A method and system for carotid artery segmentation in medical images based on deep learning

Abstract

The present invention relates to a deep learning-based medical image carotid artery segmentation algorithm and system. The method acquires and preprocesses raw neck medical image data, then inputs it into a pretrained medical image carotid artery detection network for vessel localization, generating a carotid artery localization frame. The portion of the neck medical image corresponding to the localization frame is then input into a pretrained medical image carotid artery segmentation network for vessel segmentation, generating a carotid artery segmentation result. Finally, the carotid artery segmentation result and the raw neck medical image data are visualized. Compared to existing technologies, the present invention offers the advantage of high segmentation accuracy.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method and system for medical image carotid artery segmentation based on deep learning. Background Art

[0002] Blood vessel extraction from carotid CTA (Computed Tomography Angiography) images primarily involves extracting the central path of the carotid artery and segmenting the carotid artery. This is a critical step in accurately displaying and quantifying the carotid artery in complex datasets. It is also a prerequisite for the diagnosis and surgical planning of cerebrovascular diseases such as stenosis, plaques, and aneurysms, and is crucial for the evaluation of clinical diagnostic results.

[0003] Existing studies on automatic carotid artery segmentation based on CTA images have shown that a two-dimensional convolutional network (D-SEA UNet) trained using two-dimensional multi-layer images can achieve carotid artery segmentation. Another study on automatic carotid artery segmentation based on MRA images has also shown that a two-dimensional convolutional network (UNet) trained using two-dimensional images can achieve carotid artery segmentation. These studies all rely solely on two-dimensional convolutional networks for segmentation and lack the ability to learn three-dimensional information.

[0004] After searching, a Chinese patent application, published as CN114519722A, disclosed a method for carotid artery extraction based on a convolutional neural network. The proposed model extracts and tracks patches based on a given seed point along the carotid artery's central path, then feeds this into a three-dimensional neural network to achieve carotid artery segmentation. However, this technology has the limitation of requiring manual seed points, preventing a fully automated carotid artery segmentation process. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a method and system for medical image carotid artery segmentation based on deep learning with fully automatic positioning and segmentation and high segmentation efficiency.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] According to a first aspect of the present invention, a method for segmenting carotid arteries in medical images based on deep learning is provided, the method comprising the following steps:

[0008] Step S1, data acquisition: obtaining original cervical medical imaging data I1;

[0009] Step S2, data preprocessing: After the original neck medical image data I1 undergoes a first preprocessing process P1, a unified standard second neck medical image data I2 is obtained; the second neck medical image data I2 undergoes a second preprocessing process P2 to obtain normalized third neck medical image data I3;

[0010] Step S3, blood vessel positioning: inputting the third cervical medical image data I3 into the pre-trained first attention-based U-Net network to perform blood vessel positioning, and obtaining the blood vessel positioning coordinates A1;

[0011] Step S4, blood vessel segmentation: The blood vessel positioning coordinates A1, the second neck medical image data I2, and the third neck medical image data I3 are subjected to the first data post-processing process P3 to obtain fourth neck medical image data I4; the fourth neck medical image data I4 is input into the pre-trained second attention-based U-Net network for blood vessel segmentation, and a blood vessel segmentation result A2 is obtained;

[0012] Step S5, visualization processing: combining the blood vessel segmentation result A2, the original neck medical image data I1 and the second neck medical image data I2, and undergoing the second data post-processing process P4, obtains the neck medical image and the carotid artery blood vessel segmentation combined image I5 for visualization display.

[0013] Preferably, the original cervical medical image data I1 in step S1 is CTA data.

[0014] Preferably, the first pre-processing process P1 in step S2 is specifically as follows:

[0015] Step S201: resampling the original neck medical image I1 with different pixel spacing and image size; wherein the resampling interpolation mode is nearest neighbor interpolation;

[0016] Step S202: performing center cropping or peripheral padding on the resampled cervical medical image to obtain second cervical medical image data I2 of a unified standard; wherein the Z-axis direction is not processed in the center cropping, and the peripheral padding is performed with a constant value, and the constant value is the minimum value of the input data;

[0017] The second pre-processing process P2 is specifically as follows:

[0018] Step S211: downsampling the second cervical medical image data I2 to a size of H1×W1×Z1, wherein the interpolation mode in the downsampling is nearest neighbor interpolation;

[0019] Step S212: Normalize the downsampled image data. Specifically, linearly map the voxel points with intensity values ​​in the interval [a, b] to the interval [0, 1]. Set the voxel points with intensity values ​​less than a to 0. Set the voxel points with intensity values ​​greater than b to 0.

[0020] Preferably, the first data post-processing process P3 in step S4 is specifically as follows:

[0021] Step S401, extracting the image region R1 corresponding to the blood vessel location coordinates A1 from the third cervical medical image data I3;

[0022] Step S402: resample the image region R1 to obtain coordinates A'1 in the same space as the second cervical medical image data I2; wherein the resampling interpolation mode is nearest neighbor interpolation;

[0023] Step S403 : extracting the image region corresponding to the coordinate A′ 1 from the second cervical medical image data I 2 , which is the fourth cervical medical image data I 4 .

[0024] Preferably, the U-Net network based on the attention mechanism in step S3 or step S4 includes an attention gate module, a fusion module, and a first convolution module, a first activation module, a first downsampling module, a second convolution module, a second activation module, a second downsampling module, a third convolution module, a third activation module, a third downsampling module, a fourth convolution module and a fourth activation module, a first upsampling module, a fifth convolution module, a fifth activation module, a second upsampling module, a sixth convolution module, a sixth activation module, a third upsampling module, a seventh convolution module and a seventh activation module connected in sequence;

[0025] The output of the third activation module and the output of the fourth activation module are respectively input into the attention gate module, and the output of the attention gate module and the output of the first upsampling module are fused through the fusion module as the input of the fifth convolution module; the output of the second activation module and the output of the fifth activation module are respectively input into the attention gate module, and the output of the attention gate module and the output of the second upsampling module are fused through the fusion module as the input of the sixth convolution module; the output of the first activation module and the output of the sixth activation module are respectively input into the attention gate module, and the output of the attention gate module and the output of the third upsampling module are fused through the fusion module as the input of the fifth convolution module.

[0026] Preferably, the attention gate module includes an eighth convolution module, a ninth convolution module, an eighth activation module, a tenth convolution module and a ninth activation module; the input of the eighth convolution module is a gate signal, and the input of the ninth convolution module is a feature map to be processed. After superimposing the output signals of the eighth convolution module and the ninth convolution module, the output feature map results are sequentially passed through the eighth activation module, the tenth convolution module and the ninth activation module, and the corresponding position elements of the output feature map results are multiplied with the original feature map to be processed to obtain the final processed feature map.

[0027] Preferably, the pre-training process of the U-Net network based on the attention mechanism in step S3 or step S4 includes:

[0028] 1) Obtain training set, validation set data and corresponding labels;

[0029] 2) Input the training set data into the network as a channel and set the hyperparameter values; the hyperparameter values ​​include the batch size, maximum training cycle, initial learning rate, learning rate optimizer, and loss function;

[0030] 3) Calculate the network output results of the training set data and the loss function value of the corresponding label;

[0031] 4) Adjust the parameter values ​​in the network model according to the loss function results;

[0032] 5) After the training set data has been input for a certain period of time, the validation set data and labels are input;

[0033] 6) Calculate the network output results of the validation set data and the loss function value of the corresponding label;

[0034] 7) Adjust the hyperparameter values ​​of the network model according to the loss function results;

[0035] 8) Repeat steps 2)-7) until the loss function values ​​in 3) and 6) converge and are less than the set value, and then end the training.

[0036] Preferably, the loss function is a Tversky loss function with Sigmoid.

[0037] Preferably, the step S5 specifically includes the following sub-steps:

[0038] Step S51, restoring the blood vessel segmentation result A2 to the coordinate A'1 region in the second cervical medical image data I2;

[0039] Step S52 , resampling and center cropping or surrounding padding the blood vessel segmentation result A2 to generate a blood vessel segmentation result A′2 in the same space as the original neck medical image data I1 ;

[0040] Step S53 : combining the blood vessel segmentation result A′ 2 with the original neck medical image data I 1 to form a neck medical image and carotid artery segmentation combined image I 5 for visualization.

[0041] According to a second aspect of the present invention, a deep learning-based medical image carotid artery segmentation system is provided, using any of the methods described above, the system comprising:

[0042] A data acquisition module is used to obtain original neck medical imaging data I1;

[0043] A data preprocessing module is configured to obtain standardized second cervical medical image data I2 after subjecting the original cervical medical image data I1 to a first preprocessing process P1, and obtain normalized third cervical medical image data I3 after subjecting the second cervical medical image data I2 to a second preprocessing process P2;

[0044] Blood vessel positioning module: input the third neck medical image data I3 into the pre-trained first attention-based U-Net network to perform blood vessel positioning and obtain the blood vessel positioning coordinates A1;

[0045] a blood vessel segmentation module, configured to use a pre-trained second attention-based U-Net network to perform blood vessel segmentation on fourth neck medical image data I4 obtained by performing the first data post-processing process P3 on the blood vessel positioning coordinates A1, the second neck medical image data I2, and the third neck medical image data I3, to obtain a blood vessel segmentation result A2;

[0046] The visualization module is used to visualize the neck medical image and the carotid artery segmentation combined image I5 obtained from the blood vessel segmentation result A2, the original neck medical image data I1 and the second neck medical image data I2 after the second data post-processing process P4.

[0047] Compared with the prior art, the present invention has the following advantages:

[0048] 1) By using a convolutional neural network with an attention mechanism to train data, it learns and captures features at different levels. By finding certain connections within the data, it builds a prediction model and maps the input to the output (label or predicted value), which can achieve better segmentation results.

[0049] 2) By adopting the method of the present invention, a fully automatic carotid artery positioning and segmentation task that meets clinical needs is achieved, assisting clinicians to quickly complete the fully automatic carotid artery positioning and segmentation process, and solving the problem of labor-intensive and time-consuming medical image visual inspection;

[0050] 3) The present invention uses a first preprocessing step P1 to convert images scanned by different hospitals and different equipment into the same pixel space, which helps to eliminate technical differences and avoid the uncertainty caused by technical variables from interfering with subsequent steps;

[0051] 4) The present invention adopts a second preprocessing process P2, which can improve the network processing speed while retaining the blood vessel position information through the downsampling step; the normalization step avoids the problem of subsequent network training non-convergence;

[0052] 5) The first data post-processing step P3 of the present invention utilizes the existing position information to remove other invalid information and improve the blood vessel segmentation effect.

[0053] 6) The second data post-processing process P4 of the present invention is only the inverse process of P1 and P2, which can better display the blood vessel segmentation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 Flowchart of the deep learning carotid artery segmentation method of the present invention;

[0055] Figure 2 Schematic diagram of the U-Net model with attention mechanism added;

[0056] Figure 3 Schematic diagram of the attention gate module;

[0057] Figure 4 is a flow chart of the method of the present invention;

[0058] Figure 5 Result diagrams of the cervical medical imaging data I1 to I4 in the embodiment;

[0059] Figure 6 is the blood vessel segmentation result A2 in the embodiment;

[0060] Figure 7 This is the neck medical image and the carotid artery segmentation combined image I5 in the embodiment. DETAILED DESCRIPTION

[0061] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0062] Example

[0063] like Figure 1 and Figure 4As shown, this embodiment provides a method for segmenting carotid arteries in medical images based on deep learning, which includes the following steps:

[0064] Step S1, data acquisition: obtaining original cervical medical image data I1; wherein the original cervical medical image data I1 is CTA data;

[0065] Step S2, data preprocessing: After the original neck medical image data I1 undergoes the first preprocessing process P1, a unified standard second neck medical image data I2 is obtained; the second neck medical image data I2 undergoes the second preprocessing process P2 to obtain normalized third neck medical image data I3, specifically:

[0066] The first pre-processing process P1 is specifically as follows:

[0067] Step S201: resampling the original neck medical image I1 with different pixel spacing and image size; wherein the resampling interpolation mode is nearest neighbor interpolation;

[0068] Step S202: Perform center cropping or peripheral padding on the resampled cervical medical image to obtain a second cervical medical image data I2 of a unified standard. The standardized size is H0×W0×Z0, where Z0 is a positive integer. In the center cropping, no processing is performed in the Z-axis direction. The peripheral padding is performed with a constant value, and the constant value is the minimum value of the input data. In this embodiment, H=512 and D=512.

[0069] The second pre-processing process P2 is specifically as follows:

[0070] Step S211 : downsample the second cervical medical image data I2 to a size of H1×W1×Z1, wherein the interpolation mode in the downsampling is nearest neighbor interpolation; in this embodiment, H1×W1×Z1 is 128×128×128.

[0071] Step S212: Normalize the downsampled image data. Specifically, linearly map voxels with intensity values ​​in the interval [a, b] to the interval [0, 1]. Voxels with intensity values ​​less than a are set to 0; voxels with intensity values ​​greater than b are set to 1. In this embodiment, a = -300, b = 1200.

[0072] Step S3, blood vessel positioning: inputting the third cervical medical image data I3 into the pre-trained first attention-based U-Net network to perform blood vessel positioning, and obtaining the blood vessel positioning coordinates A1;

[0073] Step S4, blood vessel segmentation: The blood vessel positioning coordinates A1, the second neck medical image data I2, and the third neck medical image data I3 are subjected to the first data post-processing process P3 to obtain fourth neck medical image data I4; the fourth neck medical image data I4 is input into the pre-trained second attention-based U-Net network for blood vessel segmentation, and a blood vessel segmentation result A2 is obtained;

[0074] The first data post-processing process P3 is specifically as follows:

[0075] Step S401, extracting the image region R1 corresponding to the blood vessel location coordinates A1 from the third cervical medical image data I3;

[0076] Step S402: resample the image region R1 to obtain coordinates A'1 in the same space as the second cervical medical image data I2; wherein the resampling interpolation mode is nearest neighbor interpolation;

[0077] Step S403 : extracting the image region corresponding to the coordinate A′ 1 from the second cervical medical image data I 2 , which is the fourth cervical medical image data I 4 .

[0078] Step S5, visualization processing: combining the blood vessel segmentation result A2, the original neck medical image data I1, and the second neck medical image data I2, and undergoing a second data post-processing process P4, which is merely the inverse process of P1 and P2, to obtain a neck medical image and a carotid artery blood vessel segmentation combined image I5 for visualization display, specifically including the following sub-steps:

[0079] Step S51, restoring the blood vessel segmentation result A2 to the coordinate A'1 region in the second cervical medical image data I2;

[0080] Step S52 , resampling and center cropping or surrounding padding the blood vessel segmentation result A2 to generate a blood vessel segmentation result A′2 in the same space as the original neck medical image data I1 ;

[0081] Step S53 : combining the blood vessel segmentation result A′ 2 with the original neck medical image data I 1 to form a neck medical image and carotid artery segmentation combined image I 5 for visualization.

[0082] like Figure 2As shown, the U-Net network based on the attention mechanism in steps S3 and S4 includes an attention gate module, a fusion module, and a first convolution module, a first activation module, a first down-sampling module, a second convolution module, a second activation module, a second down-sampling module, a third convolution module, a third activation module, a third down-sampling module, a fourth convolution module and a fourth activation module, a first up-sampling module, a fifth convolution module, a fifth activation module, a second up-sampling module, a sixth convolution module, a sixth activation module, a third up-sampling module, a seventh convolution module and a seventh activation module connected in sequence;

[0083] The output of the third activation module and the output of the fourth activation module are respectively input into the attention gate module, and the output of the attention gate module and the output of the first upsampling module are fused through the fusion module as the input of the fifth convolution module; the output of the second activation module and the output of the fifth activation module are respectively input into the attention gate module, and the output of the attention gate module and the output of the second upsampling module are fused through the fusion module as the input of the sixth convolution module; the output of the first activation module and the output of the sixth activation module are respectively input into the attention gate module, and the output of the attention gate module and the output of the third upsampling module are fused through the fusion module as the input of the fifth convolution module.

[0084] like Figure 3 As shown, the attention gate module includes an eighth convolution module, a ninth convolution module, an eighth activation module, a tenth convolution module and a ninth activation module; the input of the eighth convolution module is a gate signal, and the input of the ninth convolution module is a feature map to be processed. After superimposing the output signals of the eighth convolution module and the ninth convolution module, the output feature map results are sequentially passed through the eighth activation module, the tenth convolution module and the ninth activation module, and the corresponding position elements of the output feature map are multiplied with the original feature map to be processed to obtain the final processed feature map.

[0085] The dimensions and parameters of each module are shown in Tables 1 and 2 below.

[0086] Table 1 U-Net model based on attention mechanism

[0087]

[0088]

[0089] Table 2 Attention gate module:

[0090]

[0091] The pre-training process of the U-Net network based on the attention mechanism in step S3 or step S4 includes:

[0092] 1) Obtain training set, validation set data and corresponding labels;

[0093] 2) The training set data is input into the network as a channel, and the hyperparameter values ​​are set at the same time. The hyperparameter values ​​include the batch size, the maximum training cycle, the initial learning rate, the learning rate optimizer, and the loss function. In this embodiment, the settings are as follows:

[0094] Batch size: 8; Maximum training epochs: 100; Initial learning rate: 0.008; Learning rate optimizer: amw optimizer;

[0095] Loss function: Tversky loss function with Sigmoid, where alpha = 0.7, beta = 0.3.

[0096] For step S3, the size of I3 and A1 is 128×128×128.

[0097] For step S4, the sizes of I4 and A2 are 128×128×512.

[0098] 3) Calculate the network output results of the training set data and the loss function value of the corresponding label;

[0099] 4) Adjust the parameter values ​​in the network model according to the loss function results;

[0100] 5) After the training set data has been input for a certain period of time, the validation set data and labels are input;

[0101] 6) Calculate the network output results of the validation set data and the loss function value of the corresponding label;

[0102] 7) Adjust the hyperparameter values ​​of the network model according to the loss function results;

[0103] 8) Repeat steps 2)-7) until the loss function values ​​in 3) and 6) converge and are less than the set value, and then end the training.

[0104] In step S2, the data preprocessing process P1 specifically includes: medical images of different pixel pitches and image sizes are subjected to resampling and center cropping (or peripheral padding) steps to generate medical image data of a unified standard. This unified standard can be: In the resampling step, the interpolation mode is nearest neighbor interpolation. In the center cropping step, the Z-axis direction is not processed. In the peripheral padding step, the Z-axis direction is not processed, and the padding mode is constant value padding, where the constant value is the minimum value of the input data. In this embodiment, H = 512, D = 512.

[0105] In step S3, the data preprocessing process P2 is as follows: the input data is downsampled to a size of H1×W1×Z1 and normalized. In the downsampling step, the interpolation mode is nearest neighbor interpolation. In the normalization step, for CTA data, voxel points with intensity values ​​in the interval [a, b] are linearly mapped to the interval [0, 1]; voxel points with intensity values ​​less than a are set to 0; and voxel points with intensity values ​​greater than b are set to 0. In Example 1: a = -300, b = 1200; the obtained data size H1×W1×Z1 is 128×128×128.

[0106] In steps S3 and S4, the network structure is as follows Figure 2 、 Figure 3 , as shown in Table 1 and Table 2.

[0107] In steps 4 and 6, the hyperparameters in this embodiment are set as follows:

[0108] Step S5, the data post-processing process P3:

[0109] Step S51, extracting the image region R1 corresponding to the coordinate A1 in I3;

[0110] Step S52: resample R1 to obtain coordinates A'1 in the same space as I2;

[0111] Step S53: extract the image area corresponding to the coordinate A'1 in I2, which is I4. In the resampling step, the interpolation mode is the nearest neighbor interpolation.

[0112] In step S7, the data post-processing process P4:

[0113] Step S71, restore A2 to the coordinate A'1 area in I2;

[0114] Step S72: A2 is subjected to resampling and center cropping (or surrounding filling) steps to generate a blood vessel segmentation result A'2 in the same space as I1. A'2 is combined with I1 to form a neck medical image and a carotid artery segmentation combined image I5 for visualization.

[0115] Next, a system embodiment of the present invention is provided, which is a deep learning-based medical image carotid artery segmentation system. The system adopts the method described above and includes:

[0116] A data acquisition module is used to obtain original neck medical imaging data I1;

[0117] A data preprocessing module is configured to obtain standardized second cervical medical image data I2 after subjecting the original cervical medical image data I1 to a first preprocessing process P1, and obtain normalized third cervical medical image data I3 after subjecting the second cervical medical image data I2 to a second preprocessing process P2;

[0118] Blood vessel positioning module: input the third neck medical image data I3 into the pre-trained first attention-based U-Net network to perform blood vessel positioning and obtain the blood vessel positioning coordinates A1;

[0119] a blood vessel segmentation module, configured to use a pre-trained second attention-based U-Net network to perform blood vessel segmentation on fourth neck medical image data I4 obtained by performing the first data post-processing process P3 on the blood vessel positioning coordinates A1, the second neck medical image data I2, and the third neck medical image data I3, to obtain a blood vessel segmentation result A2;

[0120] The visualization module is used to visualize the neck medical image obtained from the blood vessel segmentation result A2, the original neck medical image data I1 and the second neck medical image data I2, and the carotid artery blood vessel segmentation combined image I5.

[0121] Figures 5 to 7 Schematic diagram of the neck medical image data segmentation process and results in the embodiment. Figure 5 I4 (right) and I4 (left) in the figure represent the right carotid artery and the left carotid artery, respectively. Figure 6 The middle part shows the segmentation results of the right carotid artery and the left carotid artery.

[0122] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A method for segmenting carotid arteries in medical images based on deep learning, characterized in that: The method comprises the following steps: Step S1: Data acquisition: obtaining original cervical medical imaging data ; Step S2: Data preprocessing: original cervical medical imaging data After the first pre-processing After that, the second cervical medical imaging data with unified standards is obtained. ; The second cervical medical imaging data After the second pre-processing , get the normalized third cervical medical imaging data ; Step S3, blood vessel positioning: the third neck medical imaging data Input the pre-trained first U-Net network based on the attention mechanism to perform vascular positioning and obtain the vascular positioning coordinates ; Step S4, blood vessel segmentation: locate the blood vessel coordinates , Second cervical medical imaging data and third cervical medical imaging data , after the first data post-processing process After that, the fourth cervical medical imaging data is obtained ; The fourth cervical medical imaging data Input the pre-trained second U-Net network based on the attention mechanism for blood vessel segmentation to obtain the blood vessel segmentation result ; Step S5: Visualization processing: combining blood vessel segmentation results , original neck medical imaging data and second cervical medical imaging data , after the second data post-processing process , get the neck medical image and carotid artery segmentation combined image , used for visual display; The first data post-processing process in step S4 , specifically: Step S401: In the third cervical medical image data Extracting blood vessel coordinates Corresponding image area ; Step S402: image area Resample and obtain the second cervical medical imaging data Coordinates in the same space ; Wherein, the resampling interpolation mode is nearest neighbor interpolation; Step S403: In the second cervical medical image data Extract coordinates from The corresponding image area is the fourth cervical medical imaging data .

2. The method for segmenting carotid arteries in medical images based on deep learning according to claim 1, characterized in that: The original cervical medical image data in step S1 For CTA data.

3. The method for segmenting carotid arteries in medical images based on deep learning according to claim 1, characterized in that: The first pre-processing process in step S2 , specifically: Step S201: original neck medical images with different pixel spacing and image sizes Resampling is performed; wherein the interpolation mode of the resampling is nearest neighbor interpolation; Step S202: perform center cropping or peripheral padding on the resampled neck medical image to obtain a second neck medical image data of a unified standard. ; wherein, the center clipping The axis direction is not processed, and the surrounding filling is filled with a constant value, and the constant value is the minimum value of the input data; The second pretreatment process , specifically: Step S211: The second cervical medical image data Downsample to The size of the downsampling is , wherein the interpolation mode in the downsampling is nearest neighbor interpolation; Step S212: Normalize the downsampled image data. Specifically, linearly map the voxel points with intensity values ​​in the interval [a, b] to the interval [0, 1]. Set the voxel points with intensity values ​​less than a to 0. Set the voxel points with intensity values ​​greater than b to 0.

4. The method for segmenting carotid arteries in medical images based on deep learning according to claim 1, characterized in that: The U-Net network based on the attention mechanism in step S3 or step S4 includes an attention gate module, a fusion module, and a first convolution module, a first activation module, a first down-sampling module, a second convolution module, a second activation module, a second down-sampling module, a third convolution module, a third activation module, a third down-sampling module, a fourth convolution module and a fourth activation module, a first up-sampling module, a fifth convolution module, a fifth activation module, a second up-sampling module, a sixth convolution module, a sixth activation module, a third up-sampling module, a seventh convolution module and a seventh activation module connected in sequence; The output of the third activation module and the output of the fourth activation module are respectively input into the attention gate module, and the output of the attention gate module and the output of the first upsampling module are fused through the fusion module as the input of the fifth convolution module; the output of the second activation module and the output of the fifth activation module are respectively input into the attention gate module, and the output of the attention gate module and the output of the second upsampling module are fused through the fusion module as the input of the sixth convolution module; the output of the first activation module and the output of the sixth activation module are respectively input into the attention gate module, and the output of the attention gate module and the output of the third upsampling module are fused through the fusion module as the input of the fifth convolution module.

5. The method for segmenting carotid arteries in medical images based on deep learning according to claim 4, characterized in that: The attention gate module includes an eighth convolution module, a ninth convolution module, an eighth activation module, a tenth convolution module and a ninth activation module; the input of the eighth convolution module is a gate signal, and the input of the ninth convolution module is a feature map to be processed. After superimposing the output signals of the eighth convolution module and the ninth convolution module, the output feature map result is sequentially passed through the eighth activation module, the tenth convolution module and the ninth activation module, and the corresponding position elements of the output feature map result are multiplied with the original feature map to be processed to obtain the final processed feature map.

6. The method for segmenting carotid arteries in medical images based on deep learning according to claim 1, characterized in that: The pre-training process of the U-Net network based on the attention mechanism in step S3 or step S4 includes: 1) Obtain training set, validation set data and corresponding labels; 2) Input the training set data into the network as a channel and set the hyperparameter values; the hyperparameter values ​​include the batch size, maximum training cycle, initial learning rate, learning rate optimizer, and loss function; 3) Calculate the network output of the training set data and the loss function value of the corresponding label; 4) Adjust the parameter values ​​in the network model according to the loss function results; 5) After the training set data has been input for a certain period of time, the validation set data and labels are input; 6) Calculate the network output results of the validation set data and the loss function value of the corresponding label; 7) Adjust the hyperparameter values ​​of the network model based on the loss function results; 8) Repeat steps 2)-7) until the loss function values ​​in 3) and 6) converge and are less than the set value, and then end the training.

7. The method for segmenting carotid arteries in medical images based on deep learning according to claim 6, characterized in that: The loss function is a Tversky loss function with Sigmoid.

8. The method for segmenting carotid arteries in medical images based on deep learning according to claim 1, characterized in that: The step S5 specifically includes the following sub-steps: Step S51: segment the blood vessels Restore to the second cervical medical imaging data Coordinates in within the area; Step S52: segment the blood vessels After resampling and center cropping or surrounding padding, the original neck medical image data is generated Blood vessel segmentation results in the same space ; Step S53: segment the blood vessels Original neck medical imaging data Combined to form a neck medical image and a carotid artery segmentation combined image , used for visual display.

9. A medical image carotid artery segmentation system based on deep learning, characterized by: The method according to any one of claims 1 to 8, wherein the system comprises: Data acquisition module, used to obtain original neck medical imaging data ; Data preprocessing module, used to process original neck medical imaging data After the first pre-processing After that, the second cervical medical imaging data with unified standards is obtained. , for the second cervical medical imaging data After the second pre-processing , get the normalized third cervical medical imaging data ; Vascular positioning module: The third neck medical imaging data Input the pre-trained first U-Net network based on the attention mechanism to perform vascular positioning and obtain the vascular positioning coordinates ; The blood vessel segmentation module is used to locate the coordinates of blood vessels using a pre-trained second attention-based U-Net network , Second cervical medical imaging data and third cervical medical imaging data After the first data post-processing process Obtained fourth cervical medical imaging data Perform blood vessel segmentation and obtain blood vessel segmentation results ; Visualization module is used to visualize the blood vessel segmentation results , original neck medical imaging data and second cervical medical imaging data The obtained neck medical image is combined with the carotid artery segmentation image Provide visual display.

Images