Peripheral neuroblastoma classification method, target classification model training method, device, storage medium, and program product

Abstract

The embodiment of the application provides a peripheral neuroblastoma classification method, a target classification model training method, equipment, a storage medium and a program product, which are applied to the field of medical image processing, the target classification model is trained through a sample pair composed of three types of images of a PET image, a CT image, and a fusion image of the PET image and the CT image, the accuracy of the target classification model is improved, and then the accuracy of the target classification model for predicting the type of peripheral neuroblastoma is improved. Moreover, by combining the judgment result of whether the training condition is met in the training process, the labeled data is updated when the training condition is not met, the labeling efficiency and rationality are improved, and then the model training efficiency is improved.

Description

Technical Field

[0001] This application relates to the field of medical image processing, and in particular to a method for classifying peripheral neuroblastoma tumors, a method for training a target classification model, an apparatus, a storage medium, and a program product. Background Technology

[0002] Peripheral neuroblastic tumors (pTNs) are clinically and biologically heterogeneous tumors, including different types such as neuroblastoma and ganglioneuroma, each corresponding to different treatment strategies.

[0003] Therefore, accurate determination of pTN type is crucial for patients to obtain early and appropriate treatment strategies, and it is also a technical problem that urgently needs to be solved. Summary of the Invention

[0004] This application provides a method for classifying peripheral neuroblastoma, a method for training and generating a target classification model, an apparatus, a storage medium, and a program product to improve the accuracy of peripheral neuroblastoma classification.

[0005] In a first aspect, this application provides a method for classifying peripheral neuroblastomas, including: Acquire a first medical image, a second medical image, and a third medical image of the target user; the first medical image is acquired using positron emission tomography (PET) technology, the second medical image is acquired using computed tomography (CT) technology, and the third medical image is obtained by fusing the first medical image and the second medical image; A target classification model is determined, which includes a feature extraction module, a feature fusion module, and a classification module. The feature extraction module includes a first feature extraction network, a second feature extraction network, and a third feature extraction network connected in parallel. The first image features of the first medical image are extracted using the first feature extraction network, the second image features of the second medical image are extracted using the second feature extraction network, and the third image features of the third medical image are extracted using the third feature extraction network. The feature fusion module is used to fuse the first image feature, the second image feature, and the third image feature to obtain fused image features; The classification module is used to classify the features of the fused image to obtain the classification result of the peripheral neuroblastoma of the target user; The target classification model is trained on a training dataset, which includes a labeled dataset and an unlabeled dataset. The labeled dataset includes multiple first sample pairs and corresponding category labels for peripheral neuroblastoma tumors. The unlabeled dataset includes multiple second sample pairs. The target model parameters of the target classification model are trained on the labeled dataset when the training conditions are met. At least one third sample pair among the multiple second sample pairs and corresponding category labels are used to update the labeled dataset when the training conditions are not met.

[0006] Secondly, this application provides a method for training a target classification model, including: A target classification model is constructed, comprising a feature extraction module, a feature fusion module, and a classification module; the feature extraction module includes a first feature extraction network, a second feature extraction network, and a third feature extraction network connected in parallel. Obtain a training dataset; the training dataset includes a labeled dataset and an unlabeled dataset. The labeled dataset includes multiple first sample pairs and corresponding category labels for peripheral neuroblastoma tumors. The unlabeled dataset includes multiple second sample pairs. Each sample pair includes a first sample image, a second sample image, and a third sample image. The first sample image is obtained from the sample user using positron emission tomography (PET) technology. The second sample image is obtained from the sample user using computed tomography (CT) technology. The third sample image is obtained by fusing the first sample image and the second sample image. The target classification model is trained using the labeled dataset to obtain the model parameters of the target classification model; If the training conditions are not met, at least one third sample pair is selected from the plurality of second sample pairs; Output annotation prompts and, in response to annotation requests, obtain the category labels corresponding to the at least one third sample pair respectively; Based on the at least one third sample pair and their corresponding category labels, update the labeled dataset and return to the step of training the target classification model using the labeled dataset until the training conditions are met; Under the condition that the training conditions are met, the target model parameters of the target classification model are determined.

[0007] Thirdly, this application provides a computing device, including a storage component and a processing component; the storage component stores one or more computer program instructions, which are invoked and executed by the processing component, and the processing component executes the one or more computer program instructions to implement the peripheral neuroblastoma classification method as described in the first aspect, or the target classification model training method as described in the second aspect.

[0008] Fourthly, this application provides a computer-readable storage medium storing a computer program that is executed by a computer to implement the peripheral neuroblastoma classification method as described in the first aspect, or the target classification model training method as described in the second aspect.

[0009] Fifthly, this application provides a computer program product storing a computer program that, when executed by a computer, implements the peripheral neuroblastoma classification method as described in the first aspect or the target classification model training method as described in the second aspect.

[0010] In this embodiment, the target classification model is trained using sample pairs composed of three types of images: PET images, CT images, and fused images of PET and CT images. Compared to training with a single type of sample, the scheme in this embodiment significantly improves the accuracy of the target classification model, thereby enhancing its accuracy in predicting the type of peripheral neuroblastoma. Furthermore, by incorporating the results of whether training conditions are met during training, the labeled data is updated when these conditions are not met. This avoids the problem of extensive pre-labeling, which leads to long processing times and high costs, thus reducing labeling time, lowering labeling costs, avoiding waste of labeling resources, improving labeling efficiency and rationality, and ultimately improving model training efficiency.

[0011] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0012] 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart of an embodiment of the target classification model training method provided in this application is shown; Figure 2 This diagram illustrates the structure of a target classification model in a practical application. Figure 3 A flowchart illustrating an embodiment of a method for classifying peripheral neuroblastomas provided in this application is shown; Figure 4 This illustration shows a schematic diagram of an embodiment of a target classification model training device provided in this application; Figure 5 This invention provides a schematic diagram of the structure of one embodiment of a peripheral neuroblastoma classification device. Figure 6 A schematic diagram of one embodiment of a computing device provided in this application is shown. Detailed Implementation

[0014] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0015] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a chronological order, nor do they limit "first" and "second" to different types.

[0016] The technical solutions of this application are applicable to the field of image processing, especially the field of medical image processing. Following the background information, currently, the determination of pNT types is achieved through invasive methods such as biopsy or pathological sections, resulting in a poor user experience.

[0017] To address the aforementioned technical problems, the inventors have proposed a technical solution for this application, comprising: acquiring a first medical image, a second medical image, and a third medical image of a target user; the first medical image is acquired using positron emission tomography (PET), the second medical image is acquired using computed tomography (CT) scanning, and the third medical image is obtained by fusing the first and second medical images; determining a target classification model, the target classification model including a feature extraction module, a feature fusion module, and a classification module, the feature extraction module including a first feature extraction network, a second feature extraction network, and a third feature extraction network connected in parallel; extracting first image features from the first medical image using the first feature extraction network, extracting second image features from the second medical image using the second feature extraction network, and extracting third image features from the third medical image using the third feature extraction network. Image features; the first image features, the second image features, and the third image features are fused using the feature fusion module to obtain fused image features; the fused image features are classified using the classification module to obtain the classification result of the peripheral neuroblastoma of the target user; wherein, the target classification model is trained based on a training dataset, the training dataset includes a labeled dataset and an unlabeled dataset, the labeled dataset includes multiple first sample pairs and corresponding category labels for peripheral neuroblastoma, the unlabeled dataset includes multiple second sample pairs, the target model parameters of the target classification model are trained based on the labeled dataset under the condition of meeting the training conditions, and at least one third sample pair among the multiple second sample pairs and corresponding category labels are used to update the labeled dataset when the training conditions are not met.

[0018] By training the target classification model using sample pairs composed of PET images, CT images, and fused PET and CT images, compared to training with single-type samples, the scheme in this application significantly improves the accuracy of the target classification model, thereby enhancing its accuracy in predicting the type of peripheral neuroblastoma. Furthermore, by incorporating the results of whether training conditions are met during training, the labeled data is updated when these conditions are not met, avoiding the problems of extensive pre-labeling processing that leads to long processing times and high costs. This reduces labeling time and costs, avoids wasting labeling resources, improves labeling efficiency and rationality, and ultimately enhances model training efficiency.

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] The technical solutions of this application can be applied to system architectures that include user terminals and servers, with the user terminal and server establishing a connection through a network. The network provides a medium for the communication link between the user terminal and the server. The network can include various connection types, such as wired or wireless communication links or fiber optic cables, etc.

[0021] The user client can interact with the server via the network to send relevant medical images or receive classification results, etc.

[0022] The user end can be a browser, an app (application), a web application such as an H5 (HyperText Markup Language 5) application, a mini-program (also known as a lightweight application), or a cloud application. The user end can be deployed on electronic devices and depends on the device or certain apps on the device to run. Electronic devices can have displays and support information browsing, such as personal mobile terminals like smartphones, tablets, and personal computers. For ease of understanding, various other types of applications can also be configured on electronic devices, such as human-computer interaction applications, model training applications, text processing applications, web browser applications, shopping applications, search applications, instant messaging tools, email clients, and social media platform software.

[0023] The server side can include servers that provide various services, such as instant messaging servers, servers used for background training that support target classification models, etc.

[0024] It should be noted that the server can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. The server can also be a server in a distributed system, or a server combined with blockchain. The server can also be a cloud server, or an intelligent cloud computing server or intelligent cloud host with artificial intelligence technology.

[0025] It should be noted that the embodiments of this application may involve the use of user data. In practical applications, user-specific personal data may be used in the scheme described herein within the scope permitted by applicable laws and regulations, provided that it complies with the applicable laws and regulations of the country (e.g., with the user's explicit consent, with the user being properly notified, etc.).

[0026] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0027] The technical solution of this application embodiment is implemented through a target classification model. The training process of the target classification model will be described below. Figure 1 The flowchart illustrates an embodiment of a target classification model training method provided in this application. The method may include the following steps.

[0028] 101: Construct a target classification model that includes a feature extraction module, a feature fusion module, and a classification module.

[0029] In this embodiment, the feature extraction module may include a first feature extraction network, a second feature extraction network, and a third feature extraction network connected in parallel. The three feature extraction networks can perform feature extraction processing on input images of different categories respectively, which can also be understood as a three-input target classification model.

[0030] Figure 2 This is a schematic diagram of the structure of a target classification model in a practical application. For example... Figure 2 As shown, the target classification model may include a feature extraction module 21, a feature fusion module 22, and a classification module 23. The feature extraction module 21 may include a first feature extraction network 211, a second feature extraction network 212, and a third feature extraction network 213 connected in parallel. The first feature extraction network 211 can be used to extract features from a first medical image, the second feature extraction network 212 can be used to extract features from a second medical image, and the third feature extraction network 213 can be used to extract features from a third medical image. The feature fusion module 22 can be used to fuse the image features extracted by the three feature extraction networks, and the classification module 23 can be used to classify the fused image features output by the feature fusion module 22 to obtain a classification result.

[0031] The three feature extraction networks can have the same structure. In a practical application, all three feature extraction networks can be implemented as 3DResNet networks (deep convolutional neural networks trained on the ImageNet model (a large-scale visual database)). Each feature extraction network can include an initial convolutional layer, a batch normalization layer, a ReLU (Rectified Linear Unit) activation function layer, a max pooling layer, and four sets of residual blocks connected in sequence. The specific structure can be referred to the traditional 3DResNet network, and will not be elaborated further.

[0032] The feature fusion module can include three pooling layers and one concatenation layer. The three pooling layers can be connected to the last set of residual blocks of each of the three feature extraction networks, and the concatenation layer can be connected to the three pooling layers. In a practical application, the pooling layers can be adaptive average pooling layers. An adaptive average pooling layer can be connected to the last set of residual blocks of a feature extraction network to perform 1×1×1 pooling on the feature map output by that network, thus unifying the dimensions of the feature maps output by each network. The concatenation layer can be connected to all three adaptive average pooling layers to concatenate the pooled features along the channel dimension, resulting in dimensional fusion features.

[0033] The classification module can include a flattening layer, three fully connected layers, two batch normalization layers, a ReLU activation function layer, and two Dropout regularization layers. The flattening layer can be connected to the concatenation layer in the feature fusion module to convert the dimensionality fusion features output by the concatenation layer into a one-dimensional vector. The first set of fully connected layers can be connected to the flattening layer, and the first set of batch normalization layers and Dropout regularization layers can be placed after the first set of fully connected layers. The second set of fully connected layers can be connected to the first set of Dropout regularization layers, and the second set of batch normalization layers and Dropout regularization layers can be placed after the second set of batch normalization layers. The third set of fully connected layers can be connected to the second set of Dropout regularization layers, and the ReLU activation function layer can be connected to the third set of fully connected layers to normalize the output of the third set of fully connected layers, obtaining the probabilities for each class. In a practical application, the momentum parameter of each batch normalization layer... It can be 0.9, the numerical stability parameter. It can be 1e-05. The dropout probability parameter p for each Dropout regularization layer can be 0.2. The ReLU activation function layer can use the softmax function (normalized exponential function).

[0034] 102: Obtain the training dataset.

[0035] In this embodiment of the application, the training dataset may include a labeled dataset (X). l Y l ) and unlabeled dataset X u The labeled dataset can include multiple first sample pairs X l And the corresponding category label Y for peripheral neuroblastoma. l Unlabeled datasets may include multiple second sample pairs X u .

[0036] In a practical application, peripheral neuroblastomas can be categorized into two types: neuroblastoma and ganglioneuroma. Based on this, the category label for peripheral neuroblastomas can include a first category and a second category; for example, if the first category is neuroblastoma, the second category is ganglioneuroma; conversely, if the first category is ganglioneuroma, the second category is neuroblastoma.

[0037] Sample pairs can be segmented based on sample users. Sample users can refer to sample patients, and sample pairs can be obtained from the medical record data generated during the diagnosis and treatment of sample patients or from the medical record cover page data recorded in the medical record cover page.

[0038] Any sample pair may include a first sample image, a second sample image, and a third sample image. The first sample image may be acquired from the sample user using positron emission tomography (PET), and the second sample image may be acquired from the sample user using computed tomography (CT). In a practical application, the first and second sample images may be obtained by scanning the sample user using a positron emission tomography system (PET / CT system, combined with fluorine-18fluorodeoxyglucose (18F-FDG) tracer).

[0039] The third sample image can be obtained by fusing the first sample image (PET sample image) and the second sample image (CT sample image). Specifically, the first sample image and the second sample image can be fused according to weights to generate a fused sample image, so as to balance structural and functional information.

[0040] To improve the accuracy of the third sample image, optionally, before fusion, the first sample image and the second sample image can be aligned in terms of anatomical structure by spatial transformation, and then the aligned first sample image and the second sample image can be fused according to weights.

[0041] To improve model training performance, image preprocessing can optionally be performed on each sample image before training. Image preprocessing may include size normalization, image enhancement, etc. Size normalization can adjust the image size of each sample image to a uniform size, while image enhancement can randomly rotate, randomly flip horizontally, or randomly flip vertically of each sample image to enrich the number of training samples and improve the model's generalization ability.

[0042] 103: Train the target classification model using the labeled dataset to obtain the model parameters of the target classification model.

[0043] This can be achieved by using multiple first sample pairs from the labeled dataset as training samples and the category label of peripheral neuroblastoma as the training label to train the target classification model. Specifically, the first sample pairs can be input into the feature extraction module to extract the first sample features of the first sample image using a first feature extraction network, the second sample features of the second sample image using a second feature extraction network, and the third sample features of the third sample image using a third feature extraction network. The feature fusion module then fuses the first, second, and third sample features to obtain fused sample features. The classification module then classifies the fused sample features to obtain the predicted category. The predicted category is compared with the category label corresponding to the first sample pair, and the target classification model is trained according to a preset loss function to obtain the model parameters of the target classification model.

[0044] In a practical application, the preset loss function can be the cross-entropy loss function, and the target classification model can be trained by minimizing the cross-entropy loss function.

[0045] The formula for the cross-entropy loss function can be: ; where n l This indicates the number of the first sample pairs, and M represents the number of categories. In this embodiment, M=2. Indicates an indicator function, This represents the predicted probability of the target classification model that the i-th sample pair belongs to the j-th class. This represents the model parameters of the target classification model.

[0046] 104: Determine if the training conditions are met. If the result is no, proceed to step 105; if the result is yes, proceed to step 108.

[0047] The training conditions may include one or more of the following: The ratio of the number of first sample pairs in the labeled dataset reaches the preset ratio; The evaluation metrics of the target classification model on the test dataset have reached the preset metric values.

[0048] Specifically, the preset ratio can be set according to actual needs, such as 80%, 90%, etc.

[0049] The test dataset may include multiple test sample pairs and their corresponding category labels for peripheral neuroblastoma. Each test sample pair may include a first test image, a second test image, and a third test image. The first test image is acquired from the test user using positron emission tomography (PET), the second test image is acquired from the test user using computed tomography (CT), and the third test image is obtained by fusing the first and second test images. The test dataset and the training dataset are independent of each other. The acquisition of the test dataset is consistent with the acquisition of the labeled dataset in the training dataset, and will not be described further.

[0050] The test sample pairs are input into the target classification model to predict the predicted category corresponding to each test sample pair. Based on the predicted category and the category label (i.e., the true category), the evaluation index value is calculated. When the calculated index value reaches the preset index value, the training condition is determined to be met.

[0051] In a practical application, evaluation metrics may include the area under the receiver operating characteristic curve (AUROC). When the AUROC value of the target classification model on the test dataset reaches a preset area value, the training conditions are considered met.

[0052] Of course, other evaluation indicators can be set according to actual needs, such as accuracy, precision, recall, etc., and this application does not impose any restrictions on this.

[0053] 105: Select at least one third sample pair from multiple second sample pairs.

[0054] 106: Output annotation prompts and, in response to annotation requests, obtain the category labels corresponding to at least one third sample pair.

[0055] When training conditions are not met, at least one third sample pair can be selected from the unlabeled dataset as new labeled data. This selection can be achieved using edge sampling, minimum confidence, or other sample selection methods, which will be described in subsequent embodiments.

[0056] Based on at least one selected third sample pair, annotation prompts can be output to guide the user in labeling the category of at least one third sample pair. For example, the annotation prompt could be: "Please label the category of peripheral neuroblastoma corresponding to each of the at least one third sample pair."

[0057] In response to a labeling request, at least one third sample pair provided by the user can be retrieved, each corresponding to a category label.

[0058] 107: Based on at least one third sample pair and their corresponding class labels, update the labeled dataset and return to step 103 to continue execution.

[0059] Based on the category labels corresponding to at least one third sample pair provided by the user, at least one third sample pair and its corresponding category labels can be added to the labeled dataset to update the labeled dataset.

[0060] For example, selecting n from multiple second sample pairs s A third sample pair is obtained, at which point the newly labeled data can be acquired. , This represents the i-th third sample pair. Let represent the class label of the i-th third sample pair. The updated labeled dataset can be... n s It can be set according to actual needs, such as 10, 20, etc.

[0061] Furthermore, based on at least one third sample pair, the unlabeled dataset can also be updated by removing the at least one third sample pair from the original unlabeled dataset to obtain the updated unlabeled dataset.

[0062] 108: Obtain the target model parameters of the target classification model.

[0063] When the training conditions are met, the model parameters obtained from the most recent training can be used as the target model parameters to end the training of the target classification model.

[0064] The trained target classification model can be used to output the classification result of peripheral neuroblastoma for a target user based on the first medical image (PET image), the second medical image (CT image), and the third medical image (a fused image of the PET and CT images). Specifically, the first feature extraction network of the target classification model can be used to extract the first image features of the first medical image, the second feature extraction network can be used to extract the second image features of the second medical image, and the third feature extraction network can be used to extract the third image features of the third medical image. The feature fusion module of the target classification model can be used to fuse the first, second, and third image features to obtain fused image features. The classification module of the target classification model can be used to classify the fused image features to obtain the classification result of peripheral neuroblastoma for the target user.

[0065] In this embodiment, the target classification model is trained using sample pairs composed of three types of images: PET images, CT images, and fused images of PET and CT images. Compared to training with a single type of sample, the scheme of this embodiment significantly improves the accuracy of the target classification model, thereby enhancing its accuracy in predicting the type of peripheral neuroblastoma. Furthermore, by incorporating the results of whether training conditions are met during training, the labeled data is updated when these conditions are not met. This avoids the problem of extensive pre-labeling, which leads to long processing times and high costs, thus reducing labeling time, lowering labeling costs, avoiding waste of labeling resources, improving labeling efficiency and rationality, and ultimately improving model training efficiency.

[0066] To further improve model training efficiency, the three feature extraction networks in the feature extraction module can be extracted from pre-trained related models. Based on this, in some embodiments, constructing a target classification model including a feature extraction module, a feature fusion module, and a classification module may include: The system extracts a first feature extraction network from a first classification model, a second feature extraction network from a second classification model, and a third feature extraction network from a third classification model. The first classification model includes a first feature extraction network and a first classification network, which are pre-trained based on a first training sample and its corresponding training label. The second classification model includes a second feature extraction network and a second classification network, which are pre-trained based on a second training sample and its corresponding training label. The third classification model includes a third feature extraction network and a third classification network, which are pre-trained based on a third training sample and its corresponding training label. The first feature extraction network, the second feature extraction network, and the third feature extraction network are connected in parallel to construct a feature extraction module; Construct a feature fusion module; the feature fusion module may include pooling layers connected to the first feature extraction network, the second feature extraction network and the third feature extraction network respectively, and a splicing layer connected to the three pooling layers respectively; A classification module is constructed. The classification module may include a fully connected layer. The weight parameters of the fully connected layer are obtained by fusing the first weight parameters of the fully connected layer in the first classification network, the second weight parameters of the fully connected layer in the second classification network, and the third weight parameters of the fully connected layer in the third classification network.

[0067] In this embodiment, the first classification model, the second classification model, and the third classification model are all single-input target classification models, and have been pre-trained based on training samples and corresponding training labels. The first classification model, the second classification model, and the third classification model can be the same or different. Each classification model includes a feature extraction network and a classification network.

[0068] It should be noted that any classification model may also include other structures such as input networks and attention mechanism modules, and this application does not impose any restrictions on this.

[0069] Specifically, extracting the first feature extraction network from the first classification model can involve extracting at least one network parameter corresponding to the first feature extraction network from the first classification model, and loading this at least one network parameter into the first feature extraction network in the feature extraction module to achieve parameter sharing. In a practical application, the at least one network parameter may include the initial convolutional layer, the batch normalization layer, and the weight parameters corresponding to the four sets of residual blocks.

[0070] Similarly, the extraction of the second and third feature extraction networks is implemented in the same way as described above, and will not be repeated here.

[0071] The specific structures of the feature fusion module and the classification module have been described in the aforementioned embodiments and will not be repeated here.

[0072] Considering that the network parameters of the feature extraction network extracted from the single-input classification model may not match those of the three-input classification model, in this embodiment, the weight parameters of the fully connected layer in the classification module can be obtained by fusing the first weight parameters of the first classification network, the second weight parameters of the second classification network, and the third weight parameters of the third classification network.

[0073] In the case where the classification module includes three sets of fully connected layers, the weight dimension of the first set of fully connected layers can be adjusted to adapt and fuse the feature dimension. That is, the weight parameters of the first fully connected layer can be obtained by fusing the first weight parameters of the first classification network, the second weight parameters of the second classification network, and the third weight parameters of the third classification network.

[0074] Taking the three classification models mentioned above as an example, a weight parameter copying method can be used. The corresponding weight parameters in one classification network are copied three times along the input channel to obtain the weight parameters of the first fully connected layer in the classification module. The weight parameters of other layers in the classification module can be extracted from the classification network and directly loaded.

[0075] By extracting the feature extraction network from the trained single-input classification model and adjusting the weight dimension of the first fully connected layer in the classification module to adapt to the fused feature dimension, the model training parameters are reduced, and the problem that the network parameters of the feature extraction network extracted from the single-input classification model may not match the three-input classification model is avoided, which further improves the model training efficiency and increases the model convergence speed.

[0076] To improve the rationality of selecting the third sample pair, in some embodiments, selecting at least one third sample pair from a plurality of second sample pairs may include: The target classification model is used to predict multiple second sample pairs respectively, and the predicted class probabilities corresponding to the multiple second sample pairs are obtained. Based on the predicted class probabilities corresponding to the multiple second sample pairs respectively, the first sample selection strategy is generated using the first policy network. Based on the predicted probabilities of the categories corresponding to multiple second sample pairs and the first sample selection strategy, the sample selection priority corresponding to the multiple second sample pairs is determined. Select at least one second sample pair with higher priority as at least one third sample pair.

[0077] The above methods may also include: Based on the predicted probabilities of the categories corresponding to multiple second sample pairs and the first sample selection strategy, the first value parameter is generated using the first value network. If the first value parameter does not meet the value condition, after the labeled dataset and the unlabeled dataset are updated, the updated target classification model is used to predict the updated second sample pairs respectively, and the predicted class probabilities corresponding to the updated second sample pairs are obtained. Based on the predicted class probabilities corresponding to the updated second sample pairs respectively, the second policy network is used to generate the second sample selection policy. Based on the updated predicted class probabilities of multiple second sample pairs and the second sample selection strategy, the second value parameters are generated using the second value network. Based on the difference information between the second value parameter and the first value parameter, the first value network and the first policy network are updated, and the process returns to the step of generating the first sample selection policy using the first policy network to continue execution.

[0078] In this embodiment, a policy-value network can be used to determine a suitable sample selection strategy to guide the selection of the third sample pair. The policy-value network (Actor-Critic network) can include a policy network and a value network. The policy network can consist of three fully connected layers, and the value network can also consist of three fully connected layers.

[0079] Specifically, the target classification model obtained from the most recent training can be used to predict multiple second sample pairs separately, thereby obtaining the predicted probability of the category corresponding to each second sample pair. , This indicates the unlabeled dataset X. u The i-th second sample pair in the middle, This represents the predicted category corresponding to the i-th second sample pair.

[0080] Construct a first probability matrix that includes the predicted probabilities of multiple second samples for their respective categories. n u This indicates the unlabeled dataset X. u The number of second sample pairs, and the elements of the first probability matrix S are the predicted class probabilities S. ij .

[0081] Input the first probability matrix S into the first policy network to obtain the first sample selection policy. , This represents the network parameters of the first strategy network.

[0082] Based on the predicted class probabilities corresponding to multiple second sample pairs and the first sample selection strategy, the sample selection priority corresponding to each of the multiple second sample pairs is determined. Specifically, this can be achieved by using a first policy network, based on the first probability matrix S and the first sample selection strategy. Generate action matrix The elements in action matrix 'a' represent the selection priorities of multiple second sample pairs. All elements of action matrix 'a' are values ​​between 0 and 1; a larger value indicates a higher priority for selection as a third sample pair. In a practical application, the values ​​in action matrix 'a' can be sorted in descending order, and the top n elements can be selected... s The second sample pair corresponding to each value is used as the third sample pair.

[0083] To evaluate the first sample selection strategy generated by the first policy network, the first probability matrix S and the first sample selection strategy can be used. Input the first value network to obtain the first value parameters. , This represents the network parameters of the first value network.

[0084] Determine whether the first value parameter Q meets the value condition. The value condition can be, for example, that the value of the first value parameter Q reaches a preset value, which can be set according to actual needs.

[0085] When the first value parameter Q does not meet the value condition, the first value network and the first policy network can be updated, and the step of generating the first sample selection policy using the first policy network can be returned to continue execution to improve the accuracy of the first sample selection policy; when the first value parameter Q meets the value condition, the current first sample selection policy can be used as the final sample selection policy.

[0086] The following section explains the update process of the first value network and the first policy network.

[0087] In this embodiment, a second policy network and a second value network can be used. The structure and principle of the second policy network are the same as those of the first policy network, and the structure and principle of the second value network are the same as those of the first value network, so they will not be described again.

[0088] After selecting at least one third sample pair according to the current first sample selection strategy, and updating the labeled and unlabeled datasets using at least one third sample pair and their corresponding class labels, the updated target classification model can be used to predict the updated second sample pairs respectively, obtain the class prediction probabilities corresponding to the updated second sample pairs respectively, and construct the corresponding updated second probability matrix S′.

[0089] The second probability matrix is ​​input into the second policy network to obtain the second sample selection policy. , This represents the network parameters of the second policy network, including the second probability matrix S′ and the second sample selection policy. Input the second value network to obtain the second value parameters. , This represents the network parameters of the second value network.

[0090] Based on the difference between the second value parameter and the first value parameter, the first value network and the first policy network can be updated.

[0091] In a practical application, the network parameters of the first value network can be updated by minimizing the expected loss based on the mean square error between the second value parameter and the first value parameter, and the network parameters of the first policy network can be updated by maximizing the expected return.

[0092] Of course, there are other ways to update, which will be described in subsequent embodiments.

[0093] By employing a policy-value network to generate sample selection strategies and dynamically updating these strategies, experience can be accumulated during continuous decision-making, and sample selection preferences can be dynamically adjusted to avoid getting trapped in local optima. Compared with static and fixed sample selection methods, this greatly improves the accuracy and rationality of sample selection strategies, enhances the accuracy and rationality of third sample pairs, and consequently improves the accuracy of subsequent model training.

[0094] To improve the accuracy of updating the first policy network and the first value network, in some embodiments, the above method may further include: The updated target classification model is used to predict at least one third sample pair, and the predicted category corresponding to at least one third sample pair is obtained. The reward parameters are determined based on the predicted category and category label corresponding to at least one third sample pair.

[0095] Based on this, updating the first value network and the first policy network based on the difference information between the second value parameter and the first value parameter can include: The target value parameter is obtained by weighting the second value parameter and the reward parameter. The first value network is updated based on the difference between the target value parameters and the predicted value parameters. The first policy network is updated based on the target value parameters and the updated first value network.

[0096] In this embodiment, the reward parameter can provide feedback on the prediction performance of the updated target classification model, and the reward parameter can be determined according to the following reward parameter generation formula.

[0097] The formula for generating reward parameters can be: Where r represents the reward parameter, n s Indicates the number of third sample pairs. k represents the predicted class of the i-th third sample pair by the updated target classification model. i This represents the category label of the i-th third sample pair, i.e., the true category.

[0098] Understandably, the reward parameter, as the current short-term gain, can be weighted with the second value parameter, which represents the long-term gain, to obtain the target value parameter. The target value parameter can then be used to update the first policy network and the first value network to improve training stability.

[0099] The formula for generating the target value parameter can be: ;in, Indicates the target value parameter. Indicates the weighting coefficient. , Let r represent the second value parameter, and r represent the reward parameter.

[0100] Based on the difference between the target value parameter and the first value parameter, the first value network can be updated first, and then the first policy network can be updated based on the target value parameter and the updated first value network.

[0101] By combining the second value parameter with the reward parameter, the long-term impact of implicitly learned samples on the model's generalization ability is realized, the marginal benefit of each newly added labeled data is improved, and better classification performance is achieved under a limited labeling budget.

[0102] In a practical application, updating the first value network based on the difference between the target value parameter and the first value parameter can include: Calculate the first expectation of the squared error between the target value parameter and the first value parameter, calculate the gradient of the first expectation with respect to the network parameters of the first value network, and update the network parameters of the first value network along the gradient descent direction; Updating the first policy network based on the updated first value network can include: Calculate the second expectation of the first value parameters corresponding to the updated first value network, and calculate the gradient of the second expectation with respect to the network parameters of the first policy network. Update the network parameters of the first policy network along the gradient ascending direction.

[0103] The first expectation can be: ;in, Represents the expectation function, Indicates the target value parameter. Indicates the first value parameter. This represents the error between the target value parameter and the first value parameter.

[0104] The network parameters of the first value network can be updated based on minimizing the expected loss.

[0105] Specifically, the gradient of the first expectation with respect to the network parameters of the first value network can be: ;in, Indicates the first value parameter Network parameters of the first value network The gradient.

[0106] The update formula for the network parameters of the first value network can be: ;in, This represents the first learning rate. In a practical application, the first learning rate can be 1e-3.

[0107] The second expectation can be: ;in, This represents the expectation function, where the network parameters of the first value network are... These are the network parameters after the above updates.

[0108] The network parameters of the first-policy network can be updated based on maximizing the expected return.

[0109] Specifically, the gradient of the second expectation with respect to the network parameters of the first policy network can be: ;in, This represents the gradient of the first value parameter Q with respect to the action matrix a. Indicates the first sample selection strategy Network parameters of the first strategy network The gradient.

[0110] The update formula for the network parameters of the first-policy network can be: ;;in, This represents the second learning rate. In a practical application, the second learning rate can be 1e-3.

[0111] Of course, in addition to the methods described above for updating the first value network based on minimizing expected loss and updating the first policy network based on maximizing expected return, other methods such as Temporal Difference (TD) error, Monte Carlo methods, and direct estimation of the Q function can also be used to update the first value network. This application does not impose any restrictions on these methods.

[0112] Optionally, the second value network and the second policy network can also be updated.

[0113] In a practical application, the update formula for the network parameters of the second value network can be: ;in, This represents the network parameters of the second value network. This represents the trade-off parameter.

[0114] The update formula for the network parameters of the second strategy network can be: ;in, The value parameters of the second policy network are represented. This represents the trade-off parameter.

[0115] like Figure 3 The diagram shown is a flowchart of an embodiment of a method for classifying peripheral neuroblastomas provided in this application. The method may include the following steps.

[0116] 301: Obtain the first, second, and third medical images of the target user.

[0117] The first medical image was acquired using positron emission tomography (PET), the second medical image was acquired using computed tomography (CT), and the third medical image was obtained by fusing the first and second medical images.

[0118] 302: Determine the target classification model. The target classification model includes a feature extraction module, a feature fusion module, and a classification module. The feature extraction module includes a first feature extraction network, a second feature extraction network, and a third feature extraction network connected in parallel.

[0119] The structure of the target classification model can be referenced. Figure 2 The diagram shown is shown in the image.

[0120] The target classification model is trained on a training dataset, which includes a labeled dataset and an unlabeled dataset. The labeled dataset includes multiple first sample pairs and their corresponding class labels for peripheral neuroblastoma, while the unlabeled dataset includes multiple second sample pairs. Specifically, the target model parameters can be obtained by training on the labeled dataset under certain training conditions. At least one third sample pair from the multiple second sample pairs, along with its corresponding class label, is used to update the labeled dataset when the training conditions are not met. The specific training process is detailed below. Figure 1 The examples shown already have corresponding descriptions, so they will not be repeated here.

[0121] 303: Extract the first image features of the first medical image using a first feature extraction network, extract the second image features of the second medical image using a second feature extraction network, and extract the third image features of the third medical image using a third feature extraction network.

[0122] 304: The feature fusion module is used to fuse the first image features, the second image features, and the third image features to obtain fused image features.

[0123] 305: The classification module is used to classify the features of the fused image to obtain the classification results of peripheral neuroblastoma of the target user.

[0124] In this embodiment, the target classification model is trained using sample pairs composed of three types of images: PET images, CT images, and fused images of PET and CT images. Compared to training with a single type of sample, the scheme in this embodiment significantly improves the accuracy of the target classification model, thereby improving the accuracy of predicting the type of peripheral neuroblastoma. Furthermore, by incorporating the judgment results of whether the training conditions are met during training, the labeled data is updated when the training conditions are not met, avoiding the problem of extensive pre-labeling processing, which leads to long processing times and high costs. This reduces labeling time and costs, avoids wasting labeling resources, improves labeling efficiency and rationality, and ultimately improves model training efficiency.

[0125] like Figure 4 The diagram shown is a structural schematic of an embodiment of a target classification model training device provided in this application. The device may include the following units.

[0126] The construction unit 401 is used to construct a target classification model including a feature extraction module, a feature fusion module, and a classification module; the feature extraction module includes a first feature extraction network, a second feature extraction network, and a third feature extraction network connected in parallel. The first acquisition unit 402 is used to acquire a training dataset. The training dataset includes a labeled dataset and an unlabeled dataset. The labeled dataset includes multiple first sample pairs and corresponding category labels for peripheral neuroblastoma tumors. The unlabeled dataset includes multiple second sample pairs. Each sample pair includes a first sample image, a second sample image, and a third sample image. The first sample image is acquired from the sample user using positron emission tomography (PET) technology. The second sample image is acquired from the sample user using computed tomography (CT) technology. The third sample image is obtained by fusing the first sample image and the second sample image. Training unit 403 is used to train the target classification model using the labeled dataset to obtain the model parameters of the target classification model; Decision unit 404 is used to determine whether the training conditions are met; Selection unit 405 is used to select at least one third sample pair from multiple second sample pairs if the result of determination unit 404 is negative. The second acquisition unit 406 is used to output annotation prompt information and, in response to the annotation request, acquire at least one third sample pair corresponding to the category label respectively; The first update unit 407 is used to update the labeled dataset based on at least one third sample pair and their corresponding class labels; The first determining unit 408 is used to determine the target model parameters of the target classification model if the result of the determining unit 404 is negative.

[0127] In some embodiments, the selection unit 405 is configured to use a target classification model to predict multiple second sample pairs respectively, obtain the category prediction probabilities corresponding to the multiple second sample pairs respectively, and generate a first sample selection strategy using a first policy network based on the category prediction probabilities corresponding to the multiple second sample pairs respectively; determine the sample selection priority corresponding to the multiple second sample pairs respectively based on the category prediction probabilities corresponding to the multiple second sample pairs respectively and the first sample selection strategy; and select at least one second sample pair with a high sample selection priority as at least one third sample pair. The above-mentioned device may further include: The first generation unit is used to generate first value parameters using the first value network based on the predicted probabilities of the categories corresponding to multiple second sample pairs and the first sample selection strategy. The second generation unit, when the first value parameter does not meet the value condition, is used to predict the updated multiple second sample pairs using the updated target classification model, obtain the class prediction probabilities corresponding to the updated multiple second sample pairs respectively, and generate a second sample selection strategy using the second policy network based on the class prediction probabilities corresponding to the updated multiple second sample pairs respectively. The third generation unit is used to predict the probability of the category corresponding to the updated multiple second sample pairs and the second sample selection strategy, and to generate the second value parameters using the second value network. The second update unit is used to update the first value network and the first policy network based on the difference information between the second value parameter and the first value parameter.

[0128] In some embodiments, the above-described apparatus may further include: The prediction unit is used to predict at least one third sample pair using the updated target classification model, and to obtain the predicted category corresponding to at least one third sample pair. The second determining unit is used to determine the reward parameters based on the predicted category and category label corresponding to at least one third sample pair respectively; The second update unit can be used to perform a weighted calculation of the second value parameter and the reward parameter to obtain the target value parameter; update the first value network based on the difference information between the target value parameter and the first value parameter; and update the first policy network based on the updated first value network.

[0129] In some embodiments, the second updating unit may be specifically used to calculate a first expectation of the squared error between the target value parameter and the first value parameter; calculate the gradient of the first expectation with respect to the network parameters of the first value network, and update the network parameters of the first value network along the gradient descent direction; calculate a second expectation of the first value parameter corresponding to the updated first value network; calculate the gradient of the second expectation with respect to the network parameters of the first policy network, and update the network parameters of the first policy network along the gradient ascent direction.

[0130] In some embodiments, the construction unit 401 can be specifically used to extract a first feature extraction network from a first classification model, a second feature extraction network from a second classification model, and a third feature extraction network from a third classification model; the first classification model includes a first feature extraction network and a first classification network, which are pre-trained based on a first training sample and corresponding training labels; the second classification model includes a second feature extraction network and a second classification network, which are pre-trained based on a second training sample and corresponding training labels; the third classification model includes a third feature extraction network and a third classification network, which are pre-trained based on a third training sample and corresponding training labels. The first feature extraction network, the second feature extraction network, and the third feature extraction network are connected in parallel to construct a feature extraction module; A feature fusion module is constructed; the feature fusion module includes pooling layers connected to the first feature extraction network, the second feature extraction network and the third feature extraction network respectively, and a splicing layer connected to the three pooling layers respectively; A classification module is constructed. The classification module includes a fully connected layer. The weight parameters of the fully connected layer are obtained by fusing the first weight parameters of the first classification network, the second weight parameters of the second classification network, and the third weight parameters of the third classification network.

[0131] Figure 4 The target classification model training device shown can be used to perform Figure 1 The implementation principle and technical effects of the target classification model training method shown will not be elaborated further. The specific methods by which each unit of the target classification model training device in the above embodiments performs its operations have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0132] like Figure 5 The diagram shown is a structural schematic of an embodiment of a peripheral neuroblastoma classification device provided in this application. The device may include the following modules.

[0133] The third acquisition unit 501 is used to acquire a first medical image, a second medical image, and a third medical image of the target user; the first medical image is acquired based on positron emission tomography (PET) technology, the second medical image is acquired based on computed tomography (CT) technology, and the third medical image is acquired based on the fusion of the first medical image and the second medical image; The third determining unit 502 is used to determine the target classification model. The target classification model includes a feature extraction module, a feature fusion module, and a classification module. The feature extraction module includes a first feature extraction network, a second feature extraction network, and a third feature extraction network connected in parallel. The extraction unit 503 is used to extract first image features of a first medical image using a first feature extraction network, extract second image features of a second medical image using a second feature extraction network, and extract third image features of a third medical image using a third feature extraction network. The fusion unit 504 is used to fuse the first image features, the second image features, and the third image features using the feature fusion module to obtain fused image features; Classification unit 505 is used to classify the fused image features using the classification module to obtain the classification result of peripheral neuroblastoma of the target user. The target classification model is trained based on the training dataset, which includes a labeled dataset and an unlabeled dataset. The labeled dataset includes multiple first sample pairs and their corresponding category labels for peripheral neuroblastoma. The unlabeled dataset includes multiple second sample pairs. The target model parameters of the target classification model are trained based on the labeled dataset when the training conditions are met. At least one third sample pair among the multiple second sample pairs and their corresponding category labels are used to update the labeled dataset when the training conditions are not met.

[0134] Figure 6 The peripheral neuroblastoma classification device shown can be used to perform... Figure 3 The implementation principle and technical effects of the peripheral neuroblastoma classification method shown will not be elaborated further. The specific operation methods of each unit in the peripheral neuroblastoma classification device in the above embodiments have been described in detail in the embodiments related to this method, and will not be elaborated further here.

[0135] Figure 6 This is a schematic diagram of the structure of one embodiment of a computing device provided in this application. Figure 6 As shown, in practical applications, the computing device may include a storage component 601 and a processing component 602.

[0136] Storage component 601 is used to store computer programs and can be configured to store various other data to support operation on a computing device. Examples of this data include instructions for any application or method used to operate on the computing device, data structures, contact data, phone book data, messages, pictures, videos, etc.

[0137] Processing component 602, coupled to storage component 601, is used to execute computer programs in storage component 601 for implementing, etc. Figure 1 The target classification model training method shown, or Figure 3 The classification method for peripheral neuroblastoma is shown.

[0138] Furthermore, such as Figure 6 As shown, the computing device may also include other components such as a communication component 603, a display component 604, a power supply component 605, and an audio component 606. Figure 6 The diagram only shows some components and does not mean that the computing device includes only these components. Figure 6 The components shown. Additionally... Figure 6 The components within the dashed box are optional, not mandatory, and their specific requirements depend on the product form of the computing device. The computing device in this embodiment can be a desktop computer, laptop computer, smartphone, or IoT (Internet of Things) device, or a server-side device such as a conventional server, cloud server, or server array. If the computing device in this embodiment is implemented as a desktop computer, laptop computer, or smartphone, it may include... Figure 6 The components within the dashed box; if the computing device in this embodiment is implemented as a conventional server, cloud server, or server array, etc., then it may not include... Figure 6 The component within the dashed box.

[0139] The processing component described above includes one or more processors to execute computer instructions to complete all or part of the steps in the method described above. Alternatively, the processing component may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the method described above.

[0140] The aforementioned storage components can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0141] The aforementioned communication component is configured to facilitate wired or wireless communication between the device housing the communication component and other devices. The device housing the communication component can access wireless networks based on communication standards, such as mobile communication networks, or combinations thereof. In one exemplary embodiment, the communication component receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel.

[0142] The aforementioned display components may include a screen, which may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen can be implemented as a touchscreen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors can sense not only the boundaries of touch or swipe actions but also the duration and pressure associated with the touch or swipe operation.

[0143] The aforementioned power supply components provide power to various components within the device in which they reside. These power supply components may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power to the device in which they reside.

[0144] The aforementioned audio component can be configured to output and / or input audio signals. For example, the audio component includes a microphone (MIC) configured to receive external audio signals when the device containing the audio component is in an operating mode, such as call mode, recording mode, or voice recognition mode. The received audio signals can be further stored in memory or transmitted via a communication component. In some embodiments, the audio component also includes a speaker for outputting audio signals.

[0145] Accordingly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to implement the steps in the above-described method embodiments. The computer-readable storage medium includes volatile or non-volatile components, or a combination thereof, and can be removable or non-removable. Examples of computer-readable storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), flash memory or other memory technologies, CD-ROM, Digital Video Disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium. Accordingly, this application also provides a computer program product, which includes a computer program or instructions that, when executed by a processor, cause the processor to implement the steps in the above method embodiments. It should be understood that each step or combination of steps in the above method flow can be implemented by the computer program or instructions. Furthermore, these computer programs or instructions can be applied to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device, enabling the processor of the general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to function as an apparatus for implementing the corresponding functions in the above method embodiments.

[0146] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0147] It should also be noted that 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 process, method, article, or apparatus. Unless otherwise specified, 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 that element.

[0148] Finally, it should be noted that the above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for classifying peripheral neuroblastomas, characterized in that, include: Acquire a first medical image, a second medical image, and a third medical image of the target user; the first medical image is acquired using positron emission tomography (PET) technology, the second medical image is acquired using computed tomography (CT) technology, and the third medical image is obtained by fusing the first medical image and the second medical image; A target classification model is determined, which includes a feature extraction module, a feature fusion module, and a classification module. The feature extraction module includes a first feature extraction network, a second feature extraction network, and a third feature extraction network connected in parallel. The first image features of the first medical image are extracted using the first feature extraction network, the second image features of the second medical image are extracted using the second feature extraction network, and the third image features of the third medical image are extracted using the third feature extraction network. The feature fusion module is used to fuse the first image feature, the second image feature, and the third image feature to obtain fused image features; The classification module is used to classify the features of the fused image to obtain the classification result of the peripheral neuroblastoma of the target user; The target classification model is trained on a training dataset, which includes a labeled dataset and an unlabeled dataset. The labeled dataset includes multiple first sample pairs and corresponding category labels for peripheral neuroblastoma tumors. The unlabeled dataset includes multiple second sample pairs. The target model parameters of the target classification model are trained on the labeled dataset when the training conditions are met. At least one third sample pair among the multiple second sample pairs and corresponding category labels are used to update the labeled dataset when the training conditions are not met.

2. A method for training a target classification model, characterized in that, include: Construct a target classification model that includes a feature extraction module, a feature fusion module, and a classification module; The feature extraction module includes a first feature extraction network, a second feature extraction network, and a third feature extraction network connected in parallel. Obtain a training dataset; the training dataset includes a labeled dataset and an unlabeled dataset. The labeled dataset includes multiple first sample pairs and corresponding category labels for peripheral neuroblastoma tumors. The unlabeled dataset includes multiple second sample pairs. Each sample pair includes a first sample image, a second sample image, and a third sample image. The first sample image is obtained from the sample user using positron emission tomography (PET) technology. The second sample image is obtained from the sample user using computed tomography (CT) technology. The third sample image is obtained by fusing the first sample image and the second sample image. The target classification model is trained using the labeled dataset to obtain the model parameters of the target classification model; If the training conditions are not met, at least one third sample pair is selected from the plurality of second sample pairs; Output annotation prompts and, in response to annotation requests, obtain the category labels corresponding to the at least one third sample pair respectively; Based on the at least one third sample pair and their corresponding category labels, update the labeled dataset and return to the step of training the target classification model using the labeled dataset until the training conditions are met; Under the condition that the training conditions are met, the target model parameters of the target classification model are determined.

3. The method according to claim 2, characterized in that, The step of selecting at least one third sample pair from the plurality of second sample pairs includes: The target classification model is used to predict the multiple second sample pairs respectively to obtain the category prediction probability corresponding to each of the multiple second sample pairs. Based on the category prediction probability corresponding to each of the multiple second sample pairs, a first sample selection strategy is generated using a first policy network. Based on the predicted probabilities of the categories corresponding to the plurality of second sample pairs and the first sample selection strategy, the sample selection priority corresponding to the plurality of second sample pairs is determined. Select at least one second sample pair with high priority as at least one third sample pair; The method further includes: Based on the predicted probabilities of the categories corresponding to the multiple second sample pairs and the first sample selection strategy, the first value parameter is generated using the first value network. If the first value parameter does not meet the value condition, the updated target classification model is used to predict the updated second sample pairs respectively, and the predicted class probabilities corresponding to the updated second sample pairs are obtained. Based on the predicted class probabilities corresponding to the updated second sample pairs respectively, the second policy network is used to generate a second sample selection strategy. Based on the updated predicted probabilities of the categories corresponding to multiple second sample pairs and the second sample selection strategy, the second value parameters are generated using the second value network. Based on the difference information between the second value parameter and the first value parameter, the first value network and the first policy network are updated, and the step of generating a first sample selection policy using the first policy network is returned to execution.

4. The method according to claim 3, characterized in that, Also includes: The updated target classification model is used to predict the at least one third sample pair respectively, and the predicted category corresponding to the at least one third sample pair is obtained respectively. The reward parameters are determined based on the predicted category and category label corresponding to the at least one third sample pair. The step of updating the first value network and the first policy network based on the difference information between the second value parameter and the first value parameter includes: The target value parameter is obtained by weighting the second value parameter with the reward parameter. The first value network is updated based on the difference information between the target value parameter and the first value parameter; The first policy network is updated based on the updated first value network.

5. The method according to claim 4, characterized in that, The step of updating the first value network based on the difference information between the target value parameter and the first value parameter includes: Calculate the first expectation of the squared error between the target value parameter and the first value parameter; Calculate the gradient of the first expectation with respect to the network parameters of the first value network, and update the network parameters of the first value network along the gradient descent direction; The step of updating the first policy network based on the updated first value network includes: Calculate the second expectation of the first value parameter corresponding to the updated first value network; Calculate the gradient of the second expectation with respect to the network parameters of the first policy network, and update the network parameters of the first policy network along the gradient ascending direction.

6. The method according to claim 3, characterized in that, The training conditions include one or more of the following: The ratio of the number of the first sample pairs in the labeled dataset reaches a preset ratio; The evaluation metric value of the target classification model on the test dataset reaches the preset metric value.

7. The method according to claim 2, characterized in that, The constructed target classification model, comprising a feature extraction module, a feature fusion module, and a classification module, includes: A first feature extraction network is extracted from a first classification model, a second feature extraction network is extracted from a second classification model, and a third feature extraction network is extracted from a third classification model; the first classification model includes the first feature extraction network and the first classification network, which are pre-trained based on a first training sample and corresponding training labels; the second classification model includes the second feature extraction network and the second classification network, which are pre-trained based on a second training sample and corresponding training labels; the third classification model includes the third feature extraction network and the third classification network, which are pre-trained based on a third training sample and corresponding training labels. The first feature extraction network, the second feature extraction network, and the third feature extraction network are connected in parallel to construct a feature extraction module; A feature fusion module is constructed; the feature fusion module includes pooling layers connected to the first feature extraction network, the second feature extraction network and the third feature extraction network respectively, and a splicing layer connected to the three pooling layers respectively; A classification module is constructed; the classification module includes a fully connected layer, and the weight parameters of the fully connected layer are obtained by fusing the first weight parameters of the fully connected layer in the first classification network, the second weight parameters of the fully connected layer in the second classification network, and the third weight parameters of the fully connected layer in the third classification network.

8. A computing device, characterized in that, It includes a storage component and a processing component; the storage component stores one or more computer program instructions, which are called and executed by the processing component, and the processing component executes the one or more computer program instructions to implement the peripheral neuroblastoma classification method as described in claim 1, or the target classification model training method as described in any one of claims 2 to 7.

9. A computer-readable storage medium, characterized in that, The system contains a computer program that is executed by a computer to implement the peripheral neuroblastoma classification method as described in claim 1, or the target classification model training method as described in any one of claims 2 to 7.

10. A computer program product, characterized in that, The system contains a computer program that, when executed by a computer, implements the peripheral neuroblastoma classification method as described in claim 1, or the target classification model training method as described in any one of claims 2 to 7.

Images