A method of identifying aortic dissection and related apparatus

Abstract

The application discloses an aortic dissection identification method and related device, which can be applied to the technical field of deep learning, extracts local features of a to-be-identified heart ultrasound image, and the local features are aortic region features. The local features of the to-be-identified heart ultrasound image are input into a trained dissection classification model to obtain a classification result output by the dissection classification model, and the classification result is used to indicate whether an aortic dissection exists in the to-be-identified heart ultrasound image. A support set at least includes local features of a first type of heart ultrasound image, the first type of heart ultrasound image is a heart ultrasound image with an aortic dissection, and a classification label is a classification result artificially annotated to indicate whether an aortic dissection exists. The classification accuracy of the dissection classification model trained based on the feature similarity between the local features of the to-be-identified heart ultrasound image and the local features of the heart ultrasound image with the aortic dissection is high, and thus, the classification result accuracy is improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and apparatus for identifying aortic dissection. Background Technology

[0002] Aortic dissection, also known as aortic aneurysm, is a serious cardiovascular disease, and early diagnosis and treatment are essential. The ultrasound appearance of aortic dissection typically shows widening of the aorta with a torn intima, appearing as a thin, membranous echo. The aorta usually divides into true and false lumens. Thrombosis may occasionally be seen within the false lumen. When the dissection involves the aortic root, color Doppler flow imaging can usually detect aortic valve regurgitation, and may also show discontinuity in the torn intima.

[0003] Therefore, aortic dissection is usually diagnosed by using deep learning models to identify the ultrasound features of aortic dissection on echocardiography. However, the accuracy of deep learning models in identifying aortic dissection is currently low.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a method and related apparatus for identifying aortic dissection, aiming to improve the accuracy of aortic dissection identification results, as follows:

[0006] A method for identifying aortic dissection includes:

[0007] Local features are extracted from the echocardiogram image to be identified, wherein the local features are aortic region features;

[0008] The local features of the echocardiogram image to be identified are input into the trained aortic dissection classification model to obtain the classification result output by the aortic dissection classification model. The classification result is used to indicate whether aortic dissection exists in the echocardiogram image to be identified.

[0009] The aortic dissection classification model is trained based on the classification labels of local features of sample echocardiogram images and the feature similarity with the support set. The support set includes at least local features of a first type of echocardiogram images, which are echocardiogram images with aortic dissection. The classification labels are manually labeled classification results used to indicate whether aortic dissection exists.

[0010] Optionally, local features of the echocardiogram image to be identified are extracted, including:

[0011] The echocardiogram image to be identified is input into a pre-trained local feature recognition model to obtain the local features output by the local feature recognition model;

[0012] The local feature recognition model is obtained by jointly training the target region detection model and the local feature extraction module.

[0013] The local feature recognition model is used to extract global features from the echocardiogram image to be recognized, and the aortic region of the echocardiogram image to be recognized is obtained based on the global feature recognition.

[0014] The local feature extraction module extracts features from the global features of the echocardiogram image to be identified based on the aortic region, and outputs the local features of the echocardiogram image to be identified.

[0015] Alternatively, the method for identifying aortic dissection also includes:

[0016] Obtain a sample image dataset, which includes multiple echocardiogram images with classification annotations;

[0017] Each echocardiogram image in the sample image dataset is input into the local feature recognition model to obtain the local features of each echocardiogram image output by the local feature recognition model;

[0018] Based on classification and labeling, the echocardiogram images in the sample image dataset are divided into the sample echocardiogram images, the first type of echocardiogram images, and the second type of echocardiogram images. The second type of echocardiogram images are echocardiogram images without aortic dissection.

[0019] Based on the segmentation results and the local features of each echocardiogram image, a query set and a support set are generated. The query set includes local features of multiple sample echocardiogram images with classification labels. The support set also includes local features of the second type of echocardiogram images.

[0020] The query set and the support set are input into the mezzanine classification model. The mezzanine classification model is trained based on the classification labels of the local features of the sample echocardiogram images and the feature similarity with the support set until the preset training completion condition is met.

[0021] Optionally, training the mezzanine classification model based on the classification labels of local features of the sample echocardiogram images and the feature similarity with the support set includes:

[0022] Calculate the feature similarity between the sample echocardiogram image and the support set;

[0023] Based on the classification labels of the sample echocardiogram images and the feature similarity, the mezzanine classification model is trained using the binary cross-entropy loss function.

[0024] Optionally, calculating the feature similarity between the sample echocardiogram image and the support set includes:

[0025] Calculate the distribution distance between the sample echocardiogram image and the support set;

[0026] Calculate the class distance between the sample echocardiogram image and the support set;

[0027] The feature similarity is obtained by fusing the distribution distance and the category distance.

[0028] A device for identifying aortic dissection, comprising:

[0029] A local feature extraction unit is used to extract local features from the echocardiogram image to be identified, wherein the local features are aortic region features;

[0030] The aortic dissection classification unit is used to input the local features of the echocardiogram image to be identified into the trained aortic dissection classification model to obtain the classification result output by the aortic dissection classification model. The classification result is used to indicate whether aortic dissection exists in the echocardiogram image to be identified.

[0031] The aortic dissection classification model is trained based on the classification labels of local features of sample echocardiogram images and the feature similarity with the support set. The support set includes at least local features of a first type of echocardiogram images, which are echocardiogram images with aortic dissection. The classification labels are manually labeled classification results used to indicate whether aortic dissection exists.

[0032] Optionally, when the local feature extraction unit is used to extract local features from the echocardiogram image to be identified, it is specifically used for:

[0033] The echocardiogram image to be identified is input into a pre-trained local feature recognition model to obtain the local features output by the local feature recognition model;

[0034] The local feature recognition model is obtained by jointly training the target region detection model and the local feature extraction module.

[0035] The local feature recognition model is used to extract global features from the echocardiogram image to be recognized, and the aortic region of the echocardiogram image to be recognized is obtained based on the global feature recognition.

[0036] The local feature extraction module extracts features from the global features of the echocardiogram image to be identified based on the aortic region, and outputs the local features of the echocardiogram image to be identified.

[0037] A computer program product includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to perform various steps of a method for identifying aortic dissection.

[0038] A device for identifying aortic dissection includes: a memory and a processor;

[0039] The memory is used to store programs;

[0040] The processor is used to execute the program to implement the various steps of the aortic dissection identification method.

[0041] A readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the various steps of a method for identifying aortic dissection.

[0042] As can be seen from the above technical solutions, the aortic dissection identification method and related apparatus provided in this application disclose an aortic dissection identification method and related apparatus that can be applied to the field of deep learning technology. The method extracts local features from the echocardiogram image to be identified, where the local features are aortic region features. The local features of the echocardiogram image to be identified are input into a trained dissection classification model to obtain the classification result output by the model. The classification result is used to indicate whether aortic dissection exists in the echocardiogram image to be identified. The dissection classification model is trained based on the classification labels of the local features of the sample echocardiogram image and the feature similarity with the support set. The support set includes at least the local features of a first type of echocardiogram image, which is an echocardiogram image containing aortic dissection. The classification labels are manually labeled classification results used to indicate whether aortic dissection exists. The dissection classification model trained based on the feature similarity between the local features of the echocardiogram image to be identified and the local features of the echocardiogram image containing aortic dissection has high classification accuracy, thereby improving the accuracy of the classification result. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 A flowchart illustrating a method for identifying aortic dissection provided in an embodiment of this application;

[0045] Figure 2 A schematic diagram of the structure of an aortic dissection identification model provided in this application embodiment;

[0046] Figure 3 A flowchart illustrating a training method for a mezzanine classification model provided in an embodiment of this application;

[0047] Figure 4A schematic diagram illustrating a local feature extraction process provided in an embodiment of this application;

[0048] Figure 5 A schematic diagram illustrating the training process of a mezzanine classification model provided in an embodiment of this application;

[0049] Figure 6 A schematic diagram of the structure of an aortic dissection identification device provided in an embodiment of this application;

[0050] Figure 7 This is a schematic diagram of the structure of an aortic dissection identification device provided in an embodiment of this application. Detailed Implementation

[0051] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0052] In recent years, with the development of deep learning, especially convolutional neural networks, the use of deep learning models for downstream computer vision tasks to recognize and diagnose echocardiograms has been widely applied in the field of medical image recognition. Compared with traditional methods, deep learning-based recognition or diagnostic methods have significant advantages in terms of accuracy and robustness.

[0053] However, current deep learning models are inaccurate in recognizing the similarity between local representations in echocardiography and aortic dissection representations, leading to inaccurate classification results based on similarity. Furthermore, sufficient sample data is required to train deep learning models during model construction. Due to the scarcity of aortic dissection cases, the amount of sample data used to train deep learning models is insufficient, resulting in inaccurate recognition of local representations in echocardiography and further contributing to the low accuracy of current deep learning models.

[0054] Based on this, the embodiments of this application provide a method for identifying aortic dissection, which aims to identify whether aortic dissection exists in echocardiogram images based on a pre-trained dissection classification model, thereby improving the accuracy of aortic dissection identification.

[0055] Figure 1 A flowchart illustrating a method for identifying aortic dissection provided in this application is shown below. Figure 1 As shown, this method specifically includes:

[0056] S101. Extract local features from the echocardiogram image to be identified.

[0057] In this embodiment, the local features are the features of the aortic region.

[0058] Specifically, regional features of the aortic region are extracted from the echocardiogram image to be identified as local features of the echocardiogram image.

[0059] It should be noted that there are various methods for extracting local features. Optionally, a pre-trained local feature recognition model can be used to extract local features from the echocardiogram image to be identified. Specifically, the echocardiogram image to be identified is input into the pre-trained local feature recognition model to obtain the local features output by the model. The local feature recognition model is obtained by jointly training a target region detection model and a local feature extraction module. The feature data training set used for joint training includes multiple sample echocardiogram images and region annotation data. The region annotation data of the sample echocardiogram images includes the aortic region of the sample echocardiogram images.

[0060] In this embodiment, the local feature recognition model is used to extract the global features of the echocardiogram image to be recognized, and the aortic region of the echocardiogram image to be recognized is obtained based on the global feature recognition; the local feature extraction module extracts features from the global features of the echocardiogram image to be recognized based on the aortic region of the echocardiogram image to be recognized, and outputs the local features of the echocardiogram image to be recognized.

[0061] S102. Input the local features of the echocardiogram image to be identified into the trained mezzanine classification model to obtain the classification result output by the mezzanine classification model.

[0062] In this embodiment, the classification result is used to indicate whether aortic dissection exists in the echocardiogram image to be identified.

[0063] In this embodiment, the aortic dissection classification model is trained based on the classification labels of local features of sample echocardiogram images and the feature similarity with the support set. The support set includes at least the local features of a first type of echocardiogram images, which are echocardiogram images with aortic dissection. The classification labels are manually labeled classification results used to indicate whether aortic dissection exists.

[0064] In this embodiment, there are several methods for calculating the feature similarity between the sample echocardiogram image and the support set. For example, the distribution distance between the sample echocardiogram image and the support set can be calculated as the feature similarity. Alternatively, the class distance between the sample echocardiogram image and the support set can be calculated as the feature similarity. Or, the calculated distribution distance and class distance can be fused together and the fusion result can be used as the feature similarity.

[0065] As can be seen from the above technical solutions, the aortic dissection identification method provided in this application outputs the extracted local features of the echocardiogram image to be identified to the dissection classification model. Since the dissection classification model is trained based on the classification labels of the local features of the sample echocardiogram image and the feature similarity with the support set, wherein the support set includes at least the echocardiogram images that are determined to have aortic dissection, the classification accuracy of the dissection classification model trained based on the feature similarity between the local features of the echocardiogram image to be identified and the local features of the echocardiogram image with aortic dissection is high. Thus, the accuracy of the classification result in indicating whether the echocardiogram image to be identified has aortic dissection is improved.

[0066] Furthermore, embodiments of this application provide an aortic dissection identification model based on deep learning. Figure 2 A schematic diagram illustrating the structure of an aortic dissection identification model provided in an embodiment of this application is shown below. Figure 2 As shown, the aortic dissection identification model includes a local feature recognition model and a dissection classification model. The local feature recognition model is used to generate local features of the aortic region in the input echocardiogram. The dissection classification model is used to obtain the similarity between the local features of the aortic region and the local representation of the aortic dissection based on the local features of the aortic region, and output the classification result based on the similarity. The classification result includes one of the following: the presence of dissection and the absence of dissection, which is used to characterize whether aortic dissection exists in the input echocardiogram.

[0067] Specifically, the local feature recognition model includes a target region detection model and a local feature extraction module. The target region detection model includes a global feature extractor (also called a feature extractor) and a target detection module (also called an aorta detection module). For example... Figure 3 As shown, the aortic detection module is used to detect the aortic region in the input echocardiogram. Specifically, the feature extractor is used to extract the global feature representation of the input echocardiogram and input it into the aortic detection module. The aortic detection module is used to detect the aortic region in the input echocardiogram based on the global features of the echocardiogram, so as to realize aortic region detection. The local feature extraction module is used to crop the global features based on the aortic region recognition result (the aortic region in the echocardiogram) to obtain the local feature representation of the aortic region.

[0068] In this embodiment, the mezzanine classification model includes a first distance calculation module, a second distance calculation module, and a mezzanine classification module. The first distance calculation module calculates the distribution distance, the second distance calculation module calculates the classification distance, and the mezzanine classification module obtains similarity based on the distribution distance and the classification distance, and outputs the classification result based on the similarity.

[0069] Figure 3This application provides a method for constructing an aortic dissection identification model, such as... Figure 3 As shown, this method includes:

[0070] S301. Obtain the feature data training set.

[0071] In this embodiment, the feature data training set includes multiple sample echocardiogram images and region annotation data, wherein the region annotation data of the sample echocardiogram images includes the aortic region of the sample echocardiogram images.

[0072] S302. Based on the feature data training set, the target region detection model and the local feature extraction module that constitute the local feature recognition model are jointly trained until the preset training completion conditions are met, and a trained local feature recognition model is obtained.

[0073] In this embodiment, the local feature recognition model is obtained by jointly training the target region detection model and the local feature extraction module. The feature data training set used for joint training includes: multiple sample echocardiogram images and region annotation data, wherein the region annotation data of the sample echocardiogram images includes the aortic region of the sample echocardiogram images.

[0074] In this embodiment, the local feature recognition model is used to extract the global features of the echocardiogram image to be recognized, and the aortic region of the echocardiogram image to be recognized is obtained based on the global feature recognition; the local feature extraction module extracts features from the global features of the echocardiogram image to be recognized based on the aortic region of the echocardiogram image to be recognized, and outputs the local features of the echocardiogram image to be recognized.

[0075] It should be noted that the local feature recognition model trained based on sample echocardiogram images and region-labeled data has the ability to extract regional features of the aortic region in the image, and the feature extraction accuracy is high.

[0076] S303. Obtain the sample image dataset.

[0077] In this embodiment, the sample image dataset includes multiple echocardiogram images with classification labels. The classification labels are manually annotated to indicate the presence of aortic dissection.

[0078] S304. Input each echocardiogram image in the sample image dataset into the local feature recognition model to obtain the local features of each echocardiogram image output by the local feature recognition model.

[0079] In this embodiment, the target region detection model in the local feature recognition model is YOLOX (You Only LookOnce-X), and the representation of local and global features is a feature map. Figure 3 The example illustrates the specific implementation flow of a local feature extraction method, as shown in the diagram. Figure 3 As shown, the method includes:

[0080] A1. Input the echocardiogram image into the global feature extractor in the target region detection model to obtain the global feature map of the echocardiogram image output by the global feature extractor.

[0081] Optionally, the global feature extractor maps the echocardiogram image to a feature map, resulting in a global feature map. It should be noted that the image size of the echocardiogram image and the size of the resulting global feature map are not necessarily the same.

[0082] For example, the image size of the input echocardiogram is... The feature map size of the global feature map of the echocardiogram image is Where C is the number of channels, H is the height, and W is the width.

[0083] A2. Input the global feature map of the echocardiogram image into the target detection module, which then identifies the aortic region of the echocardiogram image based on the global feature map.

[0084] A3. Input the aortic region of the echocardiogram image and the corresponding global feature map of the echocardiogram image into the local feature extraction module to obtain the local features of the aortic region obtained by cropping the global feature map based on the aortic region of the echocardiogram image.

[0085] In this embodiment, the local feature extraction module maps the aortic region at the image size to the feature map size, and performs cropping and interpolation on the aortic region in the global feature map to obtain local feature representation.

[0086] Optionally, for echocardiogram images Local features of the aortic region in echocardiogram images The local tensor description, including C dimensions, is as follows:

[0087]

[0088] in, It is the i-th local description of the local features, where n is the product of the length and width of the local feature map. .

[0089] S305. Based on classification labeling, the echocardiogram images in the sample image dataset are divided into sample echocardiogram images, first-class echocardiogram images, and second-class echocardiogram images.

[0090] In this embodiment, the sample echocardiogram images include echocardiogram images classified as having aortic dissection and echocardiogram images not having aortic dissection. The first type of echocardiogram images includes multiple echocardiogram images with aortic dissection, and the second type of echocardiogram images includes multiple echocardiogram images without aortic dissection.

[0091] S306. Based on the segmentation results and the local features of each echocardiogram image, generate a query set and a support set.

[0092] In this embodiment, the query set includes local features of multiple sample echocardiogram images with classification labels. The support set includes local features of first-class and second-class echocardiogram images. The echocardiogram images in the support set and the query set do not overlap.

[0093] It should be noted that the local features of the echocardiogram images are all represented by local feature maps. The query set includes the set of local features of the sample echocardiogram images, and the support set includes the set of local features of the first type of echocardiogram images and the set of local features of the second type of echocardiogram images.

[0094] It can be understood that the set of local features of the first type of echocardiogram image represents the local features of the echocardiogram image with aortic dissection, while the set of local features of the second type of echocardiogram image represents the local features of the echocardiogram image without aortic dissection.

[0095] In this embodiment, the local feature set of the support set It can be described as ,in, , , represent the vector mean and covariance matrix of a multivariate Gaussian distribution, respectively. Similarly, the local features of the query set... It can be described as In this context, the subscript S refers to the support set, and the subscript Q refers to the query set.

[0096] S307. Input the query set and support set into the mezzanine classification model.

[0097] In this embodiment, the local feature set of the support set and the local feature set of the query set are respectively input to the first distance calculation module and the second distance calculation module.

[0098] S308. Obtain the distribution distance between the sample hyperocular image and the support set calculated by the first distance calculation module.

[0099] In this embodiment, the method for calculating the distribution distance between the sample echocardiogram and the support set is as follows:

[0100]

[0101] Here, trace represents the trace of the matrix, and det represents the determinant of the matrix.

[0102] It should be noted that the distribution distance , is an asymmetric distance metric that describes the distributional distance between sample echocardiograms in the query set and the support set.

[0103] S309. Obtain the second distance calculation module to calculate the class distance between the sample echocardiogram and the support set.

[0104] In this embodiment, arrive Category distance This describes the category similarity between sample echocardiogram images in the query set and either Category 1 or Category 2 echocardiogram images in the support set that exhibit aortic dissection. Optionally, the category distance can be calculated using the following formula:

[0105]

[0106] Where topK represents selecting the K largest elements in each row of the correlation matrix; optionally, K=1.

[0107] S310. Input the class distance and distribution distance into the mezzanine classification module. Train the mezzanine classification module based on the feature similarity obtained by fusing the distribution distance and class distance and the classification label of the sample hyperocular image until the preset second training completion condition is met.

[0108] In this embodiment, the mezzanine classification module is built based on a neural network, and uses the binary cross-entropy loss function as the loss calculation for the mezzanine classification module. The binary cross-entropy loss function is as follows:

[0109]

[0110] in, Represents cross-entropy loss, This represents the probability value predicted by the model as positive. Represents the actual label.

[0111] like Figure 4As shown, the local feature extraction model is used to extract the local features Q of the sample echocardiogram images in the query set, as well as the local features S1 and S2 of the first type of echocardiogram images in the support set. Based on the first distance calculation module and the second distance calculation module, the distribution distance and category distance of the local features of the sample echocardiogram images and the local features of the first type of echocardiogram images (and / or the second type of echocardiogram images) in the support set are calculated respectively to obtain the distribution distance and category distance. The feature similarity obtained after the distribution distance and category distance are fused by concat operation is input into the mezzanine classification module so that the mezzanine classification model is trained based on the classification label and the feature similarity to obtain the mezzanine classification model used to output the classification result.

[0112] S311. In response to the achievement of the preset second training completion condition, the mezzanine classification model is obtained.

[0113] S312. Construct an aortic dissection identification model based on a local feature recognition model and a dissection classification model.

[0114] As can be seen from the above technical solution, in the construction method of the aortic dissection recognition model provided in this application embodiment, a feature extractor is first trained by using a target detection dataset with a large amount of data to extract general features in the echocardiogram image. The detected aortic region is mapped onto the feature map to obtain local feature representation. Then, through the local feature representation, the distribution distance and category distance between the image and the category in the dataset of dissection with a small amount of data are calculated. The two are then fused to finally obtain the category similarity distance, and the classification and recognition of the aortic dissection in the echocardiogram image are performed.

[0115] Therefore, this approach overcomes the problem of insufficient sample data for training deep learning models due to the scarcity of aortic dissection cases, which leads to inaccurate recognition of local representations in echocardiograms by deep learning models, and improves the training effect of deep learning models.

[0116] The aortic dissection identification model provided in this application obtains a local feature map of the aortic region by detecting the aortic region where the dissection occurs and mapping it to a global feature map for cropping. This method utilizes a feature extractor trained on a sufficiently large target detection dataset to extract general features of echocardiogram images, while avoiding the impact of irrelevant redundant features on model performance.

[0117] The distribution distance metric used in this application can better describe the distance between the feature distribution of the query set image and the feature distribution of the support set image than other distance metrics commonly used in metric learning. Therefore, it is more advantageous for identifying and classifying sandwich images in the case of small samples.

[0118] This application proposes a method that integrates category distance and distribution distance and calculates the final similarity distance using a model, which better describes the similarity distance between categories indicating the presence or absence of aortic dissection in the local feature representation of the aortic region.

[0119] Figure 6 A schematic diagram of the structure of an aortic dissection identification device provided in an embodiment of this application is shown, as follows: Figure 6 As shown, the device may include:

[0120] The local feature extraction unit 601 is used to extract local features of the echocardiogram image to be identified, wherein the local features are aortic region features;

[0121] The aortic dissection classification unit 602 is used to input the local features of the echocardiogram image to be identified into the trained aortic dissection classification model to obtain the classification result output by the aortic dissection classification model. The classification result is used to indicate whether aortic dissection exists in the echocardiogram image to be identified.

[0122] The aortic dissection classification model is trained based on the classification labels of local features of sample echocardiogram images and the feature similarity with the support set. The support set includes at least local features of a first type of echocardiogram images, which are echocardiogram images with aortic dissection. The classification labels are manually labeled classification results used to indicate whether aortic dissection exists.

[0123] Optionally, when the local feature extraction unit is used to extract local features from the echocardiogram image to be identified, it is specifically used for:

[0124] The echocardiogram image to be identified is input into a pre-trained local feature recognition model to obtain the local features output by the local feature recognition model;

[0125] The local feature recognition model is obtained by jointly training the target region detection model and the local feature extraction module.

[0126] The local feature recognition model is used to extract global features from the echocardiogram image to be recognized, and the aortic region of the echocardiogram image to be recognized is obtained based on the global feature recognition.

[0127] The local feature extraction module extracts features from the global features of the echocardiogram image to be identified based on the aortic region, and outputs the local features of the echocardiogram image to be identified.

[0128] Optionally, the aortic dissection identification device further includes a model training unit for:

[0129] Obtain a sample image dataset, which includes multiple echocardiogram images with classification annotations;

[0130] Each echocardiogram image in the sample image dataset is input into the local feature recognition model to obtain the local features of each echocardiogram image output by the local feature recognition model;

[0131] Based on classification and labeling, the echocardiogram images in the sample image dataset are divided into the sample echocardiogram images, the first type of echocardiogram images, and the second type of echocardiogram images. The second type of echocardiogram images are echocardiogram images without aortic dissection.

[0132] Based on the segmentation results and the local features of each echocardiogram image, a query set and a support set are generated. The query set includes local features of multiple sample echocardiogram images with classification labels. The support set also includes local features of the second type of echocardiogram images.

[0133] The query set and the support set are input into the mezzanine classification model. The mezzanine classification model is trained based on the classification labels of the local features of the sample echocardiogram images and the feature similarity with the support set until the preset training completion condition is met.

[0134] Optionally, when the model training unit trains the mezzanine classification model based on the classification labels of local features of the sample echocardiogram images and the feature similarity with the support set, it is specifically used for:

[0135] Calculate the feature similarity between the sample echocardiogram image and the support set;

[0136] Based on the classification labels of the sample echocardiogram images and the feature similarity, the mezzanine classification model is trained using the binary cross-entropy loss function.

[0137] Optionally, when the model training unit is used to calculate the feature similarity between the sample echocardiogram image and the support set, it is specifically used for:

[0138] Calculate the distribution distance between the sample echocardiogram image and the support set;

[0139] Calculate the class distance between the sample echocardiogram image and the support set;

[0140] The feature similarity is obtained by fusing the distribution distance and the category distance.

[0141] This application also provides a computer program product, including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to perform various steps of a method for identifying aortic dissection, as follows:

[0142] A method for identifying aortic dissection includes:

[0143] Local features are extracted from the echocardiogram image to be identified, wherein the local features are aortic region features;

[0144] The local features of the echocardiogram image to be identified are input into the trained aortic dissection classification model to obtain the classification result output by the aortic dissection classification model. The classification result is used to indicate whether aortic dissection exists in the echocardiogram image to be identified.

[0145] The aortic dissection classification model is trained based on the classification labels of local features of sample echocardiogram images and the feature similarity with the support set. The support set includes at least local features of a first type of echocardiogram images, which are echocardiogram images with aortic dissection. The classification labels are manually labeled classification results used to indicate whether aortic dissection exists.

[0146] It should be noted that the specific implementation methods of each step of the aortic dissection identification method can be found in the above embodiments, and will not be repeated here.

[0147] Figure 7 A schematic diagram of the structure of the aortic dissection identification device is shown. The device may include: at least one processor 701, at least one communication interface 702, at least one memory 703, and at least one communication bus 704.

[0148] In this embodiment of the application, the number of processor 701, communication interface 702, memory 703 and communication bus 704 is at least one, and processor 701, communication interface 702 and memory 703 communicate with each other through communication bus 704.

[0149] The processor 701 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

[0150] The memory 703 may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device;

[0151] The memory stores a program, and the processor can execute the program stored in the memory to implement the various steps of the aortic dissection identification method provided in this application embodiment, as follows:

[0152] A method for identifying aortic dissection includes:

[0153] Local features are extracted from the echocardiogram image to be identified, wherein the local features are aortic region features;

[0154] The local features of the echocardiogram image to be identified are input into the trained aortic dissection classification model to obtain the classification result output by the aortic dissection classification model. The classification result is used to indicate whether aortic dissection exists in the echocardiogram image to be identified.

[0155] The aortic dissection classification model is trained based on the classification labels of local features of sample echocardiogram images and the feature similarity with the support set. The support set includes at least local features of a first type of echocardiogram images, which are echocardiogram images with aortic dissection. The classification labels are manually labeled classification results used to indicate whether aortic dissection exists.

[0156] It should be noted that the specific implementation methods of each step of the aortic dissection identification method can be found in the above embodiments, and will not be repeated here.

[0157] This application embodiment also provides a readable storage medium that can store a computer program suitable for processor execution. When the computer program is executed by the processor, it implements the various steps of the aortic dissection identification method provided in this application embodiment, as follows:

[0158] A method for identifying aortic dissection includes:

[0159] Local features are extracted from the echocardiogram image to be identified, wherein the local features are aortic region features;

[0160] The local features of the echocardiogram image to be identified are input into the trained aortic dissection classification model to obtain the classification result output by the aortic dissection classification model. The classification result is used to indicate whether aortic dissection exists in the echocardiogram image to be identified.

[0161] The aortic dissection classification model is trained based on the classification labels of local features of sample echocardiogram images and the feature similarity with the support set. The support set includes at least local features of a first type of echocardiogram images, which are echocardiogram images with aortic dissection. The classification labels are manually labeled classification results used to indicate whether aortic dissection exists.

[0162] It should be noted that the specific implementation methods of each step of the aortic dissection identification method can be found in the above embodiments, and will not be repeated here.

[0163] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0164] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0165] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A device for identifying aortic dissection, characterized by, The method comprises the following steps: a local feature extraction unit is configured to extract local features of a to-be-identified cardiac ultrasound image, wherein the local features are aortic region features; a dissection classification unit is configured to input the local features of the to-be-identified cardiac ultrasound image into a trained dissection classification model to obtain a classification result output by the dissection classification model, wherein the classification result is used to indicate whether there is an aortic dissection in the to-be-identified cardiac ultrasound image; wherein the dissection classification model is trained based on classification labels of local features of sample cardiac ultrasound images and feature similarities with a support set, wherein the support set at least includes local features of first type cardiac ultrasound images, the first type cardiac ultrasound images are cardiac ultrasound images with aortic dissection, and the classification labels are classification results artificially labeled to indicate whether there is an aortic dissection; when the local feature extraction unit is used to extract the local features of the to-be-identified cardiac ultrasound image, the local feature extraction unit is specifically configured to input the to-be-identified cardiac ultrasound image into a pre-trained local feature recognition model to obtain local features output by the local feature recognition model; wherein the local feature recognition model is obtained by jointly training a target region detection model and a local feature extraction module; the local feature recognition model is used to extract global features of the to-be-identified cardiac ultrasound image, identify the aortic region of the to-be-identified cardiac ultrasound image based on the global features, and output the local features of the to-be-identified cardiac ultrasound image; a model training unit is configured to obtain a sample image data set, wherein the sample image data set includes a plurality of cardiac ultrasound images with band classification labels; input each cardiac ultrasound image in the sample image data set into the local feature recognition model to obtain local features of each cardiac ultrasound image output by the local feature recognition model; divide the cardiac ultrasound images in the sample image data set into the sample cardiac ultrasound images, the first type cardiac ultrasound images, and second type cardiac ultrasound images based on the classification labels, wherein the second type cardiac ultrasound images are cardiac ultrasound images without aortic dissection; generate a query set and the support set according to the division result and the local features of each cardiac ultrasound image, wherein the query set includes a plurality of local features of the sample cardiac ultrasound images with classification labels; the support set further includes local features of the second type cardiac ultrasound images; input the query set and the support set into the dissection classification model, and train the dissection classification model based on the classification labels of the local features of the sample cardiac ultrasound images and the feature similarities with the support set until a preset training completion condition is reached; when the model training unit is used to train the dissection classification model based on the classification labels of the local features of the sample cardiac ultrasound images and the feature similarities with the support set, the model training unit is specifically configured to calculate the feature similarities between the sample cardiac ultrasound images and the support set; and train the dissection classification model based on a binary cross-entropy loss function according to the classification labels of the sample cardiac ultrasound images and the feature similarities.

2. The device for identifying aortic dissection according to claim 1, characterized by, The model training unit is configured to calculate the feature similarity between the sample echocardiogram image and the support set, and specifically configured to: calculate the distribution distance between the sample echocardiogram image and the support set; calculate the category distance between the sample echocardiogram image and the support set; fuse the distribution distance and the category distance to obtain the feature similarity.

Images