Aortic dissection identification model construction method, identification method and related device

Abstract

The application discloses an aortic dissection identification model construction method and identification method and related devices, and relates to the technical field of artificial intelligence, and comprises the following steps: training a teacher network model by using an echocardiogram data set of non-aortic dissection containing a large number of data samples, so that the trained teacher network model has good generalization ability. The data samples in the echocardiogram data set of the aorta are respectively input into the trained teacher network model and a student network model, the model parameters of the student network model are adjusted according to the difference between the trained teacher network model and the student network model in a preset feature recognition result and the difference between the recognition result of the student network model and a label true value, the student network model can learn the generalization ability of the teacher network model, the student network model can be trained based on a small amount of aortic dissection data samples, and the accuracy of the student network model in identifying the aortic dissection can also be ensured.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to a method for constructing an aortic dissection identification model, an identification method, and related devices. Background Technology

[0002] Ultrasound scanning equipment is widely used in the medical field to perform ultrasound scanning imaging of human tissues. Echocardiography is an ultrasound scan of the heart, using standard two-dimensional, three-dimensional, or Doppler ultrasound to form an image of the heart.

[0003] Currently, after obtaining an echocardiogram, it can be input into a trained machine learning model. The machine learning model identifies various features of the heart and cardiovascular system to obtain recognition results for specific application scenarios, which can be used by doctors for reference. However, in some application scenarios, the recognition results of the machine learning model are less accurate and cannot meet the application requirements. Summary of the Invention

[0004] In view of this, this application provides a method for constructing an aortic dissection identification model, an identification method, and related devices to improve the robustness of the aortic dissection identification model, thereby improving the efficiency of aortic dissection identification. The specific solution is as follows:

[0005] The first aspect of this application provides a method for constructing an aortic dissection identification model, including:

[0006] Obtain echocardiographic datasets of non-aortic dissection and aortic dissection. Each data sample in the non-aortic dissection echocardiographic dataset is labeled with a preset feature label, and each data sample in the aortic dissection echocardiographic dataset is labeled with a preset feature label and an aortic dissection label.

[0007] The task is to identify the preset features of each data sample in the non-aortic dissection echocardiogram dataset. The teacher network model is trained using the non-aortic dissection echocardiogram dataset until the teacher network model converges.

[0008] The data samples from the echocardiogram dataset of the aorta are respectively input into the trained teacher network model and student network model to obtain the preset feature recognition results output by the trained teacher network model and the preset feature recognition results and aortic dissection recognition results output by the student network model.

[0009] Based on the differences between the preset feature recognition results output by the trained teacher network model and the preset feature recognition results output by the student network model, the differences between the preset feature recognition results output by the student network model and the preset feature label ground truth, and the differences between the aortic dissection recognition results output by the student network model and the aortic dissection label ground truth, the model parameters of the student network model are adjusted until the student network model converges, and the trained student network model is determined as the aortic dissection recognition model; wherein, the initial model parameters of the teacher network model and the student network model are the same.

[0010] In one possible implementation, adjusting the model parameters of the student network model based on the differences between the preset feature recognition results output by the trained teacher network model and the preset feature recognition results output by the student network model, the differences between the preset feature recognition results output by the student network model and the preset feature label ground truth, and the differences between the aortic dissection recognition results output by the student network model and the aortic dissection label ground truth, until the student network model converges, includes:

[0011] A first loss function is constructed based on the difference between the preset feature recognition results output by the trained teacher network model and the preset feature recognition results output by the student network model.

[0012] A second loss function is constructed based on the difference between the preset feature recognition results output by the student network model and the preset feature label ground truth values;

[0013] Based on the difference between the aortic dissection identification results output by the student network model and the true values ​​of the aortic dissection labels, an aortic dissection probability prediction loss function is constructed.

[0014] Based on the first loss function, the second loss function, and the aortic dissection probability prediction loss function, the loss function of the student network model is constructed.

[0015] The loss function of the student network model is optimized using gradient descent until the loss function of the student network model converges, at which point the training of the student network model ends.

[0016] In one possible implementation, the preset features include spatial and temporal features. A first loss function is constructed based on the difference between the preset feature recognition results output by the trained teacher network model and the preset feature recognition results output by the student network model, including:

[0017] Based on the first spatial feature recognition result output by the trained teacher network model, the second spatial feature recognition result output by the student network model, and the preset distillation temperature, a first spatial consistency loss function is constructed.

[0018] Based on the first time feature recognition result output by the trained teacher network model, the second time feature recognition result output by the student network model, and the preset distillation temperature, a first time consistency loss function is constructed.

[0019] The first loss function is constructed based on the first spatial consistency loss function and the first temporal consistency loss function.

[0020] In one possible implementation, the method for obtaining the first spatial feature recognition result and the second spatial feature recognition result includes:

[0021] For each data sample in the echocardiogram dataset of the aortic dissection, the forward-order image sequence in the data sample is reversed in the time dimension to obtain the reverse-order image sequence.

[0022] The reverse-order image sequence is input into the trained teacher network model to obtain the first teacher feature map, and the forward-order image sequence is input into the student network model to obtain the student feature map;

[0023] The first teacher feature map is input into the spatial feature recognition branch of the trained teacher network model to obtain the first spatial feature recognition result, and the student feature map is input into the spatial feature recognition branch of the student network model to obtain the second spatial feature recognition result.

[0024] In one possible implementation, the method for obtaining the first time feature recognition result and the second time feature recognition result includes:

[0025] For each data sample in the echocardiogram dataset of the aortic dissection, the region of interest in the aortic region is cropped from the forward sequence of images in the data sample, and the region of interest is interpolated to the input size of the trained teacher network model to obtain the aortic region image sequence.

[0026] The aortic region image sequence is input into the trained teacher network model to obtain the second teacher feature map, and the forward sequence image sequence is input into the student network model to obtain the student feature map;

[0027] The second teacher feature map is input into the time feature recognition branch of the trained teacher network model to obtain the first time feature recognition result, and the student feature map is input into the time feature recognition branch of the student network model to obtain the second time feature recognition result.

[0028] In one possible implementation, the preset features include spatial and temporal features. A second loss function is constructed based on the difference between the preset feature recognition results output by the student network model and the ground truth values ​​of the preset feature labels, including:

[0029] Based on the difference between the spatial feature recognition results output by the student network model and the true values ​​of the spatial feature labels, a second spatial consistency loss function is constructed.

[0030] Based on the difference between the time feature recognition results output by the student network model and the true values ​​of the time feature labels, a second time consistency loss function is constructed.

[0031] The second loss function is constructed based on the second spatial consistency loss function and the second temporal consistency loss function.

[0032] In one possible implementation, the preset features include spatial and temporal features. The task of identifying the preset features of each data sample in the non-aortic dissection echocardiogram dataset, and training the teacher network model using the non-aortic dissection echocardiogram dataset until the teacher network model converges, includes:

[0033] The loss function of the teacher network model is constructed based on the spatial consistency loss function and the temporal consistency loss function. The spatial consistency loss function represents the difference between the recognition result of the spatial feature by the teacher network model and the true value of the spatial feature label. The temporal consistency loss function represents the difference between the recognition result of the temporal feature by the teacher network model and the true value of the temporal feature label.

[0034] During the training of the teacher network model using the non-aortic dissection echocardiogram dataset, the gradient descent method is used to optimize the loss function of the teacher network model until the loss function of the teacher network model converges, at which point the training of the teacher network model ends.

[0035] In one possible implementation, the spatial feature is a slice category, and the temporal feature includes end-contraction frames and end-diastolic frames.

[0036] A second aspect of this application provides a method for identifying aortic dissection, comprising:

[0037] Acquire the image sequence of the echocardiogram to be processed;

[0038] The image sequence of the echocardiogram to be processed is input into the aortic dissection recognition model constructed by the aortic dissection recognition model construction method described in the first aspect or any implementation of the first aspect, and the aortic dissection recognition result output by the aortic dissection recognition model is obtained.

[0039] A third aspect of this application provides an apparatus for constructing an aortic dissection identification model, comprising:

[0040] The dataset acquisition unit is used to acquire the echocardiogram datasets of non-aortic dissection and aortic dissection. Each data sample in the echocardiogram dataset of non-aortic dissection is labeled with a preset feature label, and each data sample in the echocardiogram dataset of aortic dissection is labeled with a preset feature label and an aortic dissection label.

[0041] The teacher network model training unit is used to train the teacher network model using the non-aortic dissection echocardiogram dataset with the task of identifying preset features of each data sample in the non-aortic dissection echocardiogram dataset until the teacher network model converges.

[0042] The recognition result acquisition unit is used to input data samples from the echocardiogram dataset of the aorta into the trained teacher network model and student network model respectively, and obtain the preset feature recognition result output by the trained teacher network model and the preset feature recognition result and aortic dissection recognition result output by the student network model.

[0043] The model parameter adjustment unit is used to adjust the model parameters of the student network model based on the differences between the preset feature recognition results output by the trained teacher network model and the preset feature recognition results output by the student network model, the differences between the preset feature recognition results output by the student network model and the preset feature label ground truth, and the differences between the aortic dissection recognition results output by the student network model and the aortic dissection label ground truth, until the student network model converges, and the trained student network model is determined as the aortic dissection recognition model; wherein, the initial model parameters of the teacher network model and the student network model are the same.

[0044] A fourth aspect of this application provides an aortic dissection identification device, comprising:

[0045] The image sequence acquisition unit is used to acquire the image sequence of the echocardiogram to be processed;

[0046] The image sequence processing unit is used to input the image sequence of the echocardiogram to be processed into the aortic dissection recognition model constructed by the aortic dissection recognition model construction method described in the first aspect or any implementation of the first aspect, and to obtain the aortic dissection recognition result output by the aortic dissection recognition model.

[0047] By employing the above technical solution, the invention provides a method for constructing an aortic dissection identification model, an identification method, and related devices. First, a teacher network model is trained using a non-aortic dissection echocardiographic dataset containing a large number of data samples. This trains the teacher network model to have good generalization ability and accurately identify preset features. Then, data samples from the aortic echocardiographic dataset are input into both the trained teacher network model and the student network model. Based on the differences between the trained teacher network model and the student network model in the preset feature identification results, as well as the differences between the student network model's identification results and the ground truth labels, the model parameters of the student network model are adjusted. This allows the student network model to learn the generalization ability of the teacher network model. This achieves the goal of training the student network model with a small number of aortic dissection data samples while ensuring the accuracy and reliability of the aortic dissection identification results. Attached Figure Description

[0048] 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 embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0049] Figure 1 This is a flowchart illustrating a method for constructing an aortic dissection identification model disclosed in an embodiment of this application;

[0050] Figure 2 This is a schematic diagram of the training process of a teacher network model disclosed in an embodiment of this application;

[0051] Figure 3 This is a schematic diagram of the training process of a student network model disclosed in an embodiment of this application;

[0052] Figure 4 This is a schematic flowchart of an aortic dissection identification method disclosed in an embodiment of the present invention;

[0053] Figure 5 This is a schematic flowchart of an aortic dissection identification method based on an aortic dissection identification model disclosed in an embodiment of the present invention;

[0054] Figure 6 This is a schematic diagram of the structure of an aortic dissection identification model construction device disclosed in an embodiment of the present invention;

[0055] Figure 7 This is a schematic diagram of the structure of an aortic dissection identification device disclosed in an embodiment of the present invention. Detailed Implementation

[0056] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0057] The embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.

[0058] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0059] Aortic dissection identification is one application scenario of echocardiographic identification. However, since aortic dissection (AD) is a rare cardiovascular disease, there are few echocardiographic data samples of aortic dissection. Machine learning models trained based on a small number of data samples have low reliability and no application value.

[0060] This application provides a method for constructing an aortic dissection identification model, an identification method, and related apparatus. First, a teacher network model is trained using a non-aortic dissection echocardiographic dataset containing a large number of data samples, enabling the trained teacher network model to have good generalization ability. Data samples from the aortic echocardiographic dataset are then input into both the trained teacher network model and the student network model. Based on the differences between the trained teacher network model and the student network model in the preset feature recognition results, and the differences between the student network model's recognition results and the ground truth labels, the model parameters of the student network model are adjusted. This allows the student network model to learn the generalization ability of the teacher network model. This allows the student network model to be trained using a small number of aortic dissection data samples while still ensuring the accuracy of aortic dissection identification, providing high reference value for the diagnosis of aortic dissection.

[0061] See Figure 1 , Figure 1 This is a flowchart illustrating a method for constructing an aortic dissection identification model provided in an embodiment of this application, as shown below. Figure 1 As shown in the embodiment of this application, a method for constructing an aortic dissection identification model may include steps 101 to 104, which are described in detail below.

[0062] 101. Obtain echocardiographic datasets for non-aortic dissection and aortic dissection;

[0063] Each data sample in the non-aortic dissection echocardiogram dataset and each data sample in the aortic dissection echocardiogram dataset are an echocardiogram image sequence, which consists of multiple frames.

[0064] Each data sample in the non-aortic dissection echocardiogram dataset has been labeled with a preset feature label. The preset features are common to echocardiograms, such as section type, end-systolic frame, and end-diastolic frame.

[0065] The section type label indicates the section type to which the data sample belongs. Section types include: apical four-chamber view, parasternal left ventricular long-axis view, parasternal right ventricular long-axis view, subxiphoid four-chamber view, and parasternal short-axis view. Different section type labels can be represented by different numbers. For example, the apical four-chamber view is labeled with 1, the parasternal left ventricular long-axis view with 2, the parasternal right ventricular long-axis view with 3, the subxiphoid four-chamber view with 4, and the parasternal short-axis view with 5.

[0066] The contraction end frame label indicates whether each frame in the data sample is a contraction end frame. For example, the i-th frame is represented by 1 if it is a contraction end frame, and by 0 if it is not a contraction end frame.

[0067] The end-diastolic frame label indicates whether each frame in the data sample is an end-diastolic frame. For example, the i-th frame is represented by 1 if it is an end-diastolic frame, and by 0 if it is not.

[0068] Each data sample in the echocardiographic dataset for aortic dissection has been labeled with a predefined feature label and an aortic dissection label. The predefined features in the aortic dissection echocardiographic dataset are of the same type as those in the non-aortic dissection echocardiographic dataset, and the labeling method for the predefined feature labels is also the same. The aortic dissection label includes aortic dissection positive and aortic dissection negative; for example, aortic dissection positive is represented by 1, and aortic dissection negative is represented by 0.

[0069] It should be noted that the number of data samples in the non-aortic dissection echocardiogram dataset is much larger than the number of data samples in the aortic dissection echocardiogram dataset.

[0070] 102. With the task of identifying the preset features of each data sample in the non-aortic dissection echocardiogram dataset, the teacher network model is trained using the non-aortic dissection echocardiogram dataset until the teacher network model converges.

[0071] During the training of the teacher network model, the input data of the teacher network model is the data sample from the echocardiogram dataset of non-aortic dissection, and the output data of the teacher network model is the preset feature recognition result.

[0072] For example, a loss function for the teacher network model is constructed based on the difference between the preset feature recognition results and preset feature labels output by the teacher network model. The teacher network model is then trained using a non-aortic dissection echocardiogram dataset until the loss function of the teacher network model converges, at which point the training of the teacher network model ends.

[0073] By training a teacher network model with a relatively large number of data samples and relatively simple task labels, the disadvantage of the small number of data samples in the echocardiogram dataset of aortic dissection is compensated. Moreover, the trained teacher network model has good generalization ability and can accurately identify the preset features.

[0074] 103. Input the data samples from the aortic echocardiogram dataset into the trained teacher network model and student network model respectively, and obtain the preset feature recognition results output by the trained teacher network model and the preset feature recognition results and aortic dissection recognition results output by the student network model.

[0075] In step 102 above, the teacher network model converges, and after the training of the teacher network model ends, a trained teacher network model is obtained.

[0076] The data samples input into the trained teacher network model can be exactly the same as those input into the student network model, or they can be different. For example, the data samples input into the student network model are the original forward sequence of images, while the data samples input into the trained teacher network model are the reverse sequence of images and / or the image sequence after cropping specific regions of the data samples.

[0077] 104. Based on the differences between the preset feature recognition results output by the trained teacher network model and the preset feature recognition results output by the student network model, the differences between the preset feature recognition results output by the student network model and the preset feature label ground truth, and the differences between the aortic dissection recognition results output by the student network model and the aortic dissection label ground truth, adjust the model parameters of the student network model until the student network model converges, and determine the trained student network model as the aortic dissection recognition model.

[0078] The teacher and student network models share the same initial model parameters and structure, such as being based on CNN (Convolutional Neural Networks) or DNN (Deep Neural Networks). Considering that aortic dissection identification typically occurs in hospital emergency rooms or cardiology departments, where device performance and resources for image tensor operations are limited, a high-performance, fewer-parameter, and faster neural network model, such as the lightweight convolutional neural network MobileNet, can be chosen.

[0079] Based on the differences between the preset feature recognition results output by the trained teacher network model and the preset feature recognition results output by the student network model, the differences between the preset feature recognition results output by the student network model and the ground truth values ​​of the preset feature labels, and the differences between the aortic dissection recognition results output by the student network model and the ground truth values ​​of the aortic dissection labels, the model parameters of the student network model are adjusted. Specifically, the trained teacher network model is used to perform self-distillation on the student network model with the same structure. By employing self-distillation, the knowledge learned by the trained teacher network model is transferred to another student network model with the same structure, thereby improving the performance of the student network model. The fact that the student network model is structurally identical to the teacher network model helps it accurately learn the knowledge learned by the trained teacher network model, while also reducing the risk of overfitting.

[0080] One possible way to train a student network model is as follows: a loss function is constructed based on the differences between the preset feature recognition results output by the trained teacher network model and the preset feature recognition results output by the student network model, the differences between the preset feature recognition results output by the student network model and the preset feature label ground truth, and the differences between the aortic dissection recognition results output by the student network model and the aortic dissection label ground truth. If the loss function converges during the training process, the training of the student network model is terminated.

[0081] This embodiment discloses a method that firstly trains a teacher network model using a non-aortic dissection echocardiogram dataset containing a large number of data samples, enabling the trained teacher network model to have good generalization ability and accurately identify preset features. Then, data samples from the aortic echocardiogram dataset are input into both the trained teacher network model and the student network model. Based on the differences between the trained teacher network model and the student network model in the preset feature recognition results, as well as the differences between the student network model's recognition results and the ground truth labels, the model parameters of the student network model are adjusted. This allows the student network model to learn the generalization ability of the teacher network model, achieving the goal of training the student network model with a small number of aortic dissection data samples while ensuring the accuracy and reliability of the aortic dissection recognition results.

[0082] The following describes step 102 in the above embodiment, namely the process of training the teacher network model. One possible implementation includes the following steps A1 to A2:

[0083] A1: Based on the difference between the teacher network model's recognition results of preset features in the data samples and the true values ​​of preset feature labels, construct the loss function of the teacher network model;

[0084] The loss function of the teacher network model represents the difference between the teacher network model's recognition result of preset features in the data sample and the true value of the preset feature label.

[0085] A2: During the training of the teacher network model using the non-aortic dissection echocardiogram dataset, the gradient descent method is used to optimize the loss function of the teacher network model until the loss function of the teacher network model converges, at which point the training of the teacher network model ends.

[0086] Gradient descent is used to find the local minimum of the loss function of the teacher network model. In each iteration of the training process, a certain step size is moved along the direction of the negative gradient (i.e. the direction of the fastest descent) until the minimum value of the loss function is reached.

[0087] The mathematical representation of gradient descent is as follows:

[0088]

[0089] Where, θ 0 These are the parameters of the current teacher network and student network models, J(θ) is the loss function, and α is the learning rate. In the current parameter θ 0 The gradient at θ 1 These are the updated teacher network and student network model parameters.

[0090] Furthermore, since this application identifies aortic dissection based on echocardiography, and the image sequences in the echocardiography dataset for aortic dissection typically contain only 20% positive frames for aortic dissection, the proportion of positive frames for aortic dissection in the entire video input is relatively low. This can easily lead to overfitting during the training of the student network model, making it difficult to guarantee the model's robustness. One possible implementation of this embodiment includes preset feature labels, including spatial and temporal feature labels, and uses a self-supervised method. The trained teacher network model performs self-distillation on the student network model with the same structure in the spatial and temporal dimensions of the image sequence, improving the student network model's generalization ability in both space and time.

[0091] Based on this, another possible implementation of step 102 in the above embodiments includes the following steps B1 to B4:

[0092] B1: Construct a spatial consistency loss function based on the difference between the spatial feature recognition results of the teacher network model in the data sample and the true values ​​of the spatial feature labels;

[0093] For example, the spatial feature is the section category.

[0094] Spatial consistency loss function l space as follows:

[0095]

[0096] Where, p i y represents the probability that the data sample output by the teacher network model belongs to the i-th slice category. i represents the ground truth value of the section category label, and N represents the number of section categories in the echocardiogram dataset for non-aortic dissection.

[0097]

[0098] B2: Based on the difference between the time feature identification results in the data samples by the teacher network model and the true values ​​of the time feature labels, construct a time consistency loss function;

[0099] For example, temporal features include: end-contraction frames and end-diastolic frames.

[0100] Time consistency loss function l time as follows:

[0101] l time =l s +l d ;

[0102] Among them, l s and l d Let represent the loss functions for the end-contraction frame and the end-diastole frame, respectively.

[0103]

[0104] This represents the probability that the i-th frame of the image sequence output by the teacher network model is a frame at the end of the contraction phase. This represents the true value of the frame label at the end of the contraction period, and M is the total number of frames in the image sequence.

[0105]

[0106] Similarly,

[0107] This represents the probability that the i-th frame of the image sequence output by the teacher network model is an end-diastolic frame. This represents the true value of the frame label at the end of the contraction period, and M is the total number of frames in the image sequence.

[0108]

[0109] B3: Construct the loss function for the teacher network model based on the spatial consistency loss function and the temporal consistency loss function;

[0110] For example, a loss function l of a teacher network model teacher as follows:

[0111] l teacher =l space +l time .

[0112] B4: During the training of the teacher network model using the non-aortic dissection echocardiogram dataset, the gradient descent method is used to optimize the loss function of the teacher network model until the loss function of the teacher network model converges, at which point the training of the teacher network model ends.

[0113] The specific implementation of B4 is the same as that of A2, so it will not be repeated here.

[0114] The following describes step 104 in the above embodiment, namely the process of training the student network model. One possible implementation includes the following steps C1 to C5:

[0115] C1: Construct the first loss function based on the difference between the preset feature recognition results output by the trained teacher network model and the preset feature recognition results output by the student network model;

[0116] In one possible implementation, the preset features include spatial and temporal features. The specific implementation of C1 includes the following steps C11–C13:

[0117] C11: Construct a first spatial consistency loss function based on the first spatial feature recognition results output by the trained teacher network model, the second spatial feature recognition results output by the student network model, and the preset distillation temperature;

[0118] For example, the spatial feature is the section category.

[0119] First spatial consistency loss function l spacesoft as follows:

[0120]

[0121] in:

[0122]

[0123] T is the preset distillation temperature, which can be set to 2;

[0124] N represents the number of section categories in the echocardiographic dataset of aortic dissection;

[0125] q i p represents the probability that the data sample output by the trained teacher network model belongs to the i-th slice category. i This represents the probability that the data sample output by the student network model belongs to the i-th slice category.

[0126] C12: Construct a first-time consistency loss function based on the first-time feature recognition results output by the trained teacher network model, the second-time feature recognition results output by the student network model, and the preset distillation temperature;

[0127] For example, temporal features include: end-contraction frames and end-diastolic frames.

[0128] First-time consistency loss function l timesoft as follows:

[0129] l timesoft =l softs +l softd ;

[0130] l softs Let l represent the loss function for the final frame of the first contraction period. softs Let represent the loss function for the first end-diastolic frame.

[0131]

[0132] T is the preset distillation temperature, which can be set to 2;

[0133] M is the total number of frames in the image sequence;

[0134] q si p represents the probability that the i-th frame of the image sequence output by the teacher network model is a frame at the end of the contraction phase. sj This represents the probability that the j-th frame of the image sequence output by the student network model is a frame at the end of the contraction phase.

[0135] Constructing the loss function l for the first end-diastolic frame softd The loss function l of the first contraction terminal frame softs The principle is the same, so I won't go into details here.

[0136] C13: Construct the first loss function based on the first spatial consistency loss function and the first temporal consistency loss function.

[0137] The first loss function is λ soft l spacesoft +λ soft l timesoft , where λ soft This represents the hyperparameters of the first loss function.

[0138] C2: Construct a second loss function based on the difference between the preset feature recognition results output by the student network model and the preset feature label ground truth values;

[0139] Based on the preset features including spatial and temporal features, the specific implementation of C2 includes the following steps C21 to C23:

[0140] C21: Construct a second spatial consistency loss function based on the difference between the spatial feature recognition results output by the student network model and the true values ​​of the spatial feature labels;

[0141] For example, the spatial feature is the section category.

[0142] The second space consistency loss function is l spacehard as follows:

[0143]

[0144] Where, p iy represents the probability that the data sample output by the student network model belongs to the i-th slice category. i represents the ground truth value of the section category label, and N represents the number of section categories in the echocardiogram dataset of aortic dissection.

[0145] C22: Construct a second time consistency loss function based on the difference between the time feature recognition results output by the student network model and the true values ​​of the time feature labels;

[0146] For example, temporal features include: end-contraction frames and end-diastolic frames.

[0147] Second time consistency loss function l timehard as follows:

[0148] l timehard =l hards +l hardd ;

[0149] l hards Let l represent the loss function for the second contraction terminal frame. hardd This represents the loss function for the second end-diastolic frame.

[0150]

[0151] This represents the probability that the i-th frame of the image sequence output by the student network model is a frame at the end of the contraction phase. This represents the true value of the frame label at the end of the contraction period. This represents the probability that the i-th frame of the image sequence output by the student network model is an end-diastolic frame. This represents the true value of the frame label at the end of the contraction period, and M is the total number of frames in the image sequence.

[0152] C23: Construct a second loss function based on the second spatial consistency loss function and the second temporal consistency loss function.

[0153] The second loss function is λ. hard l spacehard +λ hard l timehard , where λ hard This represents the hyperparameters of the second loss function.

[0154] C3: Construct an aortic dissection probability prediction loss function based on the difference between the aortic dissection identification results output by the student network model and the ground truth value of the aortic dissection label;

[0155] For example, the aortic dissection probability prediction loss function l ad as follows:

[0156] l ad= -[ylogp+(1-y)log(1-p)]

[0157] Among them, l ad denoted by cross-entropy loss, p represents the aortic dissection identification result output by the student network model. For example, p represents the predicted probability of aortic dissection positivity output by the student network model, expressed as a percentage.

[0158] y is the true value for the aortic dissection label. For example, a positive aortic dissection is represented by 1, and a negative aortic dissection is represented by 0.

[0159] C4: Construct the loss function for the student network model based on the first loss function, the second loss function, and the aortic dissection probability prediction loss function;

[0160] In summary, the loss function l of the student network model student as follows:

[0161] l student =λ soft l spacesoft +λ hard l spacehard +λ soft l timesoft +λ hard l timehard +l ad

[0162] C5: Optimize the loss function of the student network model using gradient descent until the loss function of the student network model converges, then end the training of the student network model.

[0163] In the process of optimizing the loss function of the student network model using gradient descent, the exponential moving average (EMA) method can also be used to update the weights of the student network model.

[0164] V t =α·V t-1 +(1-α)θ t

[0165] Where t is the number of training steps, V t and V t-1 It is the average of the model parameters in the first t and the first t-1 steps, θ t This is the updated weight for the current training steps, α = 0.99.

[0166] This embodiment trains a teacher network model with a relatively large amount of data and relatively simple task labels. This not only compensates for the disadvantage of limited data in aortic dissection but also trains the teacher network model's feature extraction capabilities from echocardiography. Simultaneously, the two simple tasks focus on the spatial and temporal dimensions respectively. After accurately learning the knowledge acquired by the trained teacher network model, the student network model also maintains its sensitivity to the spatial and temporal dimensions of echocardiography. Compared to existing methods that directly train machine learning models using data samples, this approach requires a significantly smaller amount of actual task data.

[0167] Furthermore, in one possible implementation, the method for obtaining the first spatial feature recognition result and the second spatial feature recognition result in the above embodiments specifically includes the following steps D1 to D3:

[0168] D1: For each data sample in the echocardiogram dataset of aortic dissection, the forward sequence of the data sample is reversed in the time dimension to obtain the reverse sequence of the image.

[0169] For example, if the forward sequence of images in the data sample is [t1,t2,t3,…,tn], then the reverse sequence of images is obtained by reversing the forward sequence in the time dimension, which is [tn,…,t3,t2,t1], where n is the total number of frames in the forward and reverse sequences.

[0170] D2: Input the reverse image sequence into the trained teacher network model to obtain the first teacher feature map, and input the forward image sequence into the student network model to obtain the student feature map;

[0171] D3: Input the first teacher feature map into the spatial feature recognition branch of the trained teacher network model to obtain the first spatial feature recognition result, and input the student feature map into the spatial feature recognition branch of the student network model to obtain the second spatial feature recognition result.

[0172] When processing time-series data, such as image sequences, models typically need to understand the temporal dynamics of the data. This embodiment reverses the order of the image sequences input to the trained teacher network model. Through knowledge distillation by the trained teacher network model, the student network model learns not only to recognize forward relationships between frames but also to understand inverse relationships. This helps the student network model gain a deeper understanding of the inherent patterns in time series data. By changing the order of the input data, the student network model learns to recognize not only data in a specific order but also more generalized features, thereby improving its generalization ability when faced with new data and reducing the risk of overfitting.

[0173] Furthermore, in one possible implementation, the method for obtaining the first temporal feature recognition result and the second temporal feature recognition result in the above embodiments specifically includes the following steps E1 to E3:

[0174] E1: For each data sample in the echocardiogram dataset of aortic dissection, the region of interest in the aortic region is cropped from the forward sequence of images in the data sample, and the region of interest is interpolated to the input size of the trained teacher network model to obtain the aortic region image sequence.

[0175] For example, based on the region of interest (ROI) of the aortic region pre-labeled in the data sample, the region of interest of the aortic region is cropped from the forward sequence of images in the data sample.

[0176] In the case of aortic dissection echocardiogram datasets, manual calibration can be used to improve the accuracy of region of interest (ROI) labeling in the aortic region. Alternatively, to improve the efficiency of ROI labeling in the aortic region, data samples from the aortic dissection echocardiogram dataset can be input into a pre-built target detection model. This allows for the automatic identification of ROIs within the aortic region using target detection technology, thus achieving automatic ROI labeling in the aortic region.

[0177] Interpolation of the region of interest can be performed using any existing image interpolation algorithm, such as nearest neighbor interpolation, bilinear interpolation, cubic convolution interpolation, etc., and this invention does not impose any specific limitations.

[0178] E2: Input the aortic region image sequence into the trained teacher network model to obtain the second teacher feature map, and input the forward sequence of images into the student network model to obtain the student feature map;

[0179] E3: Input the second teacher feature map into the time feature recognition branch of the trained teacher network model to obtain the first time feature recognition result, and input the student feature map into the time feature recognition branch of the student network model to obtain the second time feature recognition result.

[0180] This embodiment performs local cropping on the image sequence input to the trained teacher network model, and uses the trained teacher network model to distill the knowledge of the student network model. This enables the student network model to not only learn the image features in the original image sequence, but also to focus on learning the image features of the aortic region, thereby improving the generalization ability of the student network model when facing new data and reducing the risk of overfitting.

[0181] In summary, this embodiment utilizes a knowledge distillation method in both time and space, employing self-distillation techniques. By reversing the temporal order and performing local pruning in space, it makes the student network model more robust to consistency after temporal and spatial transformations, resulting in better generalization ability compared to existing direct training methods.

[0182] To facilitate understanding of the aortic dissection identification model construction method provided in this application, a specific example is described below, which includes steps F1 to F3.

[0183] F1: Obtain the echocardiographic datasets for non-aortic dissection and aortic dissection.

[0184] Each data sample in the non-aortic dissection echocardiogram dataset has been labeled with the section category, end-systolic frame, and end-diastolic frame, while each data sample in the aortic dissection echocardiogram dataset has been labeled with the section category, end-systolic frame, end-diastolic frame, and aortic dissection label.

[0185] Among them, the section type represents the spatial characteristics of echocardiography, while the end-systolic frame and end-diastolic frame represent the temporal characteristics of echocardiography.

[0186] F2: Train the teacher network model using a non-aortic dissection echocardiogram dataset until the teacher network model converges.

[0187] The loss function l for constructing the teacher network model teacher :

[0188] l teacher =l space +l time ;

[0189] l space Let l represent the spatial consistency loss function. time This represents the time consistency loss function.

[0190]

[0191] Where, p i y represents the probability that the data sample output by the teacher network model belongs to the i-th slice category. i represents the ground truth value of the section category label, and N represents the number of section categories in the echocardiogram dataset for non-aortic dissection.

[0192]

[0193] l time =l s +l d ;

[0194] Among them, l s and l d Let represent the loss functions for the end-contraction frame and the end-diastole frame, respectively.

[0195]

[0196] This represents the probability that the i-th frame of the image sequence output by the teacher network model is a frame at the end of the contraction phase. This represents the true value of the frame label at the end of the contraction period, and M is the total number of frames in the image sequence.

[0197]

[0198] Similarly,

[0199] This represents the probability that the i-th frame of the image sequence output by the teacher network model is an end-diastolic frame. This represents the true value of the frame label at the end of the contraction period, and M is the total number of frames in the image sequence.

[0200]

[0201] Please see Figure 2 The diagram illustrates the training process of the teacher network model. During the training of the teacher network model, the data samples of the non-aortic dissection echocardiogram dataset input into the teacher network model can be data samples after random data augmentation. Random data augmentation includes, but is not limited to, at least one of the following random data augmentation methods:

[0202] 1. Partially crop the non-target area, which is the area excluding the aortic region;

[0203] 2. Partially cover the target area, which is the aortic region;

[0204] 3. Apply perspective transformation;

[0205] 4. Rotate by ±20 degrees;

[0206] 5. Implement vertical and horizontal compression;

[0207] 6. Use a mean filter of 1 to 2 times;

[0208] 7. Adjust the brightness, contrast, and saturation, with the range of change between 0.5 and 1.5 times.

[0209] In other words, random data augmentation can be performed by choosing one or more of the above-mentioned random data augmentation methods and applying them in combination.

[0210] Data samples from the non-aortic dissection echocardiogram dataset were input into the teacher network model to obtain teacher feature maps. The teacher feature maps were then input into the section category classification branch and the end-systolic / end-diastolic frame detection branch to obtain section category recognition results, end-systolic frame recognition results, and end-diastolic frame recognition results.

[0211] Among them, the section category recognition result is compared with the ground truth value of the section category label pre-labeled in the data sample (i.e., Figure 2 The difference between the true values ​​of the aspect categories in the table corresponds to the aspect category loss, which is the spatial consistency loss function mentioned above.

[0212] The results of the end-contraction frame identification are compared with the ground truth of the pre-labeled end-contraction frame labels in the data samples (i.e., Figure 2 The difference between the end-systolic frame ground truth (in the data sample) and the end-diastolic frame identification result corresponds to the end-systolic frame loss. Figure 2 The difference between the true value of the end-diastolic frame and the true value of the end-diastolic frame corresponds to the end-diastolic frame loss, which is the loss function between the end-contraction frame and the end-diastolic frame mentioned above.

[0213] During the training of the teacher network model using the aforementioned non-aortic dissection echocardiographic dataset, gradient descent was used to optimize the loss function l of the teacher network model. teacher The training of the teacher network model continues until the loss function of the teacher network model converges, at which point the training of the teacher network model ends.

[0214] F3: Input the data samples from the aortic echocardiogram dataset into the trained teacher network model and student network model respectively, and obtain the preset feature recognition results output by the trained teacher network model and the preset feature recognition results and aortic dissection recognition results output by the student network model. Based on the differences between the preset feature recognition results output by the trained teacher network model and the preset feature recognition results output by the student network model, the differences between the preset feature recognition results output by the student network model and the preset feature label ground truth, and the differences between the aortic dissection recognition results output by the student network model and the aortic dissection label ground truth, adjust the model parameters of the student network model until the student network model converges. The trained student network model is then determined as the aortic dissection recognition model.

[0215] That is, using a trained teacher network model to perform self-distillation on a student network model with the same structure. By using self-distillation, the knowledge learned by the trained teacher network model is transferred to another student network model with the same structure.

[0216] The loss function l for constructing the student network model student as follows:

[0217] l student =λ soft l spacesoft +λ hard l spacehard +λ soft l timesoft +λ hard l timehard +l ad .

[0218] in:

[0219] l spacesoft Represents the first spatial consistency loss function;

[0220] l timesoft This represents the first-time consistency loss function;

[0221] l spacehard Represents the second-space consistency loss function;

[0222] l timehard This represents the second time consistency loss function;

[0223] l ad This represents the loss function for predicting the probability of aortic dissection.

[0224] λ soft and λ hard Hyperparameters representing the differences in different losses.

[0225] For the first spatial consistency loss function l spacesoft Please see Figure 3 The diagram illustrates the training process of the student network model. For each data sample in the echocardiogram dataset of aortic dissection, the forward-order image sequence [t1,t2,t3,…,tn] in the data sample is reversed in time to obtain the reverse-order image sequence [tn,…,t3,t2,t1], where n is the total number of frames in the forward and reverse-order image sequences. The reverse-order image sequence [tn,…,t3,t2,t1] is input into the trained teacher network model to obtain the first teacher feature map. Then, the first teacher feature map is input into the section category classification branch of the trained teacher network model to obtain the section category recognition result [q1, q2,...,q]. i , ..., q N ]] represents the pseudo-label of the teacher for each segment category. The forward sequence of images [t1, t2, t3, ..., tn] is input into the student network model to obtain the student feature map. Then, the student feature map is input into the segment category classification branch of the student network model to obtain the segment category recognition result [p1, p2, ..., pn]. i , ..., p N] represents the predicted probability that a data sample belongs to each section category, and N represents the number of section categories in the echocardiographic dataset of aortic dissection. The section category recognition results [q1, q2, ..., q] output by the trained teacher network model are also included. i , ..., q N The results of the facet category recognition output by the student network model [p1, p2, ..., p] i , ..., p N The difference between them corresponds to the student-teacher network space consistency loss, which is the first space consistency loss function mentioned above.

[0226] First spatial consistency loss function l spacesoft as follows:

[0227]

[0228] in:

[0229]

[0230] T is the preset distillation temperature, which can be set to 2;

[0231] N represents the number of section categories in the echocardiographic dataset of aortic dissection;

[0232] q i p represents the probability that the data sample output by the trained teacher network model belongs to the i-th slice category. i This represents the probability that the data sample output by the student network model belongs to the i-th slice category.

[0233] The second space consistency loss function is l spacehard as follows:

[0234]

[0235] Where, p i y represents the probability that the data sample output by the student network model belongs to the i-th slice category. i represents the ground truth value of the section category label, and N represents the number of section categories in the echocardiogram dataset of aortic dissection.

[0236] For the first-time consistency loss function l timesoft Please see Figure 3 The diagram illustrates the training process of the student network model. For each data sample in the echocardiogram dataset of aortic dissection, a region of interest (ROI) in the aortic region is cropped from the forward-order image sequence of the data sample. This ROI is then interpolated to the input size of the trained teacher network model, resulting in an aortic region image sequence. Figure 3The aortic region image sequence is also represented by [t1,t2,t3,…,tn]. The aortic region image sequence is input into the trained teacher network model to obtain the second teacher feature map. This second teacher feature map is then input into the end-systolic / end-diastolic frame detection branch of the trained teacher network model to obtain the end-systolic frame recognition result and the end-diastolic frame recognition result. Similarly, the forward sequence image of each data sample from the aortic dissection echocardiogram dataset is input into the student network model to obtain the student feature map. This student feature map is then input into the end-systolic / end-diastolic frame detection branch of the student network model to obtain the end-systolic frame recognition result. The results include the identification of the end-diastolic frames. Specifically, the results of the identification of the end-contraction frames are output by the trained teacher network model. The results of the shrinkage terminal frame recognition output by the student network model The differences between the student and teacher network models, specifically the differences between the end-diastolic frame recognition results output by the trained teacher network model and the student network model, correspond to the student-teacher network temporal consistency loss, i.e., the first temporal consistency loss function mentioned above. timesoft .

[0237] First-time consistency loss function l timesoft as follows:

[0238] l timesoft =l softs +l softd ;

[0239] l softd Let l represent the loss function for the final frame of the first contraction period. softd Let represent the loss function for the first end-diastolic frame.

[0240]

[0241] T is the preset distillation temperature, which can be set to 2;

[0242] M is the total number of frames in the image sequence;

[0243] q si p represents the probability that the i-th frame of the image sequence output by the teacher network model is a frame at the end of the contraction phase. sj This represents the probability that the j-th frame of the image sequence output by the student network model is a frame at the end of the contraction phase.

[0244] Constructing the loss function l for the first end-diastolic frame softd The loss function l of the first contraction terminal frame softs The principle is the same, so I won't go into details here.

[0245] Second time consistency loss function l timehard as follows:

[0246] l timehard =l hards +l hardd ;

[0247] l hards Let l represent the loss function for the second contraction terminal frame. hardd This represents the loss function for the second end-diastolic frame.

[0248]

[0249] This represents the probability that the i-th frame of the image sequence output by the student network model is a frame at the end of the contraction phase. This represents the true value of the frame label at the end of the contraction period. This represents the probability that the i-th frame of the image sequence output by the student network model is an end-diastolic frame. This represents the true value of the frame label at the end of the contraction period, and M is the total number of frames in the image sequence.

[0250] For the probability prediction loss function of aortic dissection l ad Please see Figure 3 The diagram illustrates the training process of the student network model. The forward sequence of images [t1, t2, t3, ..., tn] from each data sample in the aortic dissection echocardiogram dataset is input into the student network model to obtain a student feature map. This student feature map is then input into the aortic dissection judgment branch of the student network model to obtain the aortic dissection identification result. The difference between the aortic dissection identification result and the pre-labeled aortic dissection ground truth value in the data samples corresponds to the aortic dissection probability prediction loss function l. ad .

[0251] Aortic dissection probability prediction loss function l ad as follows:

[0252] l ad = -[ylogp+(1-y)log(1-p)];

[0253] Among them, l ad denoted by cross-entropy loss, p represents the aortic dissection identification result output by the student network model. For example, p represents the predicted probability of aortic dissection positivity output by the student network model, expressed as a percentage.

[0254] Optimize the loss function of the student network model using gradient descent. student The training of the student network model continues until the loss function of the student network model converges, at which point the training of the student network model ends, and the trained student network model is determined as the aortic dissection identification model.

[0255] Based on the aortic dissection identification model construction method described in any of the above embodiments, this embodiment discloses an aortic dissection identification method. Please refer to [link to relevant documentation]. Figure 4 The method specifically includes the following steps:

[0256] 401: Acquire the image sequence of the echocardiogram to be processed;

[0257] The image sequence of the echocardiogram to be processed can be the original echocardiogram image sequence output by the ultrasound scanning device, or it can be the image sequence after preprocessing the original echocardiogram image sequence output by the ultrasound scanning device. Preprocessing includes, but is not limited to, cropping the image sequence of the original echocardiogram output by the ultrasound scanning device to the input data format size of the aortic dissection identification model.

[0258] 402: Input the image sequence of the echocardiogram to be processed into the aortic dissection recognition model to obtain the aortic dissection recognition result output by the aortic dissection recognition model.

[0259] Specifically, the echocardiogram image sequence to be processed is input into the aortic dissection judgment branch in the aortic dissection recognition model to obtain the aortic dissection recognition result output by the aortic dissection judgment branch.

[0260] like Figure 5 As shown, the echocardiogram image sequence to be processed is represented by [t1,t2,t3,…,tn]. The echocardiogram image sequence to be processed is input into the aortic dissection recognition model to obtain the feature map, and the feature map is input into the aortic dissection judgment branch to obtain the aortic dissection recognition result output by the aortic dissection judgment branch.

[0261] For example, the aortic dissection identification result is the predicted probability of a positive aortic dissection, expressed as a percentage. The higher the predicted probability of aortic dissection, the greater the probability that the echocardiogram to be processed will characterize aortic dissection.

[0262] This embodiment discloses an aortic dissection identification method. By employing a highly accurate aortic dissection identification model, the method ensures the reliability of the aortic dissection identification results, providing doctors with a reference and offering high reference value for the diagnosis of aortic dissection.

[0263] Based on the aortic dissection identification model construction method disclosed in the above embodiments, this embodiment correspondingly discloses an aortic dissection identification model construction device. Please refer to... Figure 6 The device includes:

[0264] The dataset acquisition unit 601 is used to acquire the echocardiogram dataset of non-aortic dissection and the echocardiogram dataset of aortic dissection. Each data sample in the echocardiogram dataset of non-aortic dissection has been labeled with a preset feature label, and each data sample in the echocardiogram dataset of aortic dissection has been labeled with a preset feature label and an aortic dissection label.

[0265] The teacher network model training unit 602 is used to train the teacher network model using the non-aortic dissection echocardiogram dataset with the task of identifying preset features of each data sample in the non-aortic dissection echocardiogram dataset until the teacher network model converges.

[0266] The recognition result acquisition unit 603 is used to input the data samples in the echocardiogram dataset of the aorta into the trained teacher network model and student network model respectively, and obtain the preset feature recognition result output by the trained teacher network model and the preset feature recognition result and aortic dissection recognition result output by the student network model.

[0267] The model parameter adjustment unit 604 is used to adjust the model parameters of the student network model based on the differences between the preset feature recognition results output by the trained teacher network model and the preset feature recognition results output by the student network model, the differences between the preset feature recognition results output by the student network model and the preset feature label ground truth, and the differences between the aortic dissection recognition results output by the student network model and the aortic dissection label ground truth, until the student network model converges, and the trained student network model is determined as the aortic dissection recognition model; wherein, the initial model parameters of the teacher network model and the student network model are the same.

[0268] In one possible implementation, the model parameter adjustment unit 604 includes:

[0269] The first loss function construction subunit is used to construct a first loss function based on the difference between the preset feature recognition results output by the trained teacher network model and the preset feature recognition results output by the student network model.

[0270] The second loss function construction subunit is used to construct a second loss function based on the difference between the preset feature recognition result output by the student network model and the preset feature label true value.

[0271] The third loss function construction subunit is used to construct the aortic dissection probability prediction loss function based on the difference between the aortic dissection identification result output by the student network model and the aortic dissection label true value.

[0272] The fourth loss function construction subunit is used to construct the loss function of the student network model based on the first loss function, the second loss function and the aortic dissection probability prediction loss function;

[0273] The first loss function optimization subunit is used to optimize the loss function of the student network model using gradient descent until the loss function of the student network model converges, thus ending the training of the student network model.

[0274] In one possible implementation, the preset features include spatial features and temporal features. The first loss function construction sub-unit is specifically used to construct the first spatial consistency loss function based on the first spatial feature recognition result output by the trained teacher network model, the second spatial feature recognition result output by the student network model, and the preset distillation temperature; to construct the first temporal consistency loss function based on the first temporal feature recognition result output by the trained teacher network model, the second temporal feature recognition result output by the student network model, and the preset distillation temperature; and to construct the first loss function based on the first spatial consistency loss function and the first temporal consistency loss function.

[0275] In one possible implementation, the recognition result acquisition unit 603 includes:

[0276] The first identification result acquisition subunit is used to, for each data sample in the echocardiogram dataset of the aortic dissection, reverse the forward-order image sequence in the time dimension to obtain a reverse-order image sequence; input the reverse-order image sequence into the trained teacher network model to obtain a first teacher feature map, and input the forward-order image sequence into the student network model to obtain a student feature map; input the first teacher feature map into the spatial feature recognition branch of the trained teacher network model to obtain a first spatial feature recognition result, and input the student feature map into the spatial feature recognition branch of the student network model to obtain a second spatial feature recognition result.

[0277] In one possible implementation, the recognition result acquisition unit 603 includes:

[0278] The second identification result acquisition subunit is used to, for each data sample in the echocardiogram dataset of the aortic dissection, crop out the region of interest (ROI) of the aortic region from the chronological image sequence of the data sample, and interpolate the ROI to the input size of the trained teacher network model to obtain an aortic region image sequence; input the aortic region image sequence into the trained teacher network model to obtain a second teacher feature map, and input the chronological image sequence into the student network model to obtain a student feature map; input the second teacher feature map into the time feature recognition branch of the trained teacher network model to obtain a first time feature recognition result, and input the student feature map into the time feature recognition branch of the student network model to obtain a second time feature recognition result.

[0279] In one possible implementation, the preset features include spatial features and temporal features. The second loss function construction sub-unit is specifically used to construct a second spatial consistency loss function based on the difference between the spatial feature recognition result output by the student network model and the true value of the spatial feature label; to construct a second temporal consistency loss function based on the difference between the temporal feature recognition result output by the student network model and the true value of the temporal feature label; and to construct the second loss function based on the second spatial consistency loss function and the second temporal consistency loss function.

[0280] In one possible implementation, the preset features include spatial features and temporal features. The teacher network model training unit 602 is specifically used to construct the loss function of the teacher network model based on the spatial consistency loss function and the temporal consistency loss function. The spatial consistency loss function represents the difference between the recognition result of the teacher network model for the spatial features and the ground truth value of the spatial feature label, and the temporal consistency loss function represents the difference between the recognition result of the teacher network model for the temporal features and the ground truth value of the temporal feature label. During the training of the teacher network model using the non-aortic dissection echocardiogram dataset, the gradient descent method is used to optimize the loss function of the teacher network model until the loss function of the teacher network model converges, thus ending the training of the teacher network model.

[0281] In one possible implementation, the spatial feature is a slice category, and the temporal feature includes end-contraction frames and end-diastolic frames.

[0282] This embodiment discloses an aortic dissection recognition model construction device. First, a teacher network model is trained using a non-aortic dissection echocardiogram dataset containing a large number of data samples. This trains the teacher network model to have good generalization ability and accurately identify preset features. Then, data samples from the aortic echocardiogram dataset are input into both the trained teacher network model and the student network model. Based on the differences in preset feature recognition results between the trained teacher network model and the student network model, and the differences between the student network model's recognition results and the ground truth labels, the model parameters of the student network model are adjusted. This allows the student network model to learn the generalization ability of the teacher network model. This achieves the goal of training the student network model with a small number of aortic dissection data samples while ensuring the accuracy and reliability of the aortic dissection recognition results.

[0283] Based on the aortic dissection identification method disclosed in the above embodiments, this embodiment correspondingly discloses an aortic dissection identification device. Please refer to... Figure 7 The device includes:

[0284] Image sequence acquisition unit 701 is used to acquire the image sequence of the echocardiogram to be processed;

[0285] The image sequence processing unit 702 is used to input the image sequence of the echocardiogram to be processed into the aortic dissection recognition model constructed by the aortic dissection recognition model construction method described in any of the above embodiments, and obtain the aortic dissection recognition result output by the aortic dissection recognition model.

[0286] This embodiment discloses an aortic dissection identification device that uses a highly accurate aortic dissection identification model to identify aortic dissection, ensuring the reliability of the identification results for doctors' reference and providing high reference value for the diagnosis of aortic dissection.

[0287] This embodiment also discloses a computer program product, including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement the aortic dissection identification model construction method or aortic dissection identification method as described in any of the above embodiments.

[0288] This embodiment also discloses an electronic device, including at least one processor and a memory connected to the processor, wherein:

[0289] The memory is used to store computer programs;

[0290] The processor is used to execute the computer program so that the electronic device can implement the aortic dissection identification model construction method or aortic dissection identification method as described in any of the above embodiments.

[0291] For example, electronic devices can be PC (Personal Computer) devices, such as desktop computers, laptops, mini-laptops, tablets, and ultrabooks, etc. Electronic devices can also be servers or server clusters, etc.

[0292] This embodiment also discloses a computer storage medium carrying one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement the aortic dissection identification model construction method or the aortic dissection identification method as described in any of the above embodiments.

[0293] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0294] It should also 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.

[0295] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0296] The above embodiments can be combined arbitrarily. The descriptions of the disclosed embodiments and the features recorded in the embodiments of this specification can be substituted or combined with each other, so that those skilled in the art can implement or use this application.

[0297] 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 method for constructing an aortic dissection identification model, characterized in that, include: Obtain echocardiographic datasets of non-aortic dissection and aortic dissection. Each data sample in the non-aortic dissection echocardiographic dataset has been labeled with a preset feature label, and each data sample in the aortic dissection echocardiographic dataset has been labeled with a preset feature label and an aortic dissection label. The task is to identify the preset features of each data sample in the non-aortic dissection echocardiogram dataset. The teacher network model is trained using the non-aortic dissection echocardiogram dataset until the teacher network model converges. The data samples from the echocardiogram dataset of the aortic dissection are respectively input into the trained teacher network model and student network model to obtain the preset feature recognition results output by the trained teacher network model and the preset feature recognition results and aortic dissection recognition results output by the student network model. Based on the differences between the preset feature recognition results output by the trained teacher network model and the preset feature recognition results output by the student network model, the differences between the preset feature recognition results output by the student network model and the preset feature label ground truth, and the differences between the aortic dissection recognition results output by the student network model and the aortic dissection label ground truth, the model parameters of the student network model are adjusted until the student network model converges, and the trained student network model is determined as the aortic dissection recognition model; wherein, the initial model parameters of the teacher network model and the student network model are the same.

2. The aortic dissection identification model construction method according to claim 1, characterized in that, The step of adjusting the model parameters of the student network model based on the differences between the preset feature recognition results output by the trained teacher network model and the preset feature recognition results output by the student network model, the differences between the preset feature recognition results output by the student network model and the preset feature label ground truth, and the differences between the aortic dissection recognition results output by the student network model and the aortic dissection label ground truth, until the student network model converges, includes: A first loss function is constructed based on the difference between the preset feature recognition results output by the trained teacher network model and the preset feature recognition results output by the student network model. A second loss function is constructed based on the difference between the preset feature recognition results output by the student network model and the preset feature label ground truth values; Based on the difference between the aortic dissection identification results output by the student network model and the true values ​​of the aortic dissection labels, an aortic dissection probability prediction loss function is constructed. Based on the first loss function, the second loss function, and the aortic dissection probability prediction loss function, the loss function of the student network model is constructed. The loss function of the student network model is optimized using gradient descent until the loss function of the student network model converges, at which point the training of the student network model ends.

3. The method of claim 2, wherein the method further comprises: The preset features include spatial features and temporal features. A first loss function is constructed based on the difference between the preset feature recognition results output by the trained teacher network model and the preset feature recognition results output by the student network model, including: Based on the first spatial feature recognition result output by the trained teacher network model, the second spatial feature recognition result output by the student network model, and the preset distillation temperature, a first spatial consistency loss function is constructed. Based on the first time feature recognition result output by the trained teacher network model, the second time feature recognition result output by the student network model, and the preset distillation temperature, a first time consistency loss function is constructed. The first loss function is constructed based on the first spatial consistency loss function and the first temporal consistency loss function.

4. The method of claim 3, wherein, A method for obtaining the first spatial feature recognition result and the second spatial feature recognition result includes: For each data sample in the echocardiogram dataset of the aortic dissection, the forward-order image sequence in the data sample is reversed in the time dimension to obtain the reverse-order image sequence. The reverse-order image sequence is input into the trained teacher network model to obtain the first teacher feature map, and the forward-order image sequence is input into the student network model to obtain the student feature map; The first teacher feature map is input into the spatial feature recognition branch of the trained teacher network model to obtain the first spatial feature recognition result, and the student feature map is input into the spatial feature recognition branch of the student network model to obtain the second spatial feature recognition result.

5. The method of claim 3, wherein the method further comprises: A method for obtaining the first time feature recognition result and the second time feature recognition result includes: For each data sample in the echocardiogram dataset of the aortic dissection, the region of interest in the aortic region is cropped from the forward sequence of images in the data sample, and the region of interest is interpolated to the input size of the trained teacher network model to obtain the aortic region image sequence. The aortic region image sequence is input into the trained teacher network model to obtain the second teacher feature map, and the forward sequence image sequence is input into the student network model to obtain the student feature map; The second teacher feature map is input into the time feature recognition branch of the trained teacher network model to obtain the first time feature recognition result, and the student feature map is input into the time feature recognition branch of the student network model to obtain the second time feature recognition result.

6. The method of claim 2, wherein, The preset features include spatial and temporal features. A second loss function is constructed based on the difference between the preset feature recognition results output by the student network model and the preset feature label ground truth values, including: Based on the difference between the spatial feature recognition results output by the student network model and the true values ​​of the spatial feature labels, a second spatial consistency loss function is constructed. Based on the difference between the time feature recognition results output by the student network model and the true values ​​of the time feature labels, a second time consistency loss function is constructed. The second loss function is constructed based on the second spatial consistency loss function and the second temporal consistency loss function.

7. The aortic dissection identification model construction method according to claim 1, characterized in that, The preset features include spatial and temporal features. The task of identifying the preset features of each data sample in the non-aortic dissection echocardiogram dataset, and training the teacher network model using the non-aortic dissection echocardiogram dataset until the teacher network model converges, includes: The loss function of the teacher network model is constructed based on the spatial consistency loss function and the temporal consistency loss function. The spatial consistency loss function represents the difference between the recognition result of the spatial feature by the teacher network model and the true value of the spatial feature label. The temporal consistency loss function represents the difference between the recognition result of the temporal feature by the teacher network model and the true value of the temporal feature label. During the training of the teacher network model using the non-aortic dissection echocardiogram dataset, the gradient descent method is used to optimize the loss function of the teacher network model until the loss function of the teacher network model converges, at which point the training of the teacher network model ends.

8. The method for constructing an aortic dissection identification model according to any one of claims 3 to 7, characterized in that, The spatial features are section categories, and the temporal features include end-contraction frames and end-diastolic frames.

9. An aortic dissection identification method characterized by, include: Acquire the image sequence of the echocardiogram to be processed; The image sequence of the echocardiogram to be processed is input into the aortic dissection recognition model constructed by the aortic dissection recognition model construction method according to any one of claims 1 to 8, and the aortic dissection recognition result output by the aortic dissection recognition model is obtained.

10. An aortic dissection identification model construction apparatus characterized by comprising: include: The dataset acquisition unit is used to acquire the echocardiogram datasets of non-aortic dissection and aortic dissection. Each data sample in the echocardiogram dataset of non-aortic dissection is labeled with a preset feature label, and each data sample in the echocardiogram dataset of aortic dissection is labeled with a preset feature label and an aortic dissection label. The teacher network model training unit is used to train the teacher network model using the non-aortic dissection echocardiogram dataset with the task of identifying preset features of each data sample in the non-aortic dissection echocardiogram dataset until the teacher network model converges. The identification result acquisition unit is used to input data samples from the echocardiogram dataset of the aortic dissection into the trained teacher network model and student network model respectively, and obtain the preset feature identification result output by the trained teacher network model and the preset feature identification result and aortic dissection identification result output by the student network model. The model parameter adjustment unit is used to adjust the model parameters of the student network model based on the differences between the preset feature recognition results output by the trained teacher network model and the preset feature recognition results output by the student network model, the differences between the preset feature recognition results output by the student network model and the preset feature label ground truth, and the differences between the aortic dissection recognition results output by the student network model and the aortic dissection label ground truth, until the student network model converges, and the trained student network model is determined as the aortic dissection recognition model; wherein, the initial model parameters of the teacher network model and the student network model are the same.

11. An aortic dissection identification apparatus, characterized by, include: The image sequence acquisition unit is used to acquire the image sequence of the echocardiogram to be processed; An image sequence processing unit is used to input the image sequence of the echocardiogram to be processed into the aortic dissection recognition model constructed by the aortic dissection recognition model construction method according to any one of claims 1 to 8, and obtain the aortic dissection recognition result output by the aortic dissection recognition model.

Images