A coronary artery segmentation method based on deep learning

Abstract

The present invention provides a coronary artery segmentation method based on deep learning, including the following steps: B1. obtaining an original CTA heart image; B2. normalizing and preprocessing the original CTA heart image in B1, selecting the best CT value observation window, intercepting the coronary artery area in the CTA image, suppressing non-cardiac tissues such as pulmonary veins, and improving the contrast between the target and the background; B3. image data augmentation; B4. constructing a deep learning network, using two deep neural networks Net A and Net B based on a three-layer relay supervision mechanism but with different training labels to train coronary artery CTA data, and obtaining coronary artery recognition models A and B; B5. using the coronary artery recognition model obtained in B4 to perform coronary artery segmentation. The present invention solves the technical problem that the existing coronary artery segmentation method is prone to over-segmentation.

Description

Technical Field

[0001] The present invention belongs to the field of medical image processing, and in particular relates to a coronary artery segmentation method based on CTA images for non-diagnostic purposes, which is a coronary artery segmentation method based on deep learning. Background Art

[0002] The coronary arteries are blood vessels that surround the heart and supply blood to the myocardium. Coronary artery diseases such as coronary atherosclerosis can cause part of the myocardium to not get enough oxygen and nutrients through the blood and gradually lose function, which in turn causes cardiac pulsation dysfunction. Therefore, the health of the coronary arteries plays a vital role in the normal functioning of the heart.

[0003] At present, coronary CT angiography (CTA) is mainly used in clinical diagnosis of patients with coronary heart disease. Accurate segmentation and extraction of coronary arteries based on CTA images can assist doctors in diagnosing cardiovascular diseases and formulating appropriate surgical plans. In addition, coronary segmentation is also an important basis for 3D reconstruction of blood vessels, allowing doctors to more intuitively identify the location of lesions.

[0004] Coronary segmentation technology based on traditional image processing methods generally uses the tubular characteristics of coronary vessels to design corresponding vascular filter functions, suppress background tissue, enhance vascular features, and achieve the effect of improving contrast. However, in CTA images, the grayscale values ​​of coronary vessels and other heart tissues are very similar, so over-segmentation is very likely to occur in the subsequent coronary segmentation process, greatly reducing the robustness of the algorithm and the accuracy of segmentation. Summary of the invention

[0005] The technical problem to be solved by the present invention is that the existing coronary artery segmentation method has the technical problem that over-segmentation phenomenon is very likely to occur.

[0006] In order to solve the above technical problems, the present invention provides a coronary artery segmentation method based on deep learning, which includes the following steps:

[0007] B1. Obtain original CTA cardiac images;

[0008] B2. Normalize and preprocess the original CTA cardiac image in B1. Select the best CT value observation window, intercept the coronary artery area in the CTA image, suppress non-cardiac tissues such as pulmonary veins, and improve the contrast between the target and the background; the best CT value observation window is selected manually or automatically by a computer according to medical standards.

[0009] B3. Image data augmentation;

[0010] B4. Construct a deep learning network and use two deep neural networks Net A and Net B based on a three-layer relay supervision mechanism but with different training labels to train coronary artery CTA data to obtain coronary artery recognition models A and B;

[0011] B5. Use the coronary artery recognition model obtained in B4. to perform coronary artery segmentation.

[0012] Preferably, the method of image data augmentation in step B3 includes one or more of horizontal flipping, vertical flipping, random scaling, random elastic transformation, and random gamma correction.

[0013] Preferably, the network structures of Net A and Net B both include eight layers, wherein the first layer is the input layer, the second to seventh layers are hidden layers, and the eighth layer is the output layer. The layers are connected in sequence, and the structures of the layers are as follows:

[0014] First layer: Input layer: input the three-dimensional CTA image matrix of the training set and the true label of each CTA matrix;

[0015] The second layer: hidden layer: contains two convolutional layers, each followed by a normalization layer and an activation function layer, and finally a pooling layer for downsampling; a pooling layer is connected after two convolutional layers to compress the amount of data and parameters and reduce overfitting. In short, if the input is an image, the main function of the pooling layer is to remove redundant features in the image and retain only the most important features.

[0016] The third layer: hidden layer: contains two convolutional layers, each of which is followed by a normalization layer and an activation function layer, and finally a pooling layer is used for downsampling;

[0017] The fourth layer: hidden layer: contains two convolutional layers, each of which is followed by a normalization layer and an activation function layer, and finally a pooling layer is used for downsampling;

[0018] The fifth layer: hidden layer: contains two convolution layers, each of which is followed by a normalization layer, an activation function layer, and finally an UpSampling layer. The UpSampling layer performs upsampling. The upper layer image data is normalized and activated after the first convolution in this layer, and then convolved again and then normalized and activated, and finally upsampled.

[0019] The sixth layer: hidden / relay output layer: includes a skip splicing layer, two convolution layers, each convolution is followed by a normalization layer, an activation function layer, and finally an UpSampling layer, and outputs a loss function Loss1. The UpSampling performs upsampling, and the upper layer image data is normalized and activated after the first convolution in this layer, and then convolved again and normalized and activated again, and finally upsampling is performed;

[0020] The seventh layer: hidden / relay output layer: includes a skip splicing layer, two convolution layers, each convolution is followed by a normalization layer, an activation function layer, and finally an UpSampling layer, and outputs a loss function Loss2. The UpSampling performs upsampling, and the upper layer image data is normalized and activated after the first convolution in this layer, and then convolved again and normalized and activated again, and finally upsampling is performed;

[0021] The eighth layer: hidden / relay output layer: it includes a skip splicing layer, two convolution layers, each convolution is followed by a normalization layer, an activation function layer, and finally an UpSampling layer, and outputs a loss function Loss3. The UpSampling performs upsampling, and the upper layer image data is normalized and activated after the first convolution in this layer, and then convolved again and normalized and activated again, and finally upsampling is performed;

[0022] Ninth layer: Output layer: The final Loss of the network is calculated by Loss1, Loss2, and Loss3

[0023]

[0024] The Loss1, Loss2, and Loss3 are Dice coefficient difference functions, and the calculation method is: Where X is the manually annotated label segmentation image, Y is the segmentation image predicted by the network, and |X| and |Y| represent the number of elements in X and Y, respectively. The coefficient 2 in the numerator is because the denominator repeatedly counts the common elements between X and Y.

[0025] As a preferred option, the neural network parameters of each layer of Net A and Net B are set as follows:

[0026] The second layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 32, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME;

[0027] The third layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 64, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME;

[0028] The fourth layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 128, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME;

[0029] The fifth layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 256, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, the padding mode is set to SAME, and the size of the UpSampling layer is set to 2×2×2;

[0030] The sixth layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 128, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME; the size of the UpSampling layer is set to 2×2×2;

[0031] The seventh layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 64, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME; the size of the UpSampling layer is set to 2×2×2;

[0032] The eighth layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 32, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME; the size of the UpSampling layer is set to 2×2×2;

[0033] Ninth layer: The activation function is the softmax activation function.

[0034] As a preference, the training labels of Net A and Net B are prepared as follows:

[0035] The training labels of Net A are only 0 and 1 masks, and the training labels of Net B are obtained by performing distance transformation on the training labels of Net A.

[0036] The distance transformation steps are as follows: first, determine the blood vessel surface pixels. If any of the U, D, W, E, S, and N pixel points in the six neighborhoods of a blood vessel pixel P belongs to a background pixel, then the pixel P belongs to the blood vessel surface pixel; for a blood vessel intravascular pixel Q that is not a surface pixel, first traverse the pixels in a 27-domain with a size of 3×3×3 centered on Q, and if there is a blood vessel surface pixel, calculate the Euclidean distance; if there is no blood vessel surface pixel, continue to expand the domain range until there is a blood vessel surface pixel in the domain of 5×5×5, 7×7×7, 9×9×9, 11×11×11 until the Q point, and the Euclidean distance can be calculated; traverse all blood vessel intravascular pixels, and calculate the distance between each pixel and the nearest surface pixel.

[0037] Preferably, in step B5., after the heart image is input into the coronary artery recognition models A and B, the two models respectively output the predicted probability for each pixel, the two are added and averaged, and the output after threshold processing is the final segmentation result.

[0038] The substantial effects of the present invention are: 1) The present invention adopts a relay supervision strategy, outputs an auxiliary loss function after each round of upsampling, and finally supervises the training of the network through a three-layer loss function, thereby reducing the gradient explosion phenomenon in the deep learning network training. 2) The present invention uses two networks to learn the vascular segmentation ability from two different directions. Net A is responsible for learning the 0-1 label, and Net B is responsible for learning the distance map of the blood vessel. After the results of the two are fused, it is ensured that the final segmentation result has the shape of the blood vessel. The fused network has better segmentation ability and accuracy for small branches with complex topological structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flow chart of the first embodiment.

[0040] Figure 2 This is a schematic diagram of the CTA image normalization principle in Example 1.

[0041] Figure 3 Schematic diagram of the pixel 6 area in the first embodiment.

[0042] Figure 4 FIG. 1 is a schematic diagram of the distance without clustering in the first embodiment.

[0043] Figure 5 FIG. 1 is a schematic diagram of distance after clustering in Example 1. FIG.

[0044] Figure 6This is the coronary artery segmentation result of Example 1. DETAILED DESCRIPTION

[0045] The specific implementation of the present invention will be further described below through specific examples in conjunction with the accompanying drawings.

[0046] like Figure 1 As shown in the flowchart of , the present invention provides a coronary artery segmentation method combining distance transformation and deep learning, comprising the following steps:

[0047] Step 1: Obtain original CTA cardiac images;

[0048] Step 2: Image normalization preprocessing. Select the best CT value observation window, intercept the coronary artery region of interest in the CTA image, suppress non-cardiac tissues such as pulmonary veins, and improve the contrast between the target and the background. Figure 1 As shown in the figure, after reading the window width (WW) and window level (WL) information in the Dicom header file, the CT value is intercepted. The original CTA images within the range are normalized and the CT value is less than The pixel points are set to 0, and the CT value is greater than The pixel points are set to 1; the mapping process is shown in formula (1):

[0049]

[0050] In formula (1), y represents the gray value of each pixel after normalization, and x represents the CT value of each pixel in the original CTA image;

[0051] Step 3: Data augmentation: horizontal flip, vertical flip, random scaling, random elastic transformation, and random gamma correction.

[0052] Step 4: Build a deep learning network and use two deep neural networks NetA and Net B based on a three-layer relay supervision mechanism to train coronary artery CTA data.

[0053] 4-1Net A and Net B have the same structure, except for the input training labels. Both network structures include eight layers, of which the first layer is the input layer, the second to seventh layers are hidden layers, and the eighth layer is the output layer. The structure of each layer is as follows:

[0054] First layer: Input layer: input the three-dimensional CTA image matrix of the training set and the true label of each CTA matrix;

[0055] The second layer: hidden layer: contains two convolutional layers, each of which is followed by a normalization layer and an activation function layer, and finally a pooling layer is used for downsampling;

[0056] The third layer: hidden layer: contains two convolutional layers, each of which is followed by a normalization layer and an activation function layer, and finally a pooling layer is used for downsampling;

[0057] The fourth layer: hidden layer: contains two convolutional layers, each of which is followed by a normalization layer and an activation function layer, and finally a pooling layer is used for downsampling;

[0058] The fifth layer: hidden layer: contains two convolutional layers, each of which is followed by a normalization layer and an activation function layer, and finally an UpSampling layer for upsampling;

[0059] The sixth layer: hidden / relay output layer: contains a skip splicing layer, two convolutional layers, each convolution is followed by a normalization layer and an activation function layer, and finally an UpSampling layer, and outputs a loss function Loss1;

[0060] The seventh layer: hidden / relay output layer: contains a skip splicing layer, two convolutional layers, each convolution is followed by a normalization layer and an activation function layer, and finally an UpSampling layer, and outputs a loss function Loss2;

[0061] The eighth layer: hidden / relay output layer: contains a skip splicing layer, two convolutional layers, each convolution is followed by a normalization layer and an activation function layer, and finally an UpSampling layer, and outputs a loss function Loss3;

[0062] Ninth layer: Output layer: The final loss of the network is obtained by weighting Loss1, Loss2, and Loss3, as shown in formula (3):

[0063]

[0064] 4-2 The neural network parameters of each layer of Net A and Net B are set as follows:

[0065] The second layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 32, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME;

[0066] The third layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 64, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME;

[0067] The fourth layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 128, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME;

[0068] The fifth layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 256, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, the padding mode is set to SAME, and the size of the UpSampling layer is set to 2×2×2;

[0069] The sixth layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 128, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME; the size of the UpSampling layer is set to 2×2×2;

[0070] The seventh layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 64, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME; the size of the UpSampling layer is set to 2×2×2;

[0071] The eighth layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 32, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME; the size of the UpSampling layer is set to 2×2×2;

[0072] Ninth layer: The activation function is the softmax activation function. The convolution kernel size is in pixels.

[0073] 4-3 The training labels of Net A and Net B are produced as follows:

[0074] The training labels of Net A are masks of only 0 and 1, and the training labels of Net B are obtained by performing distance transformation on the training labels of Net A. First, determine the blood vessel surface pixels, as shown in Figure (2). If any of the U, D, W, E, S, and N pixels in the six neighborhoods of a blood vessel pixel P belongs to a background pixel, then the pixel P belongs to the blood vessel surface pixel. For a blood vessel pixel Q that is not a surface pixel, first traverse the pixels in the 27 domains centered on Q and of size 3×3×3. If there are blood vessel surface pixels, calculate the Euclidean distance. If there are no blood vessel surface pixels, continue to expand the domain range to 5×5×5, 7×7×7, 9×9×9, 11×11×11 until there are blood vessel surface pixels in the domain of point Q, and the Euclidean distance can be calculated. Traverse all blood vessel pixels and calculate the distance between each pixel and the nearest surface pixel. As Figure 4 and Figure 5 As shown, for the stability of network training, Figure 4 After obtaining the distances between the pixels in the blood vessels and the nearest surface pixels, these distances are clustered into Figure 5 For the three clusters shown, the darker the color, the closer it is to the center of the blood vessel, and vice versa, the closer it is to the surface of the blood vessel.

[0075] 4-4 Training of Net A and Net B

[0076] After the training of Net A and Net B is completed, the predicted probability of each pixel is output respectively, the two are added and averaged, and the output after threshold processing is the final segmentation result.

[0077] To demonstrate the feasibility of the above method, a specific coronary artery CTA image is used as an example below.

[0078] The experimental process is as follows: First, the original CTA image sequence (such as Figure 1 Image normalization can effectively suppress non-cardiac tissues such as pulmonary blood vessels and remove most of the cardiac tissues such as atria and ventricles with grayscale values ​​similar to those of coronary arteries, thereby reducing the impact of non-coronary tissues on subsequent coronary segmentation; distance map labels are made through 0-1 segmentation labels; Net A is trained using 0-1 segmentation labels, and Net B is trained using distance map labels. After training, the predicted probability maps of the two are weighted averaged, and the final output is as follows Figure 4 The segmentation results shown. The present invention tested the algorithm on 10 CTA data and used Dice, Jaccard, and MaxSD (Maximum Surface Distance) to evaluate the algorithm segmentation results. Dice = 0.8, Jaccard = 0.66, MaxSD = 5.81 on the 10 test data.

[0079] The present invention proposes an automatic segmentation method for coronary arteries based on CTA data. First, the non-coronary tissue is effectively suppressed through image normalization preprocessing, and the contrast between the coronary artery and the background is improved; secondly, the data is augmented by flipping, scaling, elastic transformation, gamma correction, etc.; finally, the coronary artery is segmented using a deep learning network combined with distance transformation. The results show that the present invention has better segmentation ability and accuracy for small branches with uneven grayscale and complex topological structure.

[0080] The above embodiment is only a preferred solution of the present invention and does not limit the present invention in any form. There are other variations and modifications without exceeding the technical solution described in the claims.

Claims

1. A coronary artery segmentation method based on deep learning, characterized by: The steps include: B1. Obtain original CTA cardiac images; B2. Normalize and preprocess the original CTA cardiac image in B1. Select the best CT value observation window, intercept the coronary artery area in the CTA image, suppress non-cardiac tissue, and improve the contrast between the target and the background; B3. Image data augmentation; B4. Construct a deep learning network and use two deep neural networks Net A and Net B based on a three-layer relay supervision mechanism but with different training labels to train coronary artery CTA data to obtain coronary artery recognition models A and B; Net A training labels only have 0 and 1 masks, and Net B training labels are obtained by performing distance transformation on Net A training labels; The network structures of Net A and Net B are: First layer: input layer; The second to fifth layers: hidden layers: contain two convolutional layers, each followed by a normalization layer and an activation function layer. The second to fourth layers are downsampled by a pooling layer, and the fifth layer is upsampled by an UpSampling layer. The sixth to eighth layers: hidden / relay output layer: contains a skip splicing layer, two convolutional layers, each convolution is followed by a normalization layer, an activation function layer, and finally an UpSampling layer. The sixth to eighth layers output loss functions Loss1, Loss2, and Loss3 respectively; Ninth layer: Output layer: The final Loss of the network is obtained by weighting Loss1, Loss2, and Loss3; B5. Use the coronary artery recognition model obtained in B4. to perform coronary artery segmentation.

2. The method for coronary artery segmentation based on deep learning according to claim 1, characterized in that: The method of image data augmentation in step B3 includes one or more of horizontal flipping, vertical flipping, random scaling, random elastic transformation, and random gamma correction.

3. The method for coronary artery segmentation based on deep learning according to claim 1, characterized in that: The network structures of Net A and Net B both include nine layers, wherein the first layer is the input layer, the second to eighth layers are hidden layers, and the ninth layer is the output layer. The layers are connected in sequence, and the structure is as follows: First layer: Input layer: Input the 3D CTA image matrix of the training set and the true label of each 3D CTA image matrix; Fifth layer: Hidden layer: The upper image data is normalized and activated after the first convolution in this layer, and then convoluted again and normalized and activated again, and finally up-sampled; Sixth layer: hidden / relay output layer: the UpSampling performs upsampling, the upper layer image data is normalized and activated after the first convolution in this layer, and then convolution is performed again and then normalized and activated, and finally upsampling is performed; The seventh layer: hidden / relay output layer: the UpSampling performs upsampling, the upper layer image data is normalized and activated after the first convolution in this layer, and then convolution is performed again and then normalized and activated, and finally upsampling is performed; Eighth layer: hidden / relay output layer: the UpSampling performs upsampling, the upper layer image data is normalized and activated after the first convolution in this layer, and then convolution is performed again and then normalized and activated, and finally upsampling is performed; Ninth layer: Output layer: The final Loss of the network is calculated by Loss1, Loss2, and Loss3 Loss=( Loss1+2*Loss2+4*Loss3) / 7 Weighted.

4. The method for coronary artery segmentation based on deep learning according to claim 3, characterized in that: The neural network parameters of each layer of NetA and Net B are set as follows: The second layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 32, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME; The third layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 64, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME; The fourth layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 128, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME; The fifth layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 256, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, the padding mode is set to SAME, and the size of the UpSampling layer is set to 2×2×2; The sixth layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 128, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME; the size of the UpSampling layer is set to 2×2×2; The seventh layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 64, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME; the size of the UpSampling layer is set to 2×2×2; The eighth layer: The convolution kernel size of the convolution layer is 3×3×3, the stride is 1, the number of convolution kernels is 32, the activation function is the Relu activation function, the pooling layer is the maximum pooling layer, the size is 2×2×2, the stride is 2, and the padding mode is set to SAME; the size of the UpSampling layer is set to 2×2×2; Ninth layer: The activation function is the softmax activation function.

5. The method for coronary artery segmentation based on deep learning according to claim 1, characterized in that : The distance transformation steps are as follows: first, determine the blood vessel surface pixels. If any of the U, D, W, E, S, and N pixels in the six neighborhoods of a blood vessel pixel P belongs to a background pixel, then the pixel P belongs to the blood vessel surface pixel; for a blood vessel pixel Q that is not a surface pixel, first traverse the pixels in the 27 domains centered on Q and with a size of 3×3×3. If there are blood vessel surface pixels, calculate the Euclidean distance. If there are no blood vessel surface pixels, continue to expand the domain range until there are blood vessel surface pixels in the domains of 5×5×5, 7×7×7, 9×9×9, 11×11×11 until the Q point, and the Euclidean distance can be calculated; All the pixels inside the blood vessel are traversed, and the distance between each pixel and the nearest surface pixel is calculated.

6. The method for coronary artery segmentation based on deep learning according to claim 1, characterized in that : In the step B5., after the heart image is input into the coronary artery recognition models A and B, the two models respectively output the predicted probability for each pixel, the two are added and averaged, and the output after threshold processing is the final segmentation result.

Images