicted probability value p ( x, ySet a threshold δ :

[0115] (2)

[0116] in f ( x, y )=1 indicates a pixel I ( x, y ) belongs to the area of ​​coronary artery stenosis; otherwise, pixels I ( x, y This does not belong to the coronary artery stenosis area; the coronary artery stenosis area can be represented as... R ={( x, y )}| f ( x, y =1.

[0117] 6. Loss Function

[0118] The binary cross-entropy loss function is used to evaluate whether each pixel is correctly classified as a narrow region or a non-narrow region:

[0119] (3)

[0120] in yes I Medium pixels ( x, y The true label indicates whether the pixel is a region of coronary artery stenosis.

[0121] The crime assessment module performs the following operations:

[0122] enter: Coronary artery stenosis area R ;

[0123] Process: Using a logistic regression model, determine whether pixels belonging to the coronary artery stenosis area are coronary culprit vessels;

[0124] Output: Coronary culprit vessel region C and the crime probability of each pixel. Q .

[0125] The execution process of the crime assessment module includes the following steps: First, select... R exist F The corresponding value is obtained. Using a logistic regression model, the output is... Pixels I ( x, y The probability of committing the crime is:

[0126] (4)

[0127] in m =1 indicates the narrow locationI ( x, y This refers to the situation of coronary artery crime vessels; setting a crime threshold. θ ,when q ( x, y > θ At that time, the pixel was considered I ( x, y () is a culprit in the coronary artery. w T It is the weight vector of the logistic regression model. b 2 is the bias term; from this, all coronary culprit vessel regions are obtained. C and the crime probability of each pixel. Q The binary cross-entropy loss function is used to evaluate whether each pixel is correctly classified as a coronary culprit or a non-coronary culprit.

[0128] (5)

[0129] in It is a pixel I ( x, y The true label indicates whether the pixel is a coronary culprit; the total loss of the system training method proposed in this invention. L = L 1+ λL 2, of which λ This is a smoothing factor.

[0130] The visualization module performs the following operations to demonstrate:

[0131] enter: I raw , C , Q ;

[0132] Procedure: [The text abruptly shifts to a seemingly unrelated topic about coronary artery malfunction.] C In the original coronary CT images I raw The above is presented, and the probability of crime for each pixel in the coronary artery crime vessel region is labeled. Q ;

[0133] Output: Visualized diagnostic images. (e.g.) Figure 2 As shown.

[0134] set up A To utilize the coronary culprit vessel region detected in this embodiment, B The areas of coronary malpractice vessels are manually marked.

[0135] Dice similarity coefficient: This index measures the degree of overlap between coronary culprit detection results and manually labeled data. The value ranges from 0 to 1; a higher value indicates a more accurate detection result. It can be expressed as:

[0136] (6)

[0137] Intersection over Union (IoU): Evaluates the degree of overlap between segmented regions. IoU A higher value indicates a better test result. The formula is:

[0138] (7)

[0139] Accuracy: Measures whether each pixel is accurately classified; the calculation formula is:

[0140] (8)

[0141] in, TP Pixels that correctly identify a region as a coronary culprit vessel. TN Pixels that correctly classify a region as a non-coronary culprit vessel. FP and FN All represent misclassified pixels;

[0142] Specificity: Assess the model's ability to detect non-coronary culprit vessels; higher specificity means a lower false positive rate for healthy tissue; the calculation formula is:

[0143] (9)

[0144] Sensitivity: Evaluates the model's ability to detect culprit vessels in the coronary arteries; higher sensitivity means fewer false negatives; the calculation formula is:

[0145] (10)

[0146] Using the above evaluation criteria, the output results of this system are evaluated, and the overall evaluation accuracy of this system can reach over 93%.

[0147] Example 2

[0148] A method for detecting criminal vessels in coronary CT images based on multi-scale feature extraction is provided, which is based on the coronary CT image criminal vessel detection system based on multi-scale feature extraction described in Example 1. Each module in the system performs its own operation to perform detection.

Claims

1. A coronary CT image crime vessel detection system based on multi-scale feature extraction, characterized in that, Includes the following modules: Data preprocessing module: used to remove noise and correct grayscale in coronary CT images, and to identify coronary arteries; Multi-scale coronary artery stenosis identification module: It uses deep learning methods for multi-scale image generation and multi-scale feature extraction to identify the stenotic area of ​​the coronary artery; The steps are as follows: enter: ; Process: Using deep learning methods for multi-scale image generation and multi-scale feature extraction, determine whether a pixel belongs to a coronary artery stenosis area; Output: Coronary artery stenosis area R ; The specific steps include: (1) Multiscale image generation Input coronary CT images I Multi-scale images are generated using pyramid image processing techniques. I s ( s =1 , 2 , … ,S ), S The total number of scales, each scale representing an image at a different resolution; images at each scale... I s Independent input residual network ResNet is used to capture coronary artery features at different scales; (2) Multi-scale feature extraction Each scale I s The image will undergo feature extraction via a residual network; ResNet introduces residual connections, and the input to ResNet is... I s Its structure is as follows: (2-1) Initial convolutional layer Convolutional layer: kernel size is k 1× k 1. Step size is s 1; Batch normalization layer: Perform normalization; ReLU activation layer: introduces a nonlinear transformation; Max pooling layer: pooling kernel size is p × p Step size is s p ; (2-2) Residual Block 4 residual blocks: Convolutional layers: k 2× k 2. k 3× k 3 and k 4× k A 4-fold convolutional kernel is used for feature extraction; Batch normalization layer: Normalizes the output of each convolutional layer; ReLU activation layer: introduces a nonlinear transformation; (2-3) Skip connection: Perform residual connection, that is, add the input of one residual block to the output of another residual block; (2-4) Global average pooling layer: generates global feature vectors F s ; (3) Multi-scale feature fusion Will come from different scales s Feature maps output by ResNet F s The feature maps are fused to obtain the fused feature maps. F , F=concat ( F 1 , F 2 , … , F s ), F It contains multi-layered information from different scales; (4) Prediction of fully connected layers and narrow regions Feature map after fusion F Through the fully connected layer and sigmod An activation function maps high-dimensional features to the probability that each pixel is a narrow region: p ( x , y ) = σ ( w 1 F 1 ( x , y ) + w 2 F 2 ( x , y ) +…+ w s F s ( x , y ) + b 1) （1） in, w 1 ,w 2 , … ,w s These are the weight coefficients learned during training. b 1 is the bias. σ for sigmod Activation function x , y Represents the position coordinates in the image. F ( x , y ) is the location ( x , y The pixel value at () (5) Determination of coronary artery stenosis area Based on the predicted probability value p ( x,y Set a threshold δ : （2） in f ( x,y )=1 indicates a pixel I ( x,y ) belongs to the area of ​​coronary artery stenosis; otherwise, pixels I ( x,y This does not belong to the area of ​​coronary artery stenosis; the area of ​​coronary artery stenosis is represented as... R ={( x,y )}| f ( x,y ) = 1; (6) Loss function The binary cross-entropy loss function is used to evaluate whether each pixel is correctly classified as a narrow region or a non-narrow region: （3） in yes I Medium pixels ( x,y The true label indicates whether the pixel is a region of coronary artery stenosis; Crime assessment module: Using a logistic regression model, it determines whether pixels belonging to coronary artery stenosis areas belong to the coronary artery crime vessel area; the crime assessment module performs the following operations: enter: Coronary artery stenosis area R ; Process: Using a logistic regression model, determine whether pixels belonging to the coronary artery stenosis area are coronary culprit vessels; Output: Coronary culprit vessel region C and the crime probability of each pixel. Q ; The specific steps are as follows: First, select... R exist F The corresponding value is obtained. Using a logistic regression model, the output is... Pixels I ( x,y The probability of committing the crime is: （4） in m =1 indicates the narrow location I ( x,y This refers to the situation of coronary artery crime vessels; setting a crime threshold. θ ,when q ( x,y > θ At that time, the pixel was considered I ( x,y () is a culprit in the coronary artery. w T It is the weight vector of the logistic regression model. b 2 is the bias term; from this, all coronary culprit vessel regions are obtained. C and the crime probability of each pixel. Q The binary cross-entropy loss function is used to evaluate whether each pixel is correctly classified as a coronary culprit or a non-coronary culprit. （5） in It is a pixel I ( x,y The true label of a pixel indicates whether it represents a coronary artery culprit; the total loss of the system training method. L = L 1+ λL 2, of which λ It is a smoothing factor; Visualization module: Displays the culprit vessels in the coronary arteries on coronary CT images and labels their probability of committing crimes.

2. The coronary CT image crime vessel detection system based on multi-scale feature extraction according to claim 1, characterized in that, The data preprocessing module performs the following operations: noise removal and grayscale transformation, and coronary artery identification. Noise Removal: Assume the original 2D images are stacked into a 3D image. I raw Gaussian filtering is used for I raw The image is smoothed by weighted averaging of each pixel and its neighborhood to obtain the denoised image. ; Grayscale correction: Histogram equalization is used to correct the noise in the denoised image. Perform grayscale correction to adjust the grayscale histogram distribution to a uniform distribution, resulting in a denoised and grayscale-corrected image. I * ; Coronary artery identification: Using the K-means method on denoised and grayscale-corrected images I * The coronary arteries were isolated, resulting in a CT image containing only the coronary arteries. H is I The height, W is I The width, C is I The number of channels; The specific steps are as follows: enter: I * ; Process: To I * All pixels are clustered using K-means, with 2 clusters, one of which represents the coronary arteries; Output: CT images containing only coronary arteries .

3. The coronary CT image crime vessel detection system based on multi-scale feature extraction according to claim 1, characterized in that, The visualization module performs the following operations to demonstrate: enter: I raw , C , Q ; Procedure: [The text abruptly shifts to a seemingly unrelated topic about coronary artery malfunction.] C In the original coronary CT images I raw The above is presented, and the probability of crime for each pixel in the coronary artery crime vessel region is labeled. Q ; Output: Visualized diagnostic images.

4. A method for detecting criminal vessels in coronary CT images based on multi-scale feature extraction, characterized in that, Based on the coronary CT image crime vessel detection system according to any one of claims 1-3, each module in the system performs its own operation to perform detection.

Images