ANA detection method and device based on multi-task network, equipment and medium
Through multi-task network and image block slicing technology, karyotype and titer recognition of anti-nuclear antibody ANA is achieved, solving the problem that cannot be recognized by the existing technology and improving the recognition accuracy.
Patent Information
- Application Number
- CN202411488934.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-24
- Publication Date
- 2024-11-26
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The prior art cannot simultaneously recognize anti-nuclear antibody (ANA) karyotypes and titers, and cannot meet the needs of actual disease diagnosis and evaluation.
A multi-task network is used to extract richer image features by segmenting image blocks, and two task networks in the pre-trained ANA recognition model are used to identify karyotypes and titers respectively.
It improves the identification accuracy of ANA karyotype and titer, and meets the needs of actual diagnosis and evaluation.
Smart Images

Figure CN119027744A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of antinuclear antibody recognition, and in particular to a multi-task network-based ANA detection method, device, equipment and medium. Background Art
[0002] Antinuclear antibody (ANA) pattern recognition has become a widely used technology for routine ANA diagnosis and screening in clinical laboratories. In recent years, deep learning methods have made significant progress in identifying ANA patterns. The karyotype discrimination of ANA is mainly based on the microscopic morphology of cells in the image. Therefore, the recognition of ANA karyotype is very important. However, in actual work, ANA karyotype and other information are usually needed to assist in disease diagnosis and evaluation, but the existing technology can only identify various karyotypes of ANA and cannot meet actual needs. Therefore, this case proposes a multi-scale and multi-task deep convolutional neural network implementation scheme. Summary of the invention
[0003] The present application provides an ANA detection method, device, electronic device and storage medium based on a multi-task network, which can simultaneously identify the ANA karyotype and titer through two task networks, and extract richer and more accurate image features by segmenting image blocks to identify the titer and karyotype, thereby improving the recognition accuracy.
[0004] In a first aspect, an embodiment of the present application provides an ANA detection method based on a multi-task network, the method comprising: Acquire a first fluorescent image of the target user including antinuclear antibodies ANA; Dividing the first fluorescent picture into a plurality of first image blocks according to a first preset number of divisions, and determining a first image feature of the first fluorescent picture according to sub-image features of the plurality of first image blocks; The first image feature of the first fluorescence image is respectively calculated and processed by the first task network and the second task network of the pre-trained ANA recognition model, and the karyotype classification result of the first fluorescence image corresponding to the first task network and the titer classification result corresponding to the second task network are output; An ANA test result for the target user is generated according to the karyotype classification result and the titer classification result of the first fluorescent image.
[0005] In a possible implementation manner, determining the first image feature of the first fluorescent picture according to the sub-image features of the plurality of first image blocks includes: The multiple first image blocks are input into the ANA recognition model, the sub-image features of each first image block are extracted through the feature extraction network of the ANA recognition model, and the sub-image features of the multiple first image blocks are spliced to obtain the first image features of the first fluorescent image.
[0006] In a possible implementation, the first task network and the second task network include the same network architecture; the network architecture of the first task network corresponds to outputting a karyotype classification result; the network architecture of the second task network corresponds to outputting a titer classification result; The network architecture includes: a linear layer, a nominal layer, a multi-head attention mechanism layer, a convolutional layer, a nominal layer, a multi-head attention mechanism layer and a multi-layer perceptron connected in sequence.
[0007] In a possible implementation, the ANA recognition model is trained by the following method: Preprocessing candidate fluorescence images including a plurality of different karyotypes to obtain training fluorescence images, and constructing a target data set based on the training fluorescence images; Pre-constructing an original ANA recognition model, the original ANA recognition model comprising a feature extraction network and a first task processing network and a second task processing network respectively connected to the feature extraction network; The original ANA recognition model is trained using the target data set to obtain a trained ANA recognition model.
[0008] In a possible implementation, the preprocessing of candidate fluorescence images including a plurality of different karyotypes to obtain training fluorescence images includes: According to the cell distribution information in each of the candidate fluorescent images, selecting a second fluorescent image whose corresponding cell distribution information satisfies a cell distribution condition from the candidate fluorescent images; the cell distribution condition includes that at least one of the number of cells, the cell distribution characteristics and the cell clarity satisfies a corresponding preset threshold; By performing data enhancement processing on the second fluorescence image, a third fluorescence image that meets the diversity condition is obtained, and the third fluorescence image is normalized to obtain a training fluorescence image; wherein the diversity condition includes that the number of samples reaches a preset threshold and the cell distribution characteristics are different under the cell distribution condition.
[0009] In a possible implementation, selecting, according to the cell distribution information in each of the candidate fluorescence images, a second fluorescence image whose corresponding cell distribution information satisfies a cell distribution condition from the candidate fluorescence images includes: For each candidate fluorescence picture among the multiple candidate fluorescence pictures, dividing the candidate fluorescence picture into multiple second image blocks according to a second preset number of divisions, and respectively calculating cell information of each of the second image blocks; Calculating cell distribution information of the candidate fluorescence image according to the cell information of the plurality of second image blocks; It is determined whether the cell distribution information of the candidate fluorescent image meets the cell distribution condition. If not, the candidate fluorescent image is eliminated to obtain a second fluorescent image whose corresponding cell distribution information in the candidate fluorescent image meets the cell distribution condition.
[0010] In a possible implementation, performing data enhancement processing on the second fluorescent image to obtain a third fluorescent image that meets a diversity condition includes: determining a target enhancement method for the second fluorescent image according to the cell distribution characteristics of the second fluorescent image; The second fluorescent image is enhanced according to the target enhancement method; wherein the target enhancement method includes at least one of the following: geometric transformation, color transformation, noise addition and blur processing; the geometric transformation includes at least one of the following: rotation, flipping and scaling.
[0011] In a possible implementation manner, the target data set includes a training data set and a validation data set; and the training of the original ANA recognition model by using the target data set to obtain a trained ANA recognition model includes: Setting the hyperparameters of the ANA recognition model, and training the training parameters in the original ANA recognition model in a round-based manner using the training data set under the hyperparameters; the training parameters include parameters of the feature extraction network, the first task network, and the second task network; After each round of training, the ANA recognition model under the training parameters of this round is tested through the validation set. If the result of the validation data set is better than the output result of the previous round, the training parameters of the current round of the ANA recognition model are saved until a trained ANA recognition model is obtained.
[0012] In a second aspect, an embodiment of the present application further provides an ANA detection device based on a multi-task network, the device comprising: An acquisition module is used to acquire a first fluorescent image of a target user including antinuclear antibodies ANA; A segmentation module, configured to segment the first fluorescent image into a plurality of first image blocks according to a first preset segmentation quantity; A processing module, configured to determine a first image feature of the first fluorescence image according to the sub-image features of the plurality of first image blocks; respectively calculate and process the first image feature of the first fluorescence image through a first task network and a second task network of a pre-trained ANA recognition model, and output a karyotype classification result of the first fluorescence image corresponding to the first task network and a titer classification result corresponding to the second task network; A generation module is used to generate an ANA test result for the target user according to the karyotype classification result and the titer classification result of the first fluorescent image.
[0013] In a third aspect, an embodiment of the present application further provides an electronic device, comprising: a processor, a storage medium and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the storage medium through the bus, and the processor executes the machine-readable instructions to perform the steps of the ANA detection method based on a multi-task network as described in any one of the first aspects.
[0014] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the ANA detection method based on a multi-task network as described in any one of the first aspects are executed.
[0015] The present application provides the above-mentioned ANA detection method, device, equipment and medium based on multi-task network, by obtaining the first fluorescent image of the target user including antinuclear antibody ANA, and dividing the first fluorescent image into multiple first image blocks according to the first preset number of divisions, and then determining the first image feature of the first fluorescent image according to the image features of the multiple first image blocks; after that, the first image features of the first fluorescent image are respectively calculated and processed by the first task network and the second task network of the pre-trained ANA recognition model, and the karyotype classification result of the first fluorescent image corresponding to the first task network and the titer classification result corresponding to the second task network are output; finally, according to the karyotype classification result and titer classification result of the first fluorescent image, the ANA detection result for the target user is generated. The present application can simultaneously identify the ANA karyotype and titer through two task networks, and extract richer and more accurate image features for titer and karyotype identification by dividing the image blocks, thereby improving the identification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.
[0017] Figure 1 A flowchart of a first ANA detection method based on a multi-task network provided in an embodiment of the present application is shown; Figure 2 A schematic diagram of the structure of the trained ANA recognition model provided in an embodiment of the present application is shown; Figure 3 A flowchart of a second ANA detection method based on a multi-task network provided in an embodiment of the present application is shown; Figure 4 A flowchart of a third ANA detection method based on a multi-task network provided in an embodiment of the present application is shown; Figure 5 A schematic diagram showing the multi-karyotype classification performance of the ANA recognition model in one dimension is shown; Figure 6 A schematic diagram showing the multi-karyotype classification performance of the ANA recognition model in another dimension is shown; Figure 7 A schematic diagram of the titer regression performance of the ANA recognition model is shown; Figure 8 A schematic diagram of the structure of an ANA detection device based on a multi-task network provided in an embodiment of the present application is shown; Fig. 9 A schematic structural diagram of an electronic device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION
[0018] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It should be understood that the drawings in the present application only serve the purpose of explanation and description and are not used to limit the scope of protection of the present application. In addition, it should be understood that the schematic drawings are not drawn in real proportion. The flowchart used in this application shows the operations implemented according to some embodiments of the present application. It should be understood that the operations of the flowchart can be implemented out of sequence, and the steps without logical context can be reversed in order or implemented simultaneously. In addition, those skilled in the art can add one or more other operations to the flowchart under the guidance of the content of the present application, or remove one or more operations from the flowchart.
[0019] In addition, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.
[0020] It should be noted that the term "comprising" will be used in the embodiments of the present application to indicate the existence of the features declared thereafter, but does not exclude the addition of other features.
[0021] In the field of antinuclear antibody (ANA) pattern recognition, the karyotype discrimination of ANA is mainly based on the microscopic morphology of cells in the image. Therefore, the recognition of ANA karyotype is very important. However, in actual work, ANA karyotype and other information are usually needed to assist in disease diagnosis and evaluation, but the existing technology can only identify various karyotypes of ANA and cannot meet actual needs. Based on this, the embodiment of the present application provides an ANA detection method, device, electronic device and storage medium based on a multi-task network, which can simultaneously identify ANA karyotype and titer through two task networks, and extract richer and more accurate image features by segmenting image blocks to identify titer and karyotype, thereby improving recognition accuracy.
[0022] like Figure 1 As shown, an embodiment of the present application provides an ANA detection method based on a multi-task network, the method comprising: S101, obtaining a first fluorescent image of a target user including antinuclear antibodies ANA; S102, dividing the first fluorescent picture into a plurality of first image blocks according to a first preset number of divisions, and determining a first image feature of the first fluorescent picture according to sub-image features of the plurality of first image blocks; S103, respectively calculating and processing the first image feature of the first fluorescence image through the first task network and the second task network of the pre-trained ANA recognition model, and outputting the karyotype classification result of the first fluorescence image corresponding to the first task network and the titer classification result corresponding to the second task network; S104: Generate an ANA test result for the target user according to the karyotype classification result and the titer classification result of the first fluorescent image.
[0023] In the above-mentioned ANA detection method based on multi-task network, the first fluorescent image of the target user including antinuclear antibody ANA is obtained, and the first fluorescent image is divided into multiple first image blocks according to the first preset number of divisions, and then the first image features of the first fluorescent image are determined according to the image features of the multiple first image blocks; then, the first image features of the first fluorescent image are respectively calculated and processed by the first task network and the second task network of the pre-trained ANA recognition model, and the karyotype classification result of the first fluorescent image corresponding to the first task network and the titer classification result corresponding to the second task network are output; finally, the ANA detection result for the target user is generated according to the karyotype classification result and the titer classification result of the first fluorescent image. The present application can simultaneously identify the ANA karyotype and titer through two task networks, and extract richer and more accurate image features for titer and karyotype identification by dividing the image blocks, thereby improving the identification accuracy.
[0024] The above exemplary embodiments are described below respectively: S101 , obtaining a first fluorescent image of a target user including antinuclear antibodies ANA.
[0025] In an embodiment of the present application, an ANA recognition model is pre-trained. When it is necessary to diagnose and analyze an autoimmune disease for a certain user (such as a target user), a first fluorescent image of the target user including antinuclear antibodies ANA is obtained, and then the first fluorescent image is input into the above-mentioned trained ANA recognition model, and the karyotype classification result and titer classification result in the first fluorescent image are output through the ANA recognition model.
[0026] Here, the first fluorescent image of the target user can be a fluorescent image of the target user including ANA, or can be selected from multiple fluorescent images of the target user including ANA. Here, the first fluorescent image selected above is a fluorescent image whose cell distribution information satisfies the cell distribution condition among the multiple fluorescent images; the cell distribution condition satisfies the corresponding preset threshold value including at least one of the cell number, cell distribution characteristics and cell clarity; in practice, the cell number corresponds to the first preset threshold value, the cell distribution corresponds to the second preset threshold value, and the cell clarity corresponds to the third preset threshold value; the above three preset threshold values are set according to the actual situation. It should be noted that the three preset threshold values in the above cell distribution condition are determined according to the mutual influence of the above cell number, cell distribution characteristics and cell clarity; for example, the cell distribution condition is that the first preset threshold value is the cell number 5, the second preset threshold value is the distribution characteristic 1, and the third threshold value is the cell clarity 7; when the cell number 5 increases (for example, increases to 10), the requirement for distribution characteristic 1 can be reduced (for example, the requirement for distribution uniformity can be weakened, which means that it is not necessary to have uniform distribution, for example, it is allowed to be more in the upper right part of the picture and less in the lower right part), and accordingly, distribution characteristic 1 can be 0.5.
[0027] S102: Divide the first fluorescent picture into a plurality of first image blocks according to a first preset number of divisions, and determine a first image feature of the first fluorescent picture according to sub-image features of the plurality of first image blocks.
[0028] Here, the first preset number of cuts is determined according to the cell distribution characteristics of the target data set (and the data set used for the ANA recognition model) and the preset recognition accuracy, for example, it can be 9. For each first fluorescent image (i.e., a field of view), a field of view is used as a Bag, and a field of view is cut into multiple Patches according to different scales. Without loss of generality, a field of view is cut into 9 patches as Instances in the Bag.
[0029] In the embodiment of the present application, for a Bag, the image features (i.e., sub-image features) of each Instance in the Bag are first extracted respectively, and then the first image features of the first fluorescent image are determined according to the sub-image features of multiple Instances in the Bag. By dividing the image blocks as described above, more abundant and accurate image features can be extracted, thereby improving the accuracy of subsequent titer and karyotype recognition.
[0030] Afterwards, the multiple first image blocks are input into a pre-trained ANA recognition model, a sub-image feature of each first image block is extracted through a feature extraction network of the ANA recognition model, and the sub-image features of the multiple first image blocks are spliced to obtain a first image feature of the first fluorescent image; Here, if Figure 2 As shown, the pre-trained ANA recognition model includes a feature extraction network and a first task processing network and a second task processing network respectively connected to the feature extraction network; the feature extraction network is used to extract sub-image features of the first image block, and to concatenate sub-image features of multiple first image blocks in the first fluorescent image to obtain the first image features of the first fluorescent image. Specifically, each instance in each bag is input into the ANA recognition model respectively, and the ANA recognition model maps each instance to a feature embedding of a fixed length through the feature extraction network, and connects the feature embeddings of all instances in a bag to obtain the first image features of the first fluorescent image.
[0031] As an optional implementation, in the embodiment of the present application, ResNet is used as a feature extractor. ResNet maps each Instance to a feature embedding of a fixed length (specifically 1024 in this application, and the fixed length varies according to the task), and then connects all the feature embeddings to form a feature embedding of the Bag, that is, the feature token of each Instance.
[0032] In addition, in addition to ResNet, the above feature extraction network can also use ORBextiactor and the like.
[0033] S103, respectively calculating and processing the first image feature of the first fluorescence image through the first task network and the second task network of the pre-trained ANA recognition model, and outputting the karyotype classification result of the first fluorescence image corresponding to the first task network and the titer classification result corresponding to the second task network.
[0034] In an embodiment of the present application, the pre-trained ANA recognition model includes a feature extraction network, and also includes a first task processing network and a second task processing network respectively connected to the above-mentioned feature extraction network; after the feature extraction network extracts the first image feature, it will be input into the first task processing network and the second task processing network respectively, and the first task processing network and the second task processing network will respectively calculate and process the first image feature, and correspondingly, the first task processing network outputs the karyotype classification result, and the second task processing network outputs the titer classification result.
[0035] In an embodiment of the present application, the first task network and the second task network include the same network architecture; the network architecture of the first task network corresponds to the output karyotype classification result, and the network architecture of the second task network corresponds to the output titer classification result; the network architecture includes: a linear layer, a nominal layer, a multi-head attention mechanism layer, a convolutional layer, a nominal layer, a multi-head attention mechanism layer and a multi-layer perceptron connected in sequence; The first task network performs dimensionality reduction processing on the first image feature of the first fluorescent image through a linear layer, normalizes the first image feature after dimensionality reduction through a nominal layer, and through a multi-head attention mechanism layer, splits the original embedding space into multiple subspaces, and independently calculates the attention weight in each subspace, splices the attention output calculated in each subspace to form a larger vector, and converts this spliced vector back to the original space through a linear layer to obtain the final output; the convolution layer is used to use the output of the multi-head attention mechanism layer as input, obtain the context information of the first image feature through convolution kernels of different sizes, use convolution kernels of different sizes in the same layer, encode position information of different granularities, and sum up to obtain the feature moment that finally integrates the position features; through a multi-layer perceptron, the input features are transformed through one or more fully connected layers to obtain the final karyotype classification result.
[0036] The second task network performs dimensionality reduction processing on the first image feature of the first fluorescent image through a linear layer, normalizes the first image feature after dimensionality reduction through a nominal layer, and through a multi-head attention mechanism layer, splits the original embedding space into multiple subspaces, and independently calculates the attention weight in each subspace, splices the attention output calculated in each subspace to form a larger vector, and converts this spliced vector back to the original space through a linear layer to obtain the final output; the convolution layer is used to use the output of the multi-head attention mechanism layer as input, obtain the context information of the first image feature through convolution kernels of different sizes, use convolution kernels of different sizes in the same layer, encode position information of different granularities, and sum up to obtain the feature moment that finally integrates the position features; through a multi-layer perceptron, the input features are transformed through one or more fully connected layers to obtain the final titer classification result.
[0037] S104: Generate an ANA test result for the target user according to the karyotype classification result and the titer classification result of the first fluorescent image.
[0038] In an embodiment of the present application, the ANA test result of the target user may be an analysis result for the target user, which may specifically include the analysis results of the karyotype and titer of the first fluorescent image, and may also include other results, such as cell distribution, disease analysis results and suggestions, etc.
[0039] Further, such as Figure 3 As shown, in the ANA detection method based on a multi-task network provided in the embodiment of the present application, the ANA recognition model is trained by the following method: S301 , preprocessing candidate fluorescence images including a plurality of different karyotypes to obtain training fluorescence images, and constructing a target data set based on the training fluorescence images.
[0040] In practice, bad candidate fluorescence images will affect model training. Therefore, the above preprocessing includes eliminating candidate fluorescence images whose corresponding cell distribution information does not meet the cell distribution conditions to obtain a second fluorescence image. This method is also called data cleaning.
[0041] Here, when the number of the second fluorescence images meets the preset threshold (i.e., the fourth preset threshold) and meets the cell distribution feature diversity condition, the second fluorescence images are directly determined as training fluorescence images, and a training data set (i.e., a training set) and a verification data set (i.e., a verification set) are constructed based on the training fluorescence images; and when the number of the second fluorescence images is less than the preset threshold (i.e., the fourth preset threshold) and / or does not meet the cell distribution feature diversity condition, data enhancement processing is performed on the second fluorescence images to obtain third fluorescence images whose number meets the preset threshold (i.e., the fourth preset threshold) and meets the cell distribution feature diversity condition, the second fluorescence images and the third fluorescence images are determined as training fluorescence images, and a training data set (i.e., a training set) and a verification data set (i.e., a verification set) are constructed based on the training fluorescence images; S302 , pre-constructing an original ANA recognition model, wherein the original ANA recognition model includes a feature extraction network and a first task processing network and a second task processing network respectively connected to the feature extraction network.
[0042] In the embodiment of the present application, a feature extraction network of the original ANA recognition model and a first task processing network and a second task processing network respectively connected to the feature extraction network are constructed; the feature extraction network is used to extract sub-image features, and multiple sub-image features are spliced to obtain image features. Specifically, ResNet is used as a feature extractor, and ResNet maps each Instance to a feature embedding of a fixed length (specifically 1024 in this application, and the fixed length varies depending on the task), and then all the feature embeddings are connected to form a feature embedding of Bag, that is, a feature token of each Instance.
[0043] After extracting the first image features, the feature extraction network will input them into the first task processing network and the second task processing network respectively. The first task processing network and the second task processing network will respectively perform calculations on the first image features. Correspondingly, the first task processing network outputs the karyotype classification results, and the second task processing network outputs the titer classification results.
[0044] In an embodiment of the present application, the first task network and the second task network include the same network architecture; the network architecture of the first task network corresponds to the output karyotype classification result, and the network architecture of the second task network corresponds to the output titer classification result; the network architecture includes: a linear layer, a nominal layer, a multi-head attention mechanism layer, a convolutional layer, a nominal layer, a multi-head attention mechanism layer and a multi-layer perceptron connected in sequence; Here, the linear layer is the Linear Layer, which reduces the dimension of the image feature token of the fluorescent image, from 1024 to 512 here; the nominal layer is the Nom Layer, which normalizes the reduced-dimensional image features; the multi-head attention mechanism layer, namely Multi-head Self-attention, splits the original embedding space into multiple subspaces, calculates the attention weight independently in each subspace, and splices the attention outputs calculated in each subspace to form a larger vector, and then passes through a linear layer to convert the spliced vector back to the original space to obtain the final output of Multi-head Self-attention. Here, the Nyström method is used to approximate the standard Attention; the convolution layer is GConv, where different convolution layers correspond to convolution kernels of different sizes. Multiple convolution layers use the output of Multi-head Self-attention as input and obtain the context information of the token through convolution kernels of different sizes (specifically: 7*7, 5*5, and 3*3). That is, different sizes of convolution kernels are used in the same layer to encode position information of different granularities, and the sum is obtained to obtain the feature matrix that finally integrates the position features. Multilayer Perceptron is MLP Head, which further transforms the input features through one or more fully connected layers (i.e., linear layer) to obtain the final classification result. In the embodiment of the present application, a fully connected layer is used; it should be noted that the linear layer is also a fully connected layer. Karyotype corresponds to the predicted karyotype result; titer corresponds to the predicted titer result.
[0045] Based on the above structure and principle, in the embodiment of the present application, the first task network and the second task network constructed have the following functions: The first task network reduces the dimension of the image features of the fluorescent image through a linear layer, normalizes the reduced image features through a nominal layer, splits the original embedding space into multiple subspaces through a multi-head attention mechanism layer, and independently calculates the attention weight in each subspace, concatenates the attention outputs calculated in each subspace to form a larger vector, and converts this concatenated vector back to the original space through a linear layer to obtain the final output; the convolution layer is used to use the output of the multi-head attention mechanism layer as input, obtain the context information of the first image feature through convolution kernels of different sizes, use convolution kernels of different sizes in the same layer, encode position information of different granularities, and sum up to obtain the feature moment that finally integrates the position features; through a multi-layer perceptron, the input features are transformed through one or more fully connected layers to obtain the final karyotype classification result.
[0046] The second task network reduces the dimension of the image features of the fluorescent image through a linear layer, normalizes the reduced image features through a nominal layer, splits the original embedding space into multiple subspaces through a multi-head attention mechanism layer, and independently calculates the attention weight in each subspace, splices the attention outputs calculated in each subspace to form a larger vector, and converts this spliced vector back to the original space through a linear layer to obtain the final output; the convolution layer is used to use the output of the multi-head attention mechanism layer as input, obtain the context information of the first image feature through convolution kernels of different sizes, use convolution kernels of different sizes in the same layer, encode position information of different granularities, and sum up to obtain the feature moment that finally integrates the position features; through a multi-layer perceptron, the input features are transformed through one or more fully connected layers to obtain the final titer classification result.
[0047] S303: training the original ANA recognition model using the target data set to obtain a trained ANA recognition model.
[0048] After constructing a training data set (ie, training set) and a validation data set (ie, validation set) based on the training fluorescence images, the original ANA recognition model is trained based on the above training set and validation set to obtain a trained ANA recognition model.
[0049] Further, such as Figure 4 As shown, the candidate fluorescence images including a plurality of different karyotypes are preprocessed to obtain training fluorescence images, including: S401. According to the cell distribution information in each of the candidate fluorescence images, select a second fluorescence image whose corresponding cell distribution information satisfies a cell distribution condition from the candidate fluorescence images; satisfying the cell distribution condition includes at least one of the number of cells, cell distribution characteristics and cell clarity satisfying a corresponding preset threshold.
[0050] In the embodiment of the present application, the above-mentioned cell distribution information includes the number of cells, the characteristics of cell distribution and the clarity of cells, and the cell distribution conditions are pre-set. The above-mentioned cell distribution conditions include that at least one of the number of cells, the characteristics of cell distribution and the clarity of cells satisfies the corresponding preset threshold value; in practice, the number of cells corresponds to the first preset threshold value, the cell distribution corresponds to the second preset threshold value, and the clarity of cells corresponds to the third preset threshold value; the above-mentioned three preset threshold values are set according to the actual situation. It should be noted that the three preset threshold values in the above-mentioned cell distribution conditions are determined according to the mutual influence of the above-mentioned number of cells, the characteristics of cell distribution and the clarity of cells; for example, the cell distribution condition is that the first preset threshold value is the number of cells 5, the second preset threshold value is the distribution characteristic 1, and the third threshold value is the level 2 of the clarity of cells; when the number of cells 5 increases (for example, increases to 10), the requirements for the distribution characteristic 1 can be reduced (for example, the requirements for the uniformity of distribution can be weakened, which means that it is not necessary to have a uniform distribution, for example, it is allowed to be more in the upper right part of the picture and less in the lower right part). Then, based on the set cell distribution condition, it is determined whether the cell distribution information of each candidate fluorescence image meets the cell distribution condition. If not, the candidate fluorescence image is eliminated, and the remaining candidate fluorescence image is the second fluorescence image.
[0051] Specifically, the candidate fluorescent image is cut into multiple second image blocks according to the second preset number of cuts, and the cell information of each second image block is calculated respectively; the cell distribution information of the candidate fluorescent image is calculated according to the cell information of the multiple second image blocks; it is determined whether the cell distribution information of the candidate fluorescent image meets the cell distribution condition, if not, the candidate fluorescent image is eliminated, and the second fluorescent image whose corresponding cell distribution information in the candidate fluorescent image meets the cell distribution condition is obtained. For example, the image is cut into 3 * 3 patches of the same size, and for each patch, the pixel average value of the green channel is calculated. If the average value is less than the preset threshold (here, the fifth preset threshold, the exemplary fifth preset threshold is 29), it is considered that there are no cells under the patch. If more than 3 patches are determined to have no cells, the image is discarded.
[0052] In the embodiment of the present application, the reason why the cell information of the patch is calculated instead of the whole is that in actual situations, we found that in some fields of view, although the overall brightness is high, there are almost no cells, and there are one or two areas that are very bright, and other areas are very dark. Therefore, the overall calculation is biased.
[0053] S402. Performing data enhancement processing on the second fluorescence image to obtain a third fluorescence image that meets a diversity condition, and performing normalization processing on the third fluorescence image to obtain a training fluorescence image; wherein the diversity condition includes that the number of samples reaches a preset threshold and the cell distribution characteristics are different under the cell distribution condition.
[0054] Here, in this field, the number of second fluorescent images that meet the cell distribution conditions is usually insufficient, so it is necessary to perform data enhancement processing on the second fluorescent images. Here, different second fluorescent images may correspond to different enhancement processing methods. Specifically, the target enhancement method for the second fluorescent image can be determined according to the cell distribution characteristics of the second fluorescent image; then, the second fluorescent image is enhanced according to the target enhancement method; wherein the above-mentioned target enhancement method includes at least one of the following: geometric transformation, color transformation, adding noise and blur processing; the geometric transformation includes at least one of the following: rotation, flipping and scaling. In practice, during the training process, the above-mentioned data enhancement method can be matched for all training data, or the above-mentioned data enhancement method can be used randomly, in order to increase the diversity of images and improve model performance and generalization ability.
[0055] A third fluorescent image satisfying the diversity condition is obtained, and the third fluorescent image is normalized. Specifically, the pixel values of the image data of the third fluorescent image are normalized to a uniform numerical range. Specifically, the pixel values are normalized from (0-255) to (0-1). The specific normalization method adopts maximum-minimum value normalization: the image data is scaled to a range of 0 to 1, that is, each pixel value is subtracted from the minimum value and divided by the difference between the maximum value and the minimum value to obtain a training fluorescent image. The above normalization processing method can accelerate convergence, improve model generalization ability, avoid gradient vanishing and explosion, and simplify model parameter initialization.
[0056] The following specifically describes the training method of the original ANA recognition model based on the target data set in combination with the original ANA recognition model. Figure 2As shown, the first task network and the second task network include the same network architecture; the network architecture of the first task network corresponds to the output karyotype classification result; the network architecture of the second task network corresponds to the output titer classification result; the network architecture includes: a linear layer, a nominal layer, a multi-head attention mechanism layer, a convolutional layer, a nominal layer, a multi-head attention mechanism layer and a multi-layer perceptron connected in sequence; the linear layer is used to reduce the dimension of the image features of the fluorescence image; the nominal layer is used to normalize the image features after the dimensionality reduction processing; the multi-head attention mechanism layer is used to split the original embedding space into multiple subspaces, and independently in each subspace Calculate the attention weight, concatenate the attention outputs calculated in each subspace to form a larger vector, and convert the concatenated vector back to the original space through a linear layer to obtain the final output; the convolution layer is used to use the output of the multi-head attention mechanism layer as input, obtain the context information of the first image feature through convolution kernels of different sizes, use convolution kernels of different sizes in the same layer, encode position information of different granularities, and sum up to obtain the feature moment that finally integrates the position features; the multilayer perceptron is used to transform the input features through one or more fully connected layers to obtain the final karyotype classification results and titer classification results.
[0057] During the training process, a training fluorescence image is used as a field of view, and a field of view is used as a bag. A field of view is divided into multiple patches according to different scales (without loss of generality, a field of view is divided into 9 patches here) as instances in the bag. Each instance is input into the feature extraction network (the function of the network is to map each instance to a feature embedding of a fixed length and connect the feature embeddings of all instances in a bag). ResNet is used as a feature extractor here, which is not general. Each instance is mapped to a feature embedding of a fixed length (specifically 1024 in the embodiment of the present application, which varies according to different tasks), and then all the feature embeddings are connected to form the feature embedding of the bag, that is, the feature token of each instance.
[0058] In order to perform multi-core type classification and titer regression simultaneously, two classification branches are connected after the feature extractor, and the two branches use the same network structure. In each classification branch, a linear layer is used to reduce the dimension of the feature token, and the category token is randomly initialized (randomly initialized for classification) and connected to the feature token. The context information of the token is obtained through the multi-head self-attention mechanism and convolution kernels of different sizes (specifically: 7*7, 5*5, 3*3 convolution kernels), and the attention score between each pair of tokens is updated through the multi-head self-attention mechanism. Finally, the category prediction is completed through the MLP layer (Multi-Layer Perceptron).
[0059] During the network training process, the model hyperparameters are set, such as the learning rate is set to 0.0002, the weight decay is set to 0.0001, and the stochastic gradient descent is used as the optimizer. Under the hyperparameters, the training parameters in the original ANA recognition model are trained in a round-based manner through the training data set; the training parameters include the parameters of the feature extraction network, the first task network, and the second task network. In the embodiment of the present application, all the training parameters in the above original ANA recognition model are trained together, the focal loss loss function is calculated, and the gradient of the network parameters is calculated therefrom, and back propagation is performed. After one round of training, the ANA recognition model under the training parameters of this round is tested through the verification set. If the result of the verification data set is better than the output result of the previous round, the training parameters of the current round of the ANA recognition model are saved until the trained ANA recognition model is obtained. The f1-score is used to measure the quality of the verification result.
[0060] When evaluating performance based on the validation set, the multi-nucleotype classification performance is as follows: Figure 5 and 6 As shown in Figure 2, the optimal model F1 index, i.e., F1-Score, is 0.92, and the area under the curve, i.e., AUC, is 0.99. In the validation set, the titer regression performance is as follows: Figure 7 As shown, within an error range, the accuracy is 0.976.
[0061] Among them, Figure 5In the , "best model metrics" is "best model metrics", and the training cycle is "epoch", which is the training cycle of the entire training data set in the neural network; the evaluation indicators include: specificity, recall, precision, accuracy, F1 and F2 indicators, CK coefficient and area under the curve Auc; among them, F1 and F2 indicators correspond to F1-Score and F2-Score respectively, both of which are indicators used to evaluate the performance of classification models. They evaluate the performance of the model by comprehensively considering the precision and recall rate; here, F1-Score considers precision and recall rate to be equally important, while F2-Score focuses more on the importance of recall rate; CK coefficient, that is, Cohen Kappa coefficient, is an indicator to measure the consistency or accuracy of classification. In neural networks, it is used as an evaluation standard for classification accuracy in Kaggle competitions; area under the curve, Auc (Area Under the Curve), is an evaluation indicator to measure the performance of classification models. Specifically, AUC is the area under the ROC curve, which is used to evaluate the ability of the model to predict positive examples.
[0062] Among them, Figure 6 In the ROC curve, "ROC Curves" is also known as "receiver operating characteristic curve", which is a tool for evaluating the performance of classification models, especially in machine learning and deep learning. The ROC curve evaluates model performance by plotting the relationship between the true positive rate (TPR) and the false positive rate (FPR). TPR, also known as sensitivity or hit rate, measures the proportion of positive samples correctly identified by the model; FPR measures the proportion of negative samples incorrectly identified as positive by the model; "True Positive Rate" is "True Positive Rate"; "ROC curve of class0-7" means "ROC curve of classes 0-7" respectively; "micro-average ROc curve" means "micro-average ROc curve", and "area" means "area"; exist Figure 7 In the example, “Confusion matrix” is “Confusion matriX”, “Trule Labels” is “Trule Labels”, and “Predicted Labels” is “Predicted Labels”.
[0063] When the original ANA recognition model is trained by the target data set to obtain a trained ANA recognition model, the hyperparameters of the ANA recognition model are set, and the training parameters in the original ANA recognition model are trained in a round-based manner by using the training data set under the hyperparameters; the training parameters include parameters of a feature extraction network, a first task network, and a second task network; after each round of training, the ANA recognition model under the training parameters of the round is tested by using a validation set, and if the result of the validation data set is better than the output result of the previous round, the training parameters of the current round of this model are saved until a trained ANA recognition model is obtained.
[0064] The present application can simultaneously identify ANA karyotype and titer through two task networks, and extract richer and more accurate image features for titer and karyotype identification by segmenting image blocks, thereby improving the recognition accuracy.
[0065] Based on the same inventive concept, an ANA detection device based on a multi-task network corresponding to the ANA detection method based on a multi-task network is also provided in the embodiment of the present application. Since the principle of solving the problem by the device in the embodiment of the present application is similar to the above-mentioned ANA detection method based on a multi-task network in the embodiment of the present application, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be repeated.
[0066] Reference Figure 8 As shown, an ANA detection device based on a multi-task network provided in an embodiment of the present application, the device includes: An acquisition module 801 is used to acquire a first fluorescent image of a target user including antinuclear antibodies ANA; A segmentation module 802, configured to segment the first fluorescent image into a plurality of first image blocks according to a first preset segmentation quantity; The processing module 803 is used to determine the first image feature of the first fluorescence image according to the sub-image features of the multiple first image blocks; respectively calculate and process the first image feature of the first fluorescence image through the first task network and the second task network of the pre-trained ANA recognition model, and output the karyotype classification result of the first fluorescence image corresponding to the first task network and the titer classification result corresponding to the second task network; The generating module 804 is used to generate the ANA test result for the target user according to the karyotype classification result and the titer classification result of the first fluorescent image.
[0067] In a possible implementation manner, the processing module 803 determines the first image feature of the first fluorescent picture according to the sub-image features of the multiple first image blocks, including: The multiple first image blocks are input into the ANA recognition model, the sub-image features of each first image block are extracted through the feature extraction network of the ANA recognition model, and the sub-image features of the multiple first image blocks are spliced to obtain the first image features of the first fluorescent image.
[0068] In a possible implementation, the first task network and the second task network include the same network architecture; the network architecture of the first task network corresponds to outputting a karyotype classification result; the network architecture of the second task network corresponds to outputting a titer classification result; The network architecture includes: a linear layer, a nominal layer, a multi-head attention mechanism layer, a convolutional layer, a nominal layer, a multi-head attention mechanism layer and a multi-layer perceptron connected in sequence; The linear layer is used to perform dimensionality reduction processing on the first image feature of the first fluorescent image; the nominal layer is used to normalize the first image feature after dimensionality reduction processing; the multi-head attention mechanism layer is used to split the original embedding space into multiple subspaces, and independently calculate the attention weight in each subspace, splice the attention output calculated in each subspace to form a larger vector, and convert the spliced vector back to the original space through a linear layer to obtain the final output; the convolution layer is used to use the output of the multi-head attention mechanism layer as input, obtain the context information of the first image feature through convolution kernels of different sizes, use convolution kernels of different sizes in the same layer, encode position information of different granularities, and sum up to obtain the feature moment that finally integrates the position features; the multi-layer perceptron is used to transform the input features through one or more fully connected layers to obtain the final karyotype classification results and titer classification results.
[0069] In a possible implementation, the device further includes: The model training module is used to preprocess candidate fluorescence images including multiple different karyotypes to obtain training fluorescence images, and construct a target data set based on the training fluorescence images; pre-construct an original ANA recognition model, wherein the original ANA recognition model includes a feature extraction network and a first task processing network and a second task processing network respectively connected to the feature extraction network; and train the original ANA recognition model through the target data set to obtain a trained ANA recognition model.
[0070] In a possible implementation, the model training module preprocesses candidate fluorescence images including a plurality of different karyotypes to obtain training fluorescence images, including: According to the cell distribution information in each of the candidate fluorescent images, a second fluorescent image whose corresponding cell distribution information satisfies a cell distribution condition is selected from the candidate fluorescent images; the cell distribution condition includes that at least one of the number of cells, the cell distribution characteristics and the cell clarity satisfies a corresponding preset threshold; By performing data enhancement processing on the second fluorescence image, a third fluorescence image that meets the diversity condition is obtained, and the third fluorescence image is normalized to obtain a training fluorescence image; wherein the diversity condition includes that the number of samples reaches a preset threshold and the cell distribution characteristics are different under the cell distribution condition.
[0071] In a possible implementation, the model training module selects, according to the cell distribution information in each of the candidate fluorescence images, a second fluorescence image whose corresponding cell distribution information satisfies the cell distribution condition from the candidate fluorescence images, including: For each candidate fluorescence picture among the multiple candidate fluorescence pictures, dividing the candidate fluorescence picture into multiple second image blocks according to a second preset number of divisions, and respectively calculating cell information of each second image block; Calculating cell distribution information of the candidate fluorescence image according to the cell information of the plurality of second image blocks; It is determined whether the cell distribution information of the candidate fluorescent image meets the cell distribution condition. If not, the candidate fluorescent image is eliminated to obtain a second fluorescent image whose corresponding cell distribution information in the candidate fluorescent image meets the cell distribution condition.
[0072] In a possible implementation, the model training module performs data enhancement processing on the second fluorescent image to obtain a third fluorescent image that meets the diversity condition, including: determining a target enhancement method for the second fluorescent image according to the cell distribution characteristics of the second fluorescent image; The second fluorescent image is enhanced according to the target enhancement method; wherein the target enhancement method includes at least one of the following: geometric transformation, color transformation, noise addition and blur processing; the geometric transformation includes at least one of the following: rotation, flipping and scaling.
[0073] In a possible implementation, the target data set includes a training data set and a validation data set; the model training module trains the original ANA recognition model using the target data set to obtain a trained ANA recognition model, including: Setting the hyperparameters of the ANA recognition model, and training the training parameters in the original ANA recognition model in a round-based manner using the training data set under the hyperparameters; the training parameters include parameters of the feature extraction network, the first task network, and the second task network; After each round of training, the ANA recognition model under the training parameters of this round is tested through the validation set. If the result of the validation data set is better than the output result of the previous round, the training parameters of the current round of the ANA recognition model are saved until a trained ANA recognition model is obtained.
[0074] The above-mentioned ANA detection device based on the multi-task network provided in the embodiment of the present application can simultaneously identify the ANA karyotype and titer through two task networks, and extract richer and more accurate image features for titer and karyotype identification by segmenting the image blocks, thereby improving the recognition accuracy.
[0075] like Fig. 9 As shown, an electronic device 900 provided in an embodiment of the present application includes: a processor 901, a memory 902 and a bus, the memory 902 stores machine-readable instructions executable by the processor 901, and when the electronic device is running, the processor 901 communicates with the memory 902 through the bus, and the processor 901 executes the machine-readable instructions to perform the steps of the ANA detection method based on the multi-task network as described above.
[0076] Specifically, the memory 902 and the processor 901 can be general-purpose memories and processors, which are not specifically limited here. When the processor 901 runs the computer program stored in the memory 902, the ANA detection method based on the multi-task network can be executed.
[0077] Corresponding to the above-mentioned ANA detection method based on a multi-task network, an embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above-mentioned ANA detection method based on a multi-task network are executed.
[0078] The above-mentioned electronic device and storage medium provided in real time by the present application can more accurately identify image features and efficiently and accurately complete the prediction of karyotype and titer.
[0079] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the system and device described above can refer to the corresponding process in the method embodiment, and will not be repeated in this application. In the several embodiments provided in this application, it should be understood that the disclosed system, device and method can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.
[0080] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0081] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.
[0082] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard drives, ROM, RAM, magnetic disks, or optical disks.
[0083] The above are only specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
Claims
1. An ANA detection method based on a multi-task network, characterized in that: The method comprises: Acquire a first fluorescent image of the target user including antinuclear antibodies ANA; Dividing the first fluorescent picture into a plurality of first image blocks according to a first preset number of divisions, and determining a first image feature of the first fluorescent picture according to sub-image features of the plurality of first image blocks; The first image feature of the first fluorescence image is respectively calculated and processed by the first task network and the second task network of the pre-trained ANA recognition model, and the karyotype classification result of the first fluorescence image corresponding to the first task network and the titer classification result corresponding to the second task network are output; An ANA test result for the target user is generated according to the karyotype classification result and the titer classification result of the first fluorescent image.
2. The ANA detection method based on a multi-task network according to claim 1, characterized in that: The determining the first image feature of the first fluorescent picture according to the sub-image features of the plurality of first image blocks comprises: The multiple first image blocks are input into the ANA recognition model, the sub-image features of each first image block are extracted through the feature extraction network of the ANA recognition model, and the sub-image features of the multiple first image blocks are spliced to obtain the first image features of the first fluorescent image.
3. The ANA detection method based on a multi-task network according to claim 2, characterized in that: The first task network and the second task network include the same network architecture; the network architecture of the first task network corresponds to the output karyotype classification result; The network architecture of the second task network corresponds to the output titer classification result; The network architecture includes: a linear layer, a nominal layer, a multi-head attention mechanism layer, a convolutional layer, a nominal layer, a multi-head attention mechanism layer and a multi-layer perceptron connected in sequence.
4. The ANA detection method based on a multi-task network according to any one of claims 1 to 3, characterized in that: The ANA recognition model is trained by the following method: Preprocessing candidate fluorescence images including a plurality of different karyotypes to obtain training fluorescence images, and constructing a target data set based on the training fluorescence images; Pre-constructing an original ANA recognition model, the original ANA recognition model comprising a feature extraction network and a first task processing network and a second task processing network respectively connected to the feature extraction network; The original ANA recognition model is trained using the target data set to obtain a trained ANA recognition model.
5. The ANA detection method based on a multi-task network according to claim 4, characterized in that: The preprocessing of the candidate fluorescence images including a plurality of different karyotypes to obtain training fluorescence images includes: According to the cell distribution information in each of the candidate fluorescent images, selecting a second fluorescent image whose corresponding cell distribution information satisfies a cell distribution condition from the candidate fluorescent images; the cell distribution condition includes that at least one of the number of cells, the cell distribution characteristics and the cell clarity satisfies a corresponding preset threshold; By performing data enhancement processing on the second fluorescence image, a third fluorescence image that meets the diversity condition is obtained, and the third fluorescence image is normalized to obtain a training fluorescence image; wherein the diversity condition includes that the number of samples reaches a preset threshold and the cell distribution characteristics are different under the cell distribution condition.
6. The ANA detection method based on a multi-task network according to claim 5, characterized in that: The selecting, according to the cell distribution information in each of the candidate fluorescence images, a second fluorescence image whose corresponding cell distribution information satisfies the cell distribution condition from the candidate fluorescence images comprises: For each candidate fluorescence picture among the multiple candidate fluorescence pictures, dividing the candidate fluorescence picture into multiple second image blocks according to a second preset number of divisions, and respectively calculating cell information of each of the second image blocks; Calculating cell distribution information of the candidate fluorescence image according to the cell information of the plurality of second image blocks; It is determined whether the cell distribution information of the candidate fluorescent image meets the cell distribution condition. If not, the candidate fluorescent image is eliminated to obtain a second fluorescent image whose corresponding cell distribution information in the candidate fluorescent image meets the cell distribution condition.
7. The ANA detection method based on a multi-task network according to claim 5, characterized in that: The step of performing data enhancement processing on the second fluorescent image to obtain a third fluorescent image that meets a diversity condition includes: determining a target enhancement method for the second fluorescent image according to the cell distribution characteristics of the second fluorescent image; The second fluorescent image is enhanced according to the target enhancement method; wherein the target enhancement method includes at least one of the following: geometric transformation, color transformation, noise addition and blur processing; the geometric transformation includes at least one of the following: rotation, flipping and scaling.
8. The ANA detection method based on a multi-task network according to claim 4, characterized in that: The target data set includes a training data set and a verification data set; The step of training the original ANA recognition model by using the target data set to obtain a trained ANA recognition model includes: Setting the hyperparameters of the ANA recognition model, and training the training parameters in the original ANA recognition model in a round-based manner using the training data set under the hyperparameters; the training parameters include parameters of the feature extraction network, the first task network, and the second task network; After each round of training, the ANA recognition model under the training parameters of this round is tested through the validation set. If the result of the validation data set is better than the output result of the previous round, the training parameters of the current round of the ANA recognition model are saved until a trained ANA recognition model is obtained.
9. An ANA detection device based on a multi-task network, characterized in that: The device comprises: An acquisition module is used to acquire a first fluorescent image of a target user including antinuclear antibodies ANA; A segmentation module, configured to segment the first fluorescent image into a plurality of first image blocks according to a first preset segmentation quantity; A processing module, configured to determine a first image feature of the first fluorescence image according to the sub-image features of the plurality of first image blocks; respectively calculate and process the first image feature of the first fluorescence image through a first task network and a second task network of a pre-trained ANA recognition model, and output a karyotype classification result of the first fluorescence image corresponding to the first task network and a titer classification result corresponding to the second task network; A generation module is used to generate an ANA test result for the target user according to the karyotype classification result and the titer classification result of the first fluorescent image.
10. An electronic device, characterized in that: include: A processor, a storage medium and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the storage medium through the bus, and the processor executes the machine-readable instructions to perform the steps of the ANA detection method based on the multi-task network as described in any one of claims 1 to 8.
11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the ANA detection method based on a multi-task network as claimed in any one of claims 1 to 8 are executed.
Citation Information
Patent Citations
ANA fluorescent sheet automatic identification method based on machine learning and deep learning
CN109858428A
Antinuclear antibody fluorescence picture interpretation method and device, equipment and storage medium
CN115311216A
Bridge crack disease detection method and device based on crack self-segmentation model
CN117809082A
Oral operation real-time navigation method based on CT and laser oral scanning
CN117958970A
Cited By
Image recognition and detection method, image information detection method and image processing system
WO2026199807A1
