Segmentation model training method and coronary plaque segmentation method and device

Abstract

The invention provides a segmentation model training method and a coronary artery plaque segmentation method and device, and the method comprises the steps: obtaining a training sample image which is a coronary artery image containing coronary artery plaques; the training sample image is marked, a marked image is obtained, and the marked image is an image obtained by marking a coronary artery wall and a plaque area in the coronary artery wall on the training sample image; and constructing an initial segmentation model, and training the initial segmentation model based on the annotated image by taking the training sample image as an input image to obtain a target segmentation model. According to the method, the defects of original convolution feature extraction are overcome, the problem that plaque segmentation based on gray values or gradients is low in accuracy is solved, automatic plaque segmentation of coronary artery CTA is achieved, the burden of doctors can be relieved, and diagnosis and treatment efficiency can be improved.

Description

Technical Field

[0001] The present application relates to the field of medical image segmentation, and in particular, to a training method for a segmentation model, a segmentation method for coronary artery plaques, a device, an electronic device, and a storage medium. Background Art

[0002] Coronary heart disease is an important disease threatening the public health in China. In recent years, the incidence and mortality of coronary heart disease in China have been increasing year by year, and it shows a trend of getting younger. Coronary artery plaques are the main pathological manifestation of coronary heart disease. Coronary artery plaques include calcified plaques and soft plaques, which will directly lead to stenosis or even occlusion of the blood vessel lumen of the coronary artery.

[0003] CT angiography (CTA) is a non-invasive angiography synthesized by computer three-dimensional reconstruction method and has been widely used in coronary artery detection. Although using images with improved spatial quality can improve the doctor's film reading efficiency, due to the large number of CTA scan images, manual film reading will consume a lot of time and energy, and doctors are also easily affected by subjective judgment during film reading, resulting in inaccurate results. Compared with manually delineating plaques, automatically segmenting plaques by computer is faster, and can ensure the consistency and repeatability of the segmentation results while ensuring the accuracy. Traditional coronary artery plaque segmentation algorithms, such as adaptive threshold algorithms, feature-based detection methods, etc., mainly segment plaques according to the gray value and gradient change of the image, and have limited segmentation accuracy for different plaques. Therefore, accurately segmenting coronary artery plaques from medical images has important clinical significance for diagnosing heart diseases such as coronary heart disease, and there is an urgent need for an efficient and accurate automatic segmentation method for coronary artery plaque segmentation. Summary of the Invention

[0004] In view of the above problems, the present application provides a training method for a segmentation model, a segmentation method for coronary artery plaques, and a device by means of deep learning, which can automatically and accurately identify and segment coronary artery plaques, clearly show the morphology, size, distribution and other characteristics of the plaques, and assist the film reading doctor to achieve more accurate diagnosis.

[0005] The specific solutions are as follows:

[0006] The first aspect of the present application provides a training method for a segmentation model, including:

[0007] Obtain a training sample image, where the training sample image is a coronary artery image containing coronary artery plaques;

[0008] Label the training sample image to obtain a labeled image, where the labeled image is an image obtained by labeling the coronary artery wall and the plaque area therein in the training sample image;

[0009] Construct an initial segmentation model, use the training sample image as the input image, and train the initial segmentation model based on the labeled image to obtain a target segmentation model.

[0010] In a possible implementation, Dice loss is used as the loss function in the training process.

[0011] In a possible implementation, the segmentation model includes two Unet networks. Among them, Resnet is used as the backbone network of the initial feature extractor, forming a structure of "Resnet - decoder - encoder - decoder".

[0012] In a possible implementation, the encoder consists of convolutional layers, normalization layers, Relu activation functions, attention layers, and pooling layers to form residual blocks, and then the deep feature extraction of the input image is realized through the stacking of several residual blocks; the decoder is composed of the stacking of several decoding blocks, and the decoding block is composed of an upsampling layer, convolutional layers, normalization layers, attention layers, and Relu activation functions.

[0013] In a possible implementation, the attention layer includes a dual attention module embedded in the encoder and decoder structures to capture the association of feature map channels and spatial association.

[0014] In a possible implementation, it also includes using a test set to test the segmentation model to evaluate the model performance. Among them, pixel accuracy (PA), intersection over union (IOU), and Dice coefficient are used to evaluate the model performance.

[0015] The second aspect of this application provides a method for segmenting coronary artery plaques, including:

[0016] Obtain a coronary artery image to be segmented;

[0017] Input the coronary artery image to be segmented into the target segmentation model, and output the segmentation image of the coronary artery image to be segmented, where the target segmentation model is a model trained in advance according to any method provided in the first aspect above.

[0018] The third aspect of this application provides a training device for a segmentation model, which is characterized by including:

[0019] An image acquisition module, configured to acquire a training sample image and label the training sample image to obtain a labeled image;

[0020] A model training module, configured to construct an initial segmentation model, use the training sample image as the input image, and train the initial segmentation model based on the labeled image to obtain a target segmentation model.

[0021] A fourth aspect of the present application provides an electronic device, including a processor and a memory. The memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the steps of the method provided in any one of the above first aspects or to implement the steps of the method provided in the above second aspect.

[0022] A fifth aspect of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is run by a processor, it executes the steps of the method provided in any one of the above first aspects or implements the steps of the method provided in the above second aspect.

[0023] The present application has the following beneficial effects:

[0024] The training method of the segmentation model, the segmentation method and device of coronary artery plaques provided by the present application, compared with the traditional segmentation method, the present application uses a deep residual neural network to capture complex edge and texture information through multi-layer non-linear feature learning, and has stronger segmentation ability for plaques with low contrast, small size and various shapes. At the same time, it can effectively alleviate the problem of gradient disappearance, and improve the robustness and accuracy of segmentation; by constructing a two-layer Unet network to deeply extract image features, and improve the model's attention to local texture information in the image. Among them, the local texture information is closely related to the shape complexity, boundary ambiguity of the plaque and the contrast change with the surrounding tissues. By capturing these local texture features, the model can more accurately distinguish the plaque from the adjacent normal tissues, especially in areas with low contrast and variable morphology, so as to improve the accuracy and robustness of segmentation; the Resnet module is introduced as an initial feature extractor, which well solves the problems of gradient disappearance and gradient explosion during training; in order to capture the correlation between the channels of the feature map and the spatial correlation, a dual attention module is proposed to be embedded in the encoder and decoder structures to improve the quality of feature extraction. Based on the above, the solution of the present invention has the characteristics of fast detection speed and excellent detection results, thereby improving the accuracy of plaque segmentation in coronary CTA images and bringing convenience to medical detection. Description of the Drawings

[0025] Figure 1 It is a schematic flow chart of a training method of a segmentation model provided by an embodiment of the present application;

[0026] Figure 2 It is a schematic structural diagram of a segmentation model provided by an embodiment of the present application;

[0027] Figure 3 It is a schematic flow chart of a segmentation method of coronary artery plaques provided by an embodiment of the present application.

[0028] Figure 4Schematic diagram of the structure of a training device for a segmentation model provided by an embodiment of the present application.

[0029] Figure 5 Schematic diagram of the structure of a segmentation device for coronary artery plaques provided by an embodiment of the present application.

[0030] Figure 6 Schematic diagram of the structure of an electronic device provided by an embodiment of the present application. Detailed implementation manners

[0031] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the present invention, rather than limiting the present invention. In addition, it should be noted that, for the sake of convenience of description, only parts related to the present invention are shown in the drawings, rather than all the structures.

[0032] Embodiment 1

[0033] Refer to Figure 1 Fig. shows a flowchart of a method for training a segmentation model provided by Embodiment 1 of the present application. The method may include Operation S101 to Operation S103.

[0034] It should be understood that although Figure 1 the steps in the flowchart are shown in the direction of the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise clearly stated in this article, the execution of these steps has no strict order limitation, and these steps can be executed in other orders. Moreover, Figure 1 at least a part of the steps in

[0035] may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least a part of other steps or sub-steps or stages of other steps.

[0036] In the embodiment of the present disclosure, a CTA image containing coronary artery plaques can be collected by CT scanning as a training sample image.

[0037] In Operation S102, the training sample image is labeled to obtain a labeled image, where the labeled image is an image obtained by labeling the coronary artery wall and the plaque area therein in the training sample image.

[0038] In the embodiments of the present disclosure, when annotating the training sample images, the images outlined and verified by radiologists for CTA images can be used as the reference images for plaque segmentation.

[0039] Optionally, when annotating, the coronary artery structure can be segmented from the CTA image first to clarify the region of interest, reduce the interference of irrelevant information, and improve the annotation efficiency and accuracy.

[0040] Specifically, outline the coronary artery wall and the plaque region within it in the image, and focus on annotating the boundary, morphological features, and density distribution of the plaque. At the same time, ensure that the boundary between the plaque and the surrounding normal tissues is clearly visible to provide accurate reference data for the subsequent segmentation model.

[0041] Optionally, when annotating, the coordinates of the original CTA image can be converted into Cartesian coordinates. The image generated by Cartesian coordinates is more in line with the human vascular structure, can simplify the geometric transformation process of the image, and is also more convenient for annotating plaques.

[0042] Optionally, data augmentation is performed on the annotated images. The data augmentation includes at least one of the following: distortion, horizontal flipping, vertical flipping, rotation at any angle, scaling, and resampling. By performing augmentation operations on the limited dataset, more effective sample images are effectively created on the basis of the original samples, and the operation is simple and does not require additional costs.

[0043] Optionally, preprocessing can be performed on the annotated image set. The preprocessing methods include: adjusting CTA images of different sizes to the same size; and / or normalizing the data values of CTA images.

[0044] For example, when adjusting CTA images of different sizes to the same size, a standard size can be selected. If the image is larger than the standard size, an edge cropping strategy is used for processing. If the image is smaller than the standard size, an edge zero-padding strategy is used for processing.

[0045] For example, the maximum-minimum normalization method can be selected to normalize the data values of CTA images. The formula for the normalization method is:

[0046]

[0047] where X is the original image, X max is the maximum value of the pixel values in X, X min is the minimum value of the pixel values in X, X ′ is the normalized image.

[0048] Optionally, the preprocessed image set can be divided into a training set, a validation set, and a test set.

[0049] Specifically, the training set is used to train the model, the validation set is used to debug and optimize the hyperparameters in the model, and the test set is used to evaluate the model performance.

[0050] For example, the preprocessed image set is divided into a training set, a validation set, and a test set in a ratio of 60%, 20%, and 20%.

[0051] In operation S103, an initial segmentation model is constructed, with the training sample images as the input images, and the initial segmentation model is trained based on the annotated images to obtain the target segmentation model.

[0052] See Figure 2 , Figure 2 which is a schematic structural diagram of the segmentation model provided by the embodiment of the present application. The segmentation model includes two Unet networks. Among them, Resnet is used as the backbone network of the initial feature extractor, forming a structure of "Resnet - decoder - encoder - decoder".

[0053] Optionally, the segmentation model can be composed of two encoding - decoding structures to improve the model's attention to local texture information in the image. Among them, the encoder is responsible for feature extraction to learn distinguishable patch information, and the decoder is responsible for upsampling and reconstructing the feature map.

[0054] Optionally, Resnet can be used as the backbone network of the initial feature extractor, and feature transfer is performed with the corresponding decoder through skip connections.

[0055] It can be understood that during the training of deep neural network models, during the backpropagation process, the gradients decrease layer by layer, and finally the gradients of the first few layers are almost zero, making it impossible for the neural network to effectively update the parameters. Excessive network depth will also lead to the problem of gradient explosion, which is a phenomenon where the gradients are amplified layer by layer, ultimately resulting in an extremely large update amplitude of the model's weights and the training process being unable to converge. This gradient problem is mainly caused by the characteristics of the neural network itself. Especially when deep neural networks are stacked in multiple layers, gradient disappearance and gradient explosion are likely to occur, which is closely related to the network structure design and the choice of activation functions. Although the characteristics of CTA images, such as noise, low - contrast regions, or complex structures, may increase the difficulty of training, the root cause of the gradient problem lies in the expansion or attenuation of gradients during the backpropagation process of deep neural networks, especially when no appropriate initialization method or optimization strategy is adopted. Therefore, in the embodiment of the present application, Resnet is used as the backbone network of the initial feature extractor, and feature transfer is performed with the corresponding decoder through skip connections, which can effectively solve the problems of gradient disappearance and gradient explosion during training, making the training more efficient and robust.

[0056] For example, the encoder can consist of a convolutional layer, a normalization layer, a Relu activation function, an attention layer, and a pooling layer to form a residual block, and then stack several residual blocks to achieve deep feature extraction of the input image.

[0057] For example, first initialize the convolutional kernel group, set the bias parameter b to 0, and the input image X enters the convolutional layer for convolution operation. If the dimension of the input image is n×n×n c , and the dimension of the convolutional kernel is (n - f + 1)×(n - f + 1)×n c ′, where n c ′ is the number of convolutional kernels. Then add the convolution operation result to the bias parameter b and use the activation function to perform a non-linear operation on this result to obtain the final result of this layer. The activation function used in the present invention is the Relu activation function, and the formula can be expressed as:

[0058] Relu(X) = Max(0, X)

[0059] The calculation process in the convolutional layer can be expressed by the formula:

[0060] Z [l] = W [l] A [l-1] + b [l}

[0061] A [l] = Relu(Z [l] )

[0062] For example, since the original image obtains several training parameters through the operation of the convolutional layer, in order to improve the training speed of the model and make the training process more stable, and effectively avoid the problem of covariate shift in the model, it is necessary to perform standardization processing on the above weight parameters W [l] and b [l} . That is, a normalization layer is added after the convolutional layer, and its formula can be expressed as:

[0063]

[0064] Among them, m is the number of training samples, the superscript i represents the i-th sample, and ε is a constant to prevent the denominator from being 0. Considering that the input should have differences and diversities, usually Z [l-1](i) needs to be further processed, and its formula can be expressed as:

[0065]

[0066] Among them, γ and β are adjustable parameters, and in the follow-up, the values of these hyperparameters can be adjusted to make the model reach the best performance. Here, the role of γ and β is to make have both the mean and variance as any reasonable values to ensure its diversity.

[0067] For example, an attention layer can be added after the convolutional layer and the normalization layer, and a dual attention module can be introduced and embedded into the encoder and decoder structures, which can be specifically expressed as a formula:

[0068] Y ′ = X + X * σ(cha atten(X) + spa atten(X) )

[0069] Among them, σ(·) represents the ReLu activation function, cha_atten(·) represents the channel attention calculation, and spa_atten(·) represents the spatial attention calculation.

[0070] It can be understood that in a neural network, the correlation of feature map channels and spatial correlation are two different features, which respectively reflect different levels of image data. Among them, the correlation of feature map channels is commonly implemented by the channel attention mechanism, which assigns different weights to different channels to highlight important feature channels and suppress irrelevant channels; spatial correlation refers to the relationship between pixel positions in an image or the similarity of local regions, which is commonly implemented by the spatial attention mechanism. By focusing on feature maps at different positions in the image, the representation of regions with important information in the image is enhanced. The single attention mechanism mainly focuses on the attention in one dimension, that is, channel attention or spatial attention, while the dual attention mechanism combines the attention in both channel and spatial dimensions to more accurately capture the correlation of feature map channels and spatial correlation and improve the quality of feature extraction. Specifically in this solution, the channel correlation corresponds to the expression ability of different gray levels or texture patterns (such as the density difference between plaques and normal tissues) in coronary CTA images; the spatial correlation corresponds to the spatial distribution of plaques in the image, boundary features and their relationship with surrounding tissues. Through the dual attention mechanism, it is possible to simultaneously focus on the global context (spatial correlation) of a specific region and the importance of different feature channels (channel correlation) when extracting features, so as to better reflect the complex structural characteristics of the image. Compared with the single attention mechanism, the dual attention mechanism can optimize both the channel and spatial dimensions simultaneously, enhancing the comprehensive understanding of global and local information. Therefore, the dual attention mechanism can more effectively distinguish plaques from the background or other tissues, thereby improving the accuracy and robustness of feature segmentation. In this solution, improving the quality of feature extraction can enhance the ability to distinguish the plaque area from the background, especially in the case of low contrast regions and blurred boundaries; reduce the attention to irrelevant information and improve the model's ability to capture key regions (such as plaque boundaries and morphological features); improve the accuracy and consistency of the final segmentation results, providing higher quality data support for downstream plaque analysis and diagnosis.

[0071] For example, since the pooling layer is used to reduce the image size, improve the operation speed, and also reduce the influence of noise, making each feature robust, the maximum pooling method can be adopted for downsampling to reduce the influence of noise and improve the robustness of the model.

[0072] For example, the decoder can be composed of several decoding blocks stacked together, and the decoding block can be composed of an upsampling layer, a convolutional layer, a normalization layer, an attention layer, and a Relu activation function.

[0073] Specifically, the upsampling layer is located before the convolutional layer for upsampling operations. In the embodiments of the present application, transposed convolution is used as the upsampling method to more clearly magnify the feature image.

[0074] In the embodiments of the present application, through a structure of Resnet - decoder - encoder - decoder, and introducing an attention mechanism, the model can accurately extract plaque image features, and Resnet can effectively solve the problems of gradient descent and gradient disappearance caused by too many neural network layers, making the model robust.

[0075] In the embodiments of the present disclosure, the training set is input into the initial segmentation model for iterative training to obtain weight parameters, and the hyperparameters of the model are adjusted through the validation set to obtain the optimal performance; the test set is used to test the segmentation model, evaluate the model performance, and load the trained weight parameters into the segmentation model to obtain the trained model, that is, the target segmentation model.

[0076] Optionally, Diceloss is used as the loss function during the training process.

[0077] Specifically, Diceloss can be expressed by the formula:

[0078]

[0079] Where X represents the pixel labels of the true segmentation image, and Y represents the pixel categories of the predicted segmentation image.

[0080] It can be understood that in the image segmentation task, the class imbalance problem and the large difference in the number of positive and negative samples are common challenges. Specifically, in image data, there are often some classes (such as lesion areas or target objects) that are relatively scarce, while other classes (such as the background) occupy most of the pixels; this data imbalance will cause the model to be more likely to focus on the classes with a larger number during training and ignore the classes with a smaller number. Therefore, in the embodiments of the present application, Diceloss is used as the loss function to handle the class imbalance problem in the image segmentation task, making the model have a certain robustness when performing some class - imbalanced operations, being able to effectively cope with the large difference in the number of positive and negative samples in the segmentation task, and making the model pay more attention to the segmentation effect of small classes.

[0081] Optionally, the segmentation model is tested using a test set to evaluate the model performance. Among them, pixel accuracy (PA), intersection over union (IOU), and Dice coefficient can be used to evaluate the model performance. Using these three coefficients has the following advantages: Pixel accuracy (PA) can intuitively reflect the correct rate of the segmentation model in overall pixel classification and is suitable for measuring the overall performance of the model; Intersection over union (IOU) focuses on the proportion of the overlapping area between the prediction result and the ground truth label and can effectively evaluate the localization accuracy of the model for the target area; The Dice coefficient emphasizes the consistency between the prediction result and the ground truth label and is particularly suitable for the segmentation evaluation of imbalanced data. These evaluation metrics comprehensively consider the accuracy, precision, and consistency of the segmentation model, can comprehensively reflect the actual performance of the model, and provide a reliable basis for optimization and improvement.

[0082] Specifically, pixel accuracy (PA) is the proportion of correctly classified pixels in the image to the total number of pixels, and its calculation formula can be:

[0083]

[0084] Specifically, intersection over union (IOU) is the overlapping area between the predicted segmentation and the ground truth segmentation divided by their union area. IOU is an important indicator for measuring segmentation accuracy, and the higher its value, the better the segmentation effect. Its calculation formula can be:

[0085]

[0086] Specifically, the Dice coefficient is an indicator for evaluating the similarity or overlapping degree of two samples and is used to evaluate the similarity of segmentation. Its calculation formula can be:

[0087]

[0088] Among them, TP, TN, FP, and FN represent true positive, true negative, false positive, and false negative respectively.

[0089] Example 2

[0090] The second embodiment of the present application provides a method for segmenting coronary artery plaques. The target segmentation model used in this segmentation method is trained by the training method provided in the foregoing Example 1.

[0091] Please refer to Figure 3 , which is a flowchart of a method for segmenting coronary artery plaques provided by the second embodiment of the present application. The method includes:

[0092] Operation S301: Obtain a coronary artery image to be segmented.

[0093] Operation S302: Input the coronary artery image to be segmented into the target segmentation model, and output the segmented image of the coronary artery image to be segmented.

[0094] In Embodiment 2 of the present application, the target segmentation model is pre-trained according to the segmentation model training method provided in Embodiment 1.

[0095] In some embodiments of the present application, S109 includes: obtaining a coronary artery CTA image to be segmented; extracting the features of the input image and transmitting the features through skip connections to the corresponding decoder; multiplying the predicted feature map by the original input and inputting it into the second Unet network; adding the output of the first Unet and the output of the second Unet and performing upsampling to obtain the final segmentation result.

[0096] First, obtain the target coronary artery CTA image to be segmented.

[0097] Next, use the Resnet network as the initial feature extractor to extract the features of the input image and transmit the features through skip connections to the corresponding decoder. The calculation formula can be expressed as:

[0098] Y0 = Decoder(Resnet(X))

[0099] where X is the input image.

[0100] Then, multiply the predicted feature map by the original input and input it into the second encoder-decoder network, and its formula can be expressed as:

[0101] Y1 = Encoder(Y0·X)

[0102] Y2 = Decoder(Y1 + Resnet(X))

[0103] Finally, add the output of the first Unet and the output of the second Unet and perform upsampling to obtain the final segmentation result. The calculation formula can be expressed as:

[0104] Y final = Upsample(Y0 + Y2)

[0105] In summary, based on deep image processing technology and CTA, the present application constructs a deep residual neural network to realize automatic segmentation of coronary artery plaques, reduce the burden on doctors, and improve the diagnosis and treatment efficiency.

[0106] Embodiment 3

[0107] The embodiment of the present application provides a training device for a segmentation model. See Figure 4Schematic structural diagram of a training device for a segmentation model, which may include the following parts:

[0108] An image acquisition module 401, configured to acquire a training sample image and label the training sample image to obtain a labeled image;

[0109] A model training module 402, configured to construct an initial segmentation model, and use the training sample image as an input image to train the initial segmentation model based on the labeled image to obtain a target segmentation model.

[0110] Example 4

[0111] An embodiment of the present application provides a segmentation device for coronary artery plaques. Refer to Figure 5 Schematic structural diagram of a segmentation device for coronary artery plaques, which may include the following parts:

[0112] An acquisition module 501, configured to acquire a coronary artery image to be segmented.

[0113] A segmentation module 502, configured to input the coronary artery image to be segmented into the target segmentation model and output a segmentation image of the coronary artery image to be segmented; wherein, the target segmentation model is pre-trained according to the method described in Example 1.

[0114] In an implementation manner, the above device may further include a result presentation module, configured to output the segmentation result of the target segmentation model; the result presentation module may simultaneously display the original image to be segmented and the segmented image. If the image is segmented into multiple regions, the user is allowed to view a specific segmentation layer for detailed analysis.

[0115] Example 5

[0116] An embodiment of the present application provides an electronic device. Specifically, the electronic device includes a processor and a storage device; a computer program is stored on the storage device, and the computer program, when run by the processor, executes the method described in any of the above embodiments.

[0117] Figure 6 Schematic structural diagram of an electronic device provided by an embodiment of the present application. The electronic device 600 includes: a processor 60, a memory 61, a bus 62, and a communication interface 63. The processor 60, the communication interface 63, and the memory 61 are connected through the bus 62; the processor 60 is configured to execute an executable module stored in the memory 61, such as a computer program.

[0118] Among them, the memory 61 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory. The communication connection between the system network element and at least one other network element is implemented through at least one communication interface 63 (which can be wired or wireless), and the Internet, wide area network, local area network, metropolitan area network, etc. can be used.

[0119] The bus 62 can be an ISA bus, a PCI bus, an EISA bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of easy representation, Figure 6 only a bidirectional arrow is used in the figure, but it does not mean that there is only one bus or one type of bus.

[0120] Among them, the memory 61 is used to store a program. After receiving an execution instruction, the processor 60 executes the program. The method executed by the device defined by the flow process disclosed in any embodiment of the foregoing embodiments of the present invention can be applied to the processor 60 or implemented by the processor 60.

[0121] The processor 60 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit in the hardware of the processor 60 or by an instruction in software form. The above-mentioned processor 60 can be a general-purpose processor, including a central processing unit (CPU for short), a network processor (NP for short), etc.; it can also be a digital signal processor (DSP for short), an application specific integrated circuit (ASIC for short), a field-programmable gate array (FPGA for short), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. It can implement or execute the various methods, steps and logic block diagrams disclosed in the embodiments of the present invention. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as being executed and completed by a hardware decoding processor, or by a combination of hardware and software modules in the decoding processor. The software module can be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. This storage medium is located in the memory 61, and the processor 60 reads the information in the memory 61 and combines its hardware to complete the steps of the above method.

[0122] Example 6

[0123] This embodiment provides a computer-readable storage medium, on which a computer program is stored, and the computer program is executed by a processor to perform the method described in any of the foregoing embodiments.

[0124] The computer program product of the readable storage medium provided by the embodiments of the present application includes a computer-readable storage medium storing program codes, and the instructions included in the program codes can be used to execute the method described in the foregoing method embodiments. For specific implementation, reference can be made to the foregoing method embodiments, which will not be elaborated herein.

[0125] If the above functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0126] Those skilled in the art should understand that the embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects.

[0127] Those skilled in the art can easily understand that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A training method for a segmentation model, characterized in that Including: Obtain training sample images, where the training sample images are coronary artery images containing coronary plaques; Annotate the training sample images to obtain annotated images, where the annotated images are images obtained by annotating the coronary artery wall and the plaque regions therein in the training sample images; Construct an initial segmentation model, use the training sample images as input images, and train the initial segmentation model based on the annotated images to obtain a target segmentation model.

2. The training method according to claim 1, wherein Use Dice loss as the loss function in the training process.

3. The training method according to claim 1, wherein The segmentation model includes two Unet networks, where Resnet is used as the backbone network of the initial feature extractor, forming a structure of "Resnet - decoder - encoder - decoder".

4. The training method according to claim 3, wherein The encoder consists of convolutional layers, normalization layers, Relu activation functions, attention layers, and pooling layers to form residual blocks, and then realizes the extraction of deep features of the input images through the stacking of several residual blocks; The decoder consists of the stacking of several decoding blocks, and the decoding blocks are composed of upsampling layers, convolutional layers, normalization layers, attention layers, and Relu activation functions.

5. The training method according to claim 4, characterized in that The attention layer includes dual attention modules embedded in the encoder and decoder structures to capture the association of feature map channels and spatial association.

6. The training method according to claim 1, characterized in that It also includes using a test set to test the segmentation model to evaluate the model performance, where pixel accuracy (PA), intersection over union (IOU), and Dice coefficient are used to evaluate the model performance.

7. A method for segmenting coronary artery plaques, characterized in that, Including: Obtain the coronary artery image to be segmented; Input the coronary artery image to be segmented into the target segmentation model, and output the segmentation image of the coronary artery image to be segmented, where the target segmentation model is a model pre-trained according to the method described in any one of claims 1 to 6.

8. A training device for a segmentation model, characterized in that, Including: An image acquisition module, configured to obtain training sample images and annotate the training sample images to obtain annotated images; A model training module, configured to construct an initial segmentation model, use the training sample images as input images, and train the initial segmentation model based on the annotated images to obtain a target segmentation model.

9. An electronic device, characterized in that, Including a processor and a memory, where the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement the steps of the method described in any one of claims 1 to 6 or to implement the steps of the method described in claim 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it executes the steps of the method described in any one of claims 1 to 6 or executes the steps of the method described in claim 7.

Images