CN113643317B - Coronary artery segmentation method based on deep geometric evolution model - Google Patents

Abstract

The invention discloses a coronary artery segmentation method based on a deep geometric evolution model, and relates to the field of deep learning image segmentation. Carry out new CCTA image training and other steps, and obtain two results. The pre-segmentation results are used to extract the skeleton, and the distance change results are used to calculate the radius of the coronary artery. The spherical model is constructed by the skeleton and the radius to obtain the evolved coronary artery and the pre-segmentation result. Obtain the final segmentation result; step 6, new data prediction; the present invention reconstructs the relationship between the coronary artery, the pixel and the spatial structure by using the skeleton and the radius of the coronary artery geometric structure of the CCTA data, improving the effectiveness of the coronary artery segmentation, and accurately giving the coronary artery segmentation. The doctor provides the basis for judgment.

Description

Coronary artery segmentation method based on deep geometric evolution model

technical field

The invention relates to the field of image segmentation of deep learning, in particular to a coronary artery segmentation method based on a deep geometric evolution model.

Background technique

Cardiovascular disease is one of the main diseases that cause human death; the "World Health Report" pointed out that the global annual mortality rate of cardiovascular-related diseases and its complications has exceeded the sum of cancer mortality; studies have shown that cardiovascular disease Most of the diseases are caused by coronary artery disease; early detection of coronary heart disease is of great significance for clinicians to formulate reasonable treatment plans; coronary computed tomography angiography (CCTA) is a commonly used clinical diagnosis of cardiovascular disease. It is also the gold standard for the diagnosis of cardiovascular disease; rapid, accurate and semi-automatic segmentation of coronary arteries from CCTA images by machine learning methods is a very meaningful auxiliary medical method, which plays a key role in the early detection of coronary heart disease. It will effectively provide doctors with a reliable basis for diagnosis.

Due to the particularity and complexity of the CCTA image itself, for example: (1) the coronary structure is complex, there are very branched small tubes, (2) the grayscale of the coronary artery is uneven, the contrast with the surrounding tissue is low, and the peripheral part of the blood vessel The boundary is blurred, (3) the coronary arteries contain various complex lesions; the segmentation algorithm for CCTA images has become a research hotspot in recent years; while the neural network-based method is to transform the segmentation problem into a classification problem, Nekovei uses BP The neural network performs segmentation training on blood vessels; in recent years, the Convolutional Neural Network (CNN) has achieved good results in the image segmentation network, of which the U-net network is currently the most widely used in the field of two-dimensional medical image segmentation. A network structure is proposed by Ronneberger on the basis of FCN. The network structure is characterized in that the network presents a "U" shape, and the number of upsampling layers and downsampling layers is equal; Nasr-Esfahani et al. The blood vessels and background areas are sent into the CNN network, which combines the global information and local information, and injects the canny edge detector for training, so as to obtain good results; and Jo proposes a method of selective feature mapping, using To segment the left front descending backbone of cardiovascular images; in order to overcome the limitation that U-Net has only one set of cascaded layers between the encoding and decoding blocks, Jun et al. quoted a new network of T-Net, adding pooling during encoding. Layers and upper layers are used to make the predicted masks more accurate.

Existing CCTA segmentation methods are based on traditional medical image segmentation methods, which are not only very cumbersome to design, but also difficult to guarantee the effectiveness of segmentation; neural network-based methods can automatically learn coronary arteries from CCTA images using neural networks. The essential characteristics of arteries, such as Unet or Unet++, do not comprehensively consider the full-text information of CCTA data, ignore the correlation between pixels, do not comprehensively consider the geometric information of blood vessels, and classify some small blood vessels in CCTA data. not good.

SUMMARY OF THE INVENTION

Based on the above problems, the present invention provides a coronary artery segmentation method based on a deep geometric evolution model, which can fully consider the spatial consistency of CCTA data, fully consider the geometric structure of coronary arteries, build a multi-object segmentation model, and use the skeleton and radius. To reconstruct coronary arteries, fully consider the relationship between pixels and spatial structures, so as to improve the effectiveness of coronary artery segmentation.

For solving the above technical problems, the technical scheme adopted in the present invention is as follows:

Coronary artery segmentation method based on deep geometric evolution model, including the following steps:

Step 1. Obtain the dataset;

Step 2, data preprocessing;

Step 3. Build a geometric evolution segmentation model;

The deep geometric evolution segmentation model includes an encoding module, a pre-segmentation module, and a distance change module. After the pre-processed data in step 2 is input into the deep geometric evolution segmentation model, the distance change results and pre-segmentation results are output; the pre-segmentation model extracts the skeleton, and the distance changes The module calculates the radius of the corresponding point of the skeleton; the coding module is composed of four coding blocks, each of which is composed of atrous convolution, batch normalization layer, and maximum pooling layer; the pre-segmentation module is composed of four segmentation blocks, each of which The segmentation block is composed of hole convolution, batch normalization layer, and upsampling layer; the distance change module is composed of four distance change blocks, and each distance change block is composed of hole convolution, batch normalization layer, and upsampling layer, and After the output result is obtained, it should be compared with the input image after calculating the distance change. The distance transform module (Distance transform, DT for short) is used to calculate the distance between each non-zero point in the image and its nearest zero point; the encoding module and the pre-segmentation module and the distance Each corresponding block of the change module is connected by skip connections, so that more high-frequency information is obtained. In order to obtain a better distance change result, the present invention uses the information of each segment of the pre-segmentation module to calculate the position attention module of the feature map as auxiliary information of the distance change module, so as to obtain richer feature information of the distance change module , better segmentation results will be obtained when the segmentation is evolved using the distance variation module.

The fourth step is to train and judge the CCTA image of the constructed multi-target segmentation model; the dice loss function is used for the pre-segmentation module, and the cross-entropy loss function is used for the distance change module; the fourth step is to train and evaluate the constructed multi-target segmentation model. Judgment; use the dice loss function for the pre-segmentation module, and use the cross-entropy loss function for the distance change module; use the joint loss function to train the multi-objective model, where the pre-segmentation module uses the activation function based on the dice coefficient, DT is a range between 0 and The value between 1, the goal of the network is to maximize the dice coefficient, the definition formula (3) is as follows:

(3)

(3)

表示预测结果，

表示标注图；in,

represents the prediction result,

Represents an annotation diagram;

Another distance change module uses the cross-entropy loss function, which is defined as follows:

(4)

(4)

表示距离变化离散后的类别数量，

表示CCTA数据上像素i等于距离变化离散后k的概率,如果像素i的真实类别等于k则取1,否则取0;

为CCTA数据上像素i属于距离变化离散后k的预测概率。in,

represents the number of categories after the distance change is discrete,

Represents the probability that the pixel i on the CCTA data is equal to k after the distance change is discrete, if the true category of the pixel i is equal to k , it takes 1, otherwise it takes 0;

is the predicted probability that the pixel i on the CCTA data belongs to k after the distance change is discrete.

The joint loss (6) of the final deep geometric evolutionary segmentation model is:

(6)

(6)

Step 5. Geometric evolution;

，在对骨架上提取每一个点坐标

，用对应冠状动脉的半径

和点坐标

通过球模型演化得到演化后的冠状动脉，演化后的冠状动脉与骨架提取的其它分割结果通过并集模块求并集得到最终的分割结果。The distance change results DT and pre-segmentation results are obtained from the CCTA images after training and evaluation. The results of the pre-segmentation module are extracted by the skeleton extraction module, and the distance change results are obtained through the distance change module to find the radius of the coronary arteries.

, extract the coordinates of each point on the pair of skeletons

, with the radius of the corresponding coronary artery

and point coordinates

The evolved coronary artery is obtained through the evolution of the spherical model, and the final segmentation result is obtained by the union of the evolved coronary artery and other segmentation results extracted from the skeleton.

Step 6. New data prediction;

是像素

是否是冠状动脉的概率；

是像素

属于

类的概率。对于一个像素

，它的半径

可以计算为

，它的骨架

可以利用skimage的skeletonize计算得到。对于骨架上的每一个点

，利用半径和坐标构建球模型演化得到冠脉，计算公式(1)如下:After the new CCTA data passes through the encoding module, the pre-segmentation module and the distance change module, two results (pre-segmentation result and distance change result) will be obtained, among which,

is the pixel

the probability of whether it is a coronary artery;

is the pixel

belong

class probability. for one pixel

, its radius

can be calculated as

, its skeleton

It can be calculated using the skeletonize of skimage. for every point on the skeleton

, using the radius and coordinates to construct the spherical model evolution to obtain the coronary artery, the calculation formula (1) is as follows:

(1)

(1)

可以由演化后的冠状动脉和骨架提取的其它分割结果求并集而得，计算公式(2)如下:Segmentation results of the last coronary artery

It can be obtained by the union of other segmentation results extracted from the evolved coronary artery and skeleton, and the calculation formula (2) is as follows:

(2)

(2)

Further, in the first step, the data set includes CCTA data and label data, and the label data needs to be manually labeled with 3DSlicer software for each case of CCTA data.

Further, the step 2 specifically includes the following:

The original CCTA data is read to obtain the window width and window level of the CCTA data, the window width represents the selected CT value range when displaying the image, the window level represents the center position of the image grayscale, and then the window width Set it to 1000, set the window level to 200, and finally normalize the CCTA data.

Further, in the third step, the encoding module adopts ReLU as the activation function, and the upsampling layer of the pre-segmentation module and the distance change module adopts bilinear interpolation.

Further, the specific construction process of the multi-target segmentation model is as follows:

Step 1: Divide the CCTA data into a plurality of sub-blocks through a block operation, and each sub-block has a corresponding size label, which is sent to the encoding module.

Step 2: Perform atrous convolution, batch normalization, and maximum pooling operations on the data in the encoding module to obtain the maximum receptive field.

的距离变化DT可以由每一个像素点

由公式(5)计算如下：Step 3: The features obtained in the encoding module are sent to the pre-segmentation module and the distance change module respectively. Both modules use bilinear interpolation to interpolate each point, and then perform convolution and batch normalization. Layer operation to complete the upsampling, in which the distance change module finally outputs the result and compares it with the input image after calculating the distance change, distance transform (DT). DT is used to calculate the distance from each non-zero point in the image to the nearest zero point. For CCTA image labels

The distance change DT can be determined by each pixel

It is calculated by formula (5) as follows:

(5)

(5)

是距离变化DT，

是像素点

和

的欧式距离，

表示该像素点在冠状动脉上。本发明通过将DT用one-hot编码的方式离散化到

类，离散后的DT由回归问题转为离散问题。in,

is the distance change DT,

is the pixel

and

the Euclidean distance,

Indicates that the pixel is on the coronary artery. The present invention discretizes DT to

class, the discrete DT is converted from a regression problem to a discrete problem.

Step 4: Connect the four encoding blocks in the encoding module, the four segmentation blocks in the pre-segmentation module, and the four pairs of distance changing blocks in the distance changing module to each corresponding block through a cross-layer connection structure, so as to obtain: More high frequency information.

Step 5: The information of each segmented block of the pre-segmentation module is used to calculate the position attention module of the feature map as auxiliary information for the distance transformation block of the distance transformation module.

Further, the step 4 specifically includes the following steps:

Step 41, data augmentation;

Step 42, network training;

Step 43, model evaluation.

Further, in the step 41, the CCTA data is augmented by methods of rotation, cropping and adding noise.

Further, in the step 42, the augmented CCTA data is divided into a plurality of sub-blocks through a block operation, and sent to the segmentation model in batches, the network sets the learning rate to 0.001, and the learning rate is every 20 learning iterations. After decaying ten times, the convolution weights are initialized with a Gaussian distribution, a training batch is set to 4, and the number of learning iterations is 200. The network training uses the BP backpropagation algorithm to calculate the gradient and update the weights. The network learning is for each batch. The parameters are updated every time. After each iterative learning, the pre-segmentation module judges the evaluation result of the segmentation, and the distance transformation module judges the evaluation result of the distance change. If the current error is less than the error of the previous iteration, the current multi-object segmentation model is saved, and then continue Train until the maximum number of iterations is reached.

Further, in the step 43, the joint loss is used for measurement, the overlapped part of the output CCTA data and the marked part of the real CCTA data is compared, and the multi-object segmentation model with the optimal evaluation index is saved.

Further, the step 5 specifically includes the following steps:

Step 51: Send the new CCTA data to the pre-trained model in step 4 for testing. The new CCTA data will obtain two results (segmentation result and distance change result) after passing through the encoding module, the pre-segmentation module and the distance change module.

Step 52: Perform skeleton extraction on the result of the pre-segmentation module.

Step 53: The radius of the coronary artery can be obtained by calculating argmax from the result of the distance change.

Step 54: For each point on the skeleton, use the corresponding radius and point coordinates to construct a spherical model for evolution to obtain an evolved coronary artery.

Step 55 , obtain a final coronary artery segmentation result by summing the evolved coronary artery and the result of the pre-segmentation module.

Compared with the prior art, the beneficial effects of the present invention are:

1. The present invention considers the coronary artery as a tubular structure by comprehensively analyzing the characteristics of the tubular arteries affected by CCTA, which can be obtained by evolution with a skeleton and a radius, and the radius can be obtained by distance transformation, so a deep geometric evolution segmentation model is constructed, thereby obtaining a more accurate segmentation model. for accurate segmentation results.

2. The present invention uses the information of each segmented block of the pre-segmentation module to calculate the location attention of the feature map as auxiliary information of the corresponding block of the distance variation module, thereby obtaining a more accurate distance variation result, thereby improving the accuracy of segmentation.

3. The present invention directly starts from the segmentation of three-dimensional data, which can relatively well ensure the correlation before and after the three-dimensional data of CCTA, and adopts the corresponding evaluation index, so that the result of the segmentation is more accurate, and it provides a better diagnosis for subsequent doctors. basis.

4. For the model that has been trained, it can quickly segment the coronary artery evolution to realize the data segmentation of batch CCTA, which can realize unattended batch operation, and the speed is fast, saving the manpower and material resources of coronary artery labeling, and providing doctors with auxiliary diagnosis. stronger basis.

Description of drawings

Fig. 1 is the flow chart of the deep geometric evolution segmentation model constructed for this implementation;

FIG. 2 is a frame diagram of a deep geometric evolution segmentation model constructed for this implementation.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings. Embodiments of the present invention include but are not limited to the following examples.

As shown in Figure 1, the coronary artery segmentation method based on the deep geometric evolution model includes the following steps:

Step 1. Obtain the dataset;

Step 2, data preprocessing;

Step 3: Build a deep geometric evolution segmentation model;

The deep geometric evolution segmentation model includes an encoding module, a pre-segmentation module, and a distance change module. After the pre-processed data in step 2 is input into the deep geometric evolution segmentation model, the distance change results and pre-segmentation results are output; the pre-segmentation model extracts the skeleton, and the distance changes The module calculates the radius of the corresponding point of the skeleton; the coding module is composed of four coding blocks, each of which is composed of atrous convolution, batch normalization layer, and maximum pooling layer; the pre-segmentation module is composed of four segmentation blocks, each of which The segmentation block is composed of hole convolution, batch normalization layer, and upsampling layer; the distance change module is composed of four distance change blocks, and each distance change block is composed of hole convolution, batch normalization layer, and upsampling layer, and After the output result is obtained, it should be compared with the input image after calculating the distance change. The distance transform module (Distance transform, DT for short) is used to calculate the distance between each non-zero point in the image and its nearest zero point; the encoding module and the pre-segmentation module and the distance Each corresponding block of the change module is connected by skip connections, so that more high-frequency information is obtained. In order to obtain a better distance change result, the present invention uses the information of each segment of the pre-segmentation module to calculate the position attention module of the feature map as auxiliary information of the distance change module, so as to obtain richer feature information of the distance change module , better segmentation results will be obtained when the segmentation is evolved using the distance variation module.

The present invention discretizes the DT into k classes by one-hot coding, and the discretized DT is transformed from a regression problem to a discrete problem.

Step 4: Train and judge the constructed multi-target segmentation model; use the dice loss function for the pre-segmentation module, and use the cross-entropy loss function for the distance change module; use the joint loss function to train the multi-target model, where the pre-segmentation module uses the The activation function of the dice coefficient, DT is a value between 0 and 1, the goal of the network is to maximize the dice coefficient, and formula (3) is defined as follows:

(3)

(3)

表示预测结果，

表示标注图；in,

represents the prediction result,

Represents an annotation diagram;

Another distance change module uses the cross-entropy loss function, which is defined as follows (4):

(4)

(4)

表示距离变化离散后的类别数量，

表示CCTA数据上像素i等于距离变化离散后k的概率,如果像素i的真实类别等于k则取1,否则取0;

为CCTA数据上像素i属于距离变化离散后k的预测概率。in,

represents the number of categories after the distance change is discrete,

Represents the probability that the pixel i on the CCTA data is equal to k after the distance change is discrete, if the true category of the pixel i is equal to k , it takes 1, otherwise it takes 0;

is the predicted probability that the pixel i on the CCTA data belongs to k after the distance change is discrete.

The joint loss (6) of the final deep geometric evolutionary segmentation model is:

(6)

(6)

，在对骨架上提取每一个点坐标

，用对应冠状动脉的半径

和点坐标

通过球模型演化得到演化后的冠状动脉，演化后的冠状动脉与骨架提取的其它分割结果通过并集模块求并集得到最终的分割结果；Step 5: Geometric evolution; obtain the distance change result DT and the pre-segmentation result from the CCTA images after training and evaluation, perform the skeleton extraction on the result of the pre-segmentation module through the skeleton extraction module, and obtain the coronary artery size for the distance change result through the distance change module. radius

, extract the coordinates of each point on the pair of skeletons

, with the radius of the corresponding coronary artery

and point coordinates

The evolved coronary artery is obtained through the evolution of the spherical model, and the final segmentation result is obtained by the union of the evolved coronary artery and other segmentation results extracted from the skeleton;

Step 6. New data prediction;

是像素

是否是冠状动脉的概率；

是像素

属于

类的概率。对于一个像素

，它的半径

可以计算为

，它的骨架

可以利用skimage的skeletonize计算得到。对于骨架上的每一个点

，利用半径和坐标构建球模型演化得到冠脉，计算公式(1)如下：After the new CCTA data passes through the encoding module, the pre-segmentation module and the distance change module, two results (pre-segmentation result and distance change result) will be obtained, among which,

is the pixel

the probability of whether it is a coronary artery;

is the pixel

belong

class probability. for one pixel

, its radius

can be calculated as

, its skeleton

It can be calculated using the skeletonize of skimage. for every point on the skeleton

, using the radius and coordinates to construct the spherical model evolution to obtain the coronary artery, the calculation formula (1) is as follows:

(1)

(1)

可以由演化后的冠状动脉和骨架提取的其它分割结果求并集而得，计算公式（2）如下：Segmentation results of the last coronary artery

It can be obtained from the union of other segmentation results extracted from the evolved coronary artery and the skeleton, and the calculation formula (2) is as follows:

(2)

(2)

Further, in the first step, the data set includes CCTA data and labels, and the label data needs to be manually labeled with 3D Slicer software for each case of CCTA data.

Further, the step 2 specifically includes the following:

The original CCTA data is read to obtain the window width and window level of the CCTA data, the window width represents the selected CT value range when displaying the image, the window level represents the center position of the image grayscale, and then the window width Set it to 1000, set the window level to 200, and finally normalize the CCTA data.

Further, in the third step, the encoding module adopts ReLU as the activation function, and the upsampling layer of the pre-segmentation module and the distance change module adopts bilinear interpolation.

Further, the specific construction process of the multi-target segmentation model is as follows:

Step 1: Divide the CCTA data into a plurality of sub-blocks through a block operation, and each sub-block has a corresponding size label, which is sent to the encoding module.

Step 2: Perform atrous convolution, batch normalization, and maximum pooling operations on the data in the lower encoding module to obtain the maximum receptive field.

的距离变化DT可以由每一个像素点

由公式(5)计算如下：Step 3: The features obtained in the encoding module are sent to the pre-segmentation module and the distance change module respectively. Both modules use bilinear interpolation to interpolate each point, and then perform convolution and batch normalization. layer operations to complete the upsampling. Among them, the final output result of the distance change module should be compared with the input image after calculating the distance change, the distance change (Distance transform, referred to as DT). DT is used to calculate the distance from each non-zero point in the image to the nearest zero point. For CCTA image labels

The distance change DT can be determined by each pixel

It is calculated by formula (5) as follows:

(5)

(5)

是距离变化DT，

是像素点

和

的欧式距离，

表示该像素点在冠状动脉上。本发明通过将DT用one-hot编码的方式离散化到

类，离散后的DT由回归问题转为离散问题。in

is the distance change DT,

is the pixel

and

the Euclidean distance,

Indicates that the pixel is on the coronary artery. The present invention discretizes DT to

class, the discrete DT is converted from a regression problem to a discrete problem.

Step 4: Connect the four encoding blocks in the encoding module, the four segmentation blocks in the pre-segmentation module, and the four pairs of distance changing blocks in the distance changing module to each corresponding block through a cross-layer connection structure, so as to obtain: More high frequency information.

Step 5: The information of each segmented block of the pre-segmentation module is used to calculate the location attention of the feature map as auxiliary information of the distance transformation block for the distance transformation module.

Further, the step 4 specifically includes the following steps:

Step 41, data augmentation;

Step 42, network training;

Step 43, model evaluation.

Further, in the step 41, the CCTA data is augmented by methods of rotation, cropping and adding noise.

Further, in the step 42, the augmented CCTA data is divided into a plurality of sub-blocks through a block operation, and sent to the segmentation model in batches, the network sets the learning rate to 0.001, and the learning rate is every 20 learning iterations. After decaying ten times, the convolution weights are initialized with a Gaussian distribution, a training batch is set to 4, and the number of learning iterations is 200. The network training uses the BP backpropagation algorithm to calculate the gradient and update the weights. The network learning is for each batch. The parameters are updated every time. After each iterative learning, the pre-segmentation module judges the evaluation result of the segmentation, and the distance transformation module judges the evaluation result of the distance change. If the current error is less than the error of the previous iteration, the current multi-object segmentation model is saved, and then continue Train until the maximum number of iterations is reached.

Further, in the step 43, the joint loss is used for measurement, the overlapped part of the output CCTA data and the marked part of the real CCTA data is compared, and the multi-object segmentation model with the optimal evaluation index is saved.

Further, the step 5 specifically includes the following steps:

Step 51: Send the new CCTA data to the pre-trained model in step 4 for testing. The new CCTA data will obtain two results (segmentation result and distance change result) after passing through the encoding module, the pre-segmentation module and the distance change module.

Step 52: Perform skeleton extraction on the result of the pre-segmentation module.

Step 53: The radius of the coronary artery can be obtained by calculating argmax from the result of the distance change.

Step 54: For each point on the skeleton, use the corresponding radius and point coordinates to construct a spherical model for evolution to obtain an evolved coronary artery.

The above is an embodiment of the present invention. The above examples and the specific parameters in the examples are only to clearly describe the inventor's invention verification process, not to limit the scope of the patent protection of the present invention. The scope of the patent protection of the present invention is still based on the claims. Equivalent structural changes made in the contents of the description and drawings of the present invention shall be included within the protection scope of the present invention.

Claims (8)

1. the coronary artery segmentation method based on the deep geometric evolution model, is characterized in that, comprises the steps: Step 1. Obtain the dataset; Step 2, data preprocessing; Step 3: Build a deep geometric evolution segmentation model; The deep geometric evolution segmentation model includes an encoding module, a pre-segmentation module, and a distance change module. After the pre-processed data in step 2 is input into the deep geometric evolution segmentation model, the distance change results and pre-segmentation results are output; the pre-segmentation model extracts the skeleton and the distance changes. The module obtains the radius of the corresponding point of the skeleton; the encoding module is composed of four encoding blocks, each of which is composed of a hole convolution layer, a batch normalization layer and a maximum pooling layer; the pre-segmentation module is composed of four segmentation blocks, Each segmentation block is composed of a hole convolution layer, a batch normalization layer and an upsampling layer; the distance change module is composed of four distance change blocks, and each distance change block is composed of a hole convolution layer, a batch normalization layer and an upper sampling layer. The sampling layer is formed, and the result predicted by the distance change module should be compared with the result of calculating the distance change of the input image. The distance change module is used to calculate the distance between each non-zero point in the image and the nearest zero point; Each coding block is connected with each corresponding block of the pre-segmentation module and the distance change module through skip connections, so that the pre-segmentation module and the distance change module can obtain more high-frequency information to improve the spatial information. The output is to predict the radius corresponding to each skeleton point; Step 4: Train and judge the CCTA image of the constructed deep geometric evolution segmentation model; use the dice loss function for the pre-segmentation module, and use the cross-entropy loss function for the distance change module; Step 5. Geometric evolution; Use the constructed multi-target segmentation model to train new CCTA images, obtain distance change results and pre-segmentation results, extract the skeleton from the results of the pre-segmentation module, and calculate the radius of the coronary artery from the distance change results. For each point of , use the radius and point coordinates of the corresponding coronary artery to construct a spherical model to evolve the evolved coronary artery, and the pre-segmentation result extracted from the evolved coronary artery and the skeleton is obtained through the union module to obtain the final segmentation result; Step 6. New data prediction; In the steps 5-6, the spherical model obtains the coronary artery by formula (1),

is the pixel

the probability of whether it is a coronary artery;

is the pixel

belong

class probability, for a pixel

, its radius

can be calculated as

, its skeleton

It can be calculated by skimage's skeletonize, for each point on the skeleton

, the coronary artery is obtained by radius evolution, and the calculation formula (1) is as follows:

(1) in,

represents the radius of the coronary artery,

Represents each point coordinate extracted on the skeleton, U represents the set of all point coordinates on the skeleton; Finally, the segmentation results of the coronary arteries

It can be obtained from the union of the evolved coronary artery and the segmentation result, and the calculation formula (2) is as follows:

(2). 2. The coronary artery segmentation method based on the deep geometric evolution model according to claim 1, wherein the step 2 specifically comprises the following: The data set includes CCTA data, and for each case of CCTA data, a label needs to be manually set, the original CCTA data is read, and the window width and window level of the CCTA data are obtained, and the window width represents the selected CT when displaying the image. value range, the window level represents the center position of the grayscale of the image, then the window width is set to 1000, the window level is set to 200, and finally the CCTA data is normalized. 3. the coronary artery segmentation method based on deep geometric evolution model according to claim 1, is characterized in that: in described step 3, coding module adopts ReLU as activation function, and the upsampling layer of pre-segmentation module, distance variation module adopts Bilinear interpolation; in the fourth step, a joint loss function is used to train a multi-target model, wherein the pre-segmentation module uses an activation function based on the dice coefficient, DT is a value between 0 and 1, and the goal of the network is To maximize the dice coefficient, the following formula (3) is defined:

(3) in,

represents the prediction result,

Represents an annotation diagram; Another distance change module uses the cross-entropy loss function, which is defined as follows:

(4) in,

represents the number of categories after the distance change is discrete,

Represents the probability that the pixel i on the CCTA data is equal to k after the distance change is discrete, if the true category of the pixel i is equal to k , it takes 1, otherwise it takes 0;

is the predicted probability that the pixel i on the CCTA data belongs to k after the distance change is discrete. 4. the coronary artery segmentation method based on deep geometric evolution model according to claim 3, is characterized in that: the concrete construction process of described segmentation model is as follows: Step 1: Divide the CCTA data into multiple sub-blocks through the block operation, and each sub-block has a corresponding size label, which is sent to the encoding module; Step 2: Perform atrous convolution, batch normalization, and maximum pooling operations on the data in the encoding module to obtain the maximum receptive field; Step 3: The features obtained in the encoding module are sent to the pre-segmentation module and the distance change module respectively. Both modules use bilinear interpolation to interpolate each point, and then perform convolution and batch normalization. Layer operation to complete the upsampling, where the distance change module finally outputs the result and compares it with the input image after calculating the distance change, and the distance change DT is used to calculate the closest distance between each non-zero point in the image and its own zero point; Step 4: After each sub-block passes through four encoding modules and upsampling layers, the final output layer has two branches, one branch is the segmentation result, and the other branch calculates the distance change DT, DT is used to calculate each non-zero point in the image The distance to the nearest zero point to yourself, for CCTA image labels

The distance change DT can be determined by each pixel

It is calculated by formula (5) as follows:

(5) in,

is the distance change DT,

is the pixel

and

the Euclidean distance,

Indicates that the pixel is on the coronary artery, discretized by one-hot encoding of DT to

class, the discrete DT is converted from a regression problem to a discrete problem; Step 5: On the model prediction after training, the constructed segmentation model is used to segment the new CCTA data. The new CCTA data will get two results after passing through the encoding module, the pre-segmentation module and the distance change module, and calculate two results respectively. result of the branch. 5. The coronary artery segmentation method based on the deep geometric evolution model according to claim 1, wherein the step 4 specifically comprises the following steps: Step 41, data augmentation; Step 42, network training; Step 43, model evaluation. 6 . The coronary artery segmentation method based on the deep geometric evolution model according to claim 5 , wherein in the step 41 , the CCTA data is augmented by methods of rotation, cropping, and noise addition. 7 . 7. The coronary artery segmentation method based on the deep geometric evolution model according to claim 5, characterized in that: in the step 42, the augmented CCTA data is divided into a plurality of sub-blocks by a block operation, and is divided into batches. The network sets the learning rate to 0.001, and the learning rate decays ten times after every 20 learning iterations. The convolution weights are initialized with Gaussian distribution. The first training batch is set to 4, and the number of learning iterations is 200. The network training uses the BP backpropagation algorithm to calculate the gradient and update the weights. The network learning updates the parameters once for each batch. After each iteration of learning, the segmentation model judges the evaluation result of the segmentation. If the current error is smaller than the error of the previous iteration , save the current segmentation model and continue training until the maximum number of iterations is reached. 8. The coronary artery segmentation method based on deep geometric evolution model according to claim 5, it is characterized in that: in described step 43, utilize joint loss to measure, compare the overlapping part of output CCTA data and real CCTA data labeling part , and save the segmentation model with the best evaluation index.

Images