Method, device, equipment and storage medium for interpreting fluorescence images of antinuclear antibodies

Abstract

The embodiment of the present invention discloses a method and apparatus, device and storage medium for interpreting fluorescent images of antinuclear antibodies. The method includes: based on a preset image size, performing slider segmentation on N fluorescent images to obtain a number of segmented image blocks; stacking the image blocks vertically according to the channel dimension of the image channel, and inputting the stacked image blocks into a preset ANA recognition model for prediction processing to determine the predicted label of each fluorescent image. The labels include karyotype and titer. The ANA recognition model is obtained by inputting training sample data that conforms to the target user's ANA image interpretation habits into a pre-trained ANA recognition model for transfer learning; performing result inspection processing based on the predicted label of each fluorescent image to determine the target interpretation result of the sample to be identified. In this way, the ANA recognition model that conforms to the target user's interpretation habits can be used to interpret ANA fluorescent images, with higher interpretation efficiency and more accurate results.

Description

Technical Field

[0001] The present invention relates to the technical field of antinuclear antibody testing, and in particular to a method and device for interpreting fluorescence images of antinuclear antibodies, equipment and storage medium. Background Art

[0002] Currently, some manufacturers' instruments include the ability to automatically interpret fluorescent images, such as the AESKU Helios fully automated immunofluorescence instrument and the Europattern fully automated immunofluorescence instrument. While these manufacturers' instruments can automatically complete the entire indirect immunofluorescence process and interpret fluorescent images of antinuclear antibodies (ANA), they are only capable of interpreting certain karyotypes and are unable to interpret mixed karyotypes. Furthermore, the interpretation criteria of these devices are immutable, making it impossible to calibrate the detection model based on the actual interpretation results or interpretation habits of the physician in their daily work. These manufacturers' devices can only recognize some common karyotypes and cannot self-calibrate the model based on historical interpretation data.

[0003] Because the model interpretation is too rigid and does not conform to the habits of the target users, most laboratory physicians still use manual interpretation, which has low interpretation efficiency. Therefore, there is an urgent need for a means to improve the efficiency of ANA interpretation. Summary of the Invention

[0004] The main purpose of the present invention is to provide a method and device, equipment and storage medium for interpreting fluorescent images of antinuclear antibodies, which can solve the problem of low interpretation efficiency in the prior art.

[0005] To achieve the above objectives, the present invention provides a method for interpreting a fluorescent image of an antinuclear antibody in a first aspect, the method comprising:

[0006] Obtain N fluorescence images of the target user's sample to be identified, obtained through ANA testing, where N is a positive integer;

[0007] Based on a preset image size, performing slider segmentation on the N fluorescence images to obtain a plurality of segmented image blocks;

[0008] The image blocks are stacked vertically according to the channel dimension of the image channel, and the stacked image blocks are input into a preset ANA recognition model for prediction processing to determine a predicted label for each of the fluorescence images, wherein the label includes a karyotype and a titer. The ANA recognition model is obtained by inputting training sample data that conforms to the target user's ANA image interpretation habits into a pre-trained ANA recognition model for transfer learning;

[0009] The result verification process is performed according to the predicted label of each of the fluorescent images to determine the target interpretation result of the sample to be identified.

[0010] In one possible implementation, transfer learning is performed as follows, including:

[0011] Obtaining a target error database for the target user, the target error database including correspondences between a first number of first images and first labels; the target error database being a training sample database consistent with the target user's ANA image interpretation habits, the first images being ANA fluorescence images whose predicted labels differ from their actual labels, and the first labels being primary labels selected using the actual labels of the first images;

[0012] Determining a migration training sample dataset according to the error database and a preset migration training sample selection rule, wherein the migration training sample dataset includes a second number of correspondences between second images and second labels, where the second number is greater than the first number;

[0013] The migration training sample data set is input into the pre-trained ANA recognition model for migration training to obtain a trained ANA recognition model.

[0014] In a feasible implementation, the obtaining of the target error database of the target user includes:

[0015] Determining the karyotype type of the actual label of each first image in a preset error database;

[0016] If the karyotype type is a single karyotype or negative, the actual label of the first image is used as the first label of the first image;

[0017] If the karyotype type is a composite karyotype, the label with the highest titer included in the actual labels of the first image is used as the first label of the first image;

[0018] If the karyotype type is a composite karyotype, and the actual labels of the first image of the composite karyotype include two or more labels with the highest titers, then randomly select one label with the highest titer from the two or more labels with the highest titer as the first label of the first image;

[0019] The target error database is obtained based on the first image and the first label.

[0020] In a feasible implementation, determining the migration training sample data set according to the error database and a preset migration training sample selection rule includes:

[0021] Determining the proportion of each first label in the error database;

[0022] Randomly extracting the first image and the first label from the error database according to the proportion to obtain candidate first images and first labels;

[0023] If the image and label in the transfer training sample dataset are the same as the candidate first image and first label, performing random data augmentation processing on the candidate first image and first label, and adding the processed candidate first image and first label to the transfer training sample dataset;

[0024] If the images and labels in the transfer training sample dataset do not overlap with the candidate first images and first labels, directly adding the candidate first images and first labels to the transfer training sample dataset;

[0025] Repeating the step of randomly extracting the first image and the first label from the error database according to the proportion to obtain candidate first images and first labels until the number of samples in the migration training sample dataset reaches a third number, which is less than the second number;

[0026] When the number of samples in the migration training sample data set reaches a third number, traversing the second labels in the migration training sample data set to determine label category data for each label in the migration training sample data set;

[0027] Determine the relationship between the total number of each type of label category data and the first numerical threshold;

[0028] When there is first label category data whose total number is less than a first threshold value, a third label and a third image of the same label category as the first label category data are randomly selected from a preset sample database and added to the migration training sample dataset until the total number of the first label category data equals the first threshold value;

[0029] When there is second label category data whose total number is greater than the first threshold, randomly deleting the correspondence between the fourth image and the fourth label in the second label category data until the total number of the second label category data is equal to the first threshold;

[0030] Repeat the step of determining the relationship between the total number of the label category data and the first number threshold until the number of samples in the migration training sample data set reaches a second number, thereby obtaining a final migration training sample data set.

[0031] In a feasible implementation, if N is greater than 1, then performing result verification processing based on the predicted label of each of the fluorescent images to determine the target interpretation result of the sample to be identified includes:

[0032] Classifying the predicted labels of the N fluorescence images to obtain a category set for each label category, wherein the category set includes a correspondence between the label category and the number of images;

[0033] Determine the maximum number of images in all the category sets;

[0034] Determine a target quantity threshold for the N fluorescence images using N and a preset quantity threshold algorithm, wherein the quantity threshold algorithm is to obtain the target quantity threshold by adding 1 to the product of N and one-half;

[0035] If the maximum value is greater than or equal to the target quantity threshold, the predicted label corresponding to the maximum value is used as the target interpretation result of the sample to be identified;

[0036] If the maximum value is less than the target quantity threshold, a warning indication is output, where the warning indication indicates that the prediction result has failed and manual interpretation is recommended.

[0037] In a feasible implementation, the step of using the predicted label corresponding to the maximum value as the interpretation label of the target interpretation result of the sample to be identified further includes:

[0038] Determining whether the interpretation label of the target interpretation result is negative;

[0039] When the judgment label is negative, the pixel value of the R channel of each fluorescence image of the ANA test is obtained;

[0040] Determining the average fluorescence intensity of each of the fluorescence images according to the pixel values ​​and a preset average value algorithm;

[0041] Performing a single-sample T test using the average fluorescence intensity of the N fluorescence images and the preset average fluorescence intensity of negative data to determine the test value of each fluorescence image;

[0042] If the test value of each of the fluorescent images is less than the preset confidence interval percentage, the judgment label is determined to be a true negative;

[0043] If the inspection value of any of the fluorescent images is greater than or equal to the preset confidence interval percentage, the output warning indication is executed, and the warning indication indicates that the identification label has failed and manual judgment is recommended.

[0044] To achieve the above object, the second aspect of the present invention provides a method for training an ANA recognition model, the method comprising:

[0045] Obtaining an ANA training sample set, where the ANA training sample set includes a correspondence between a plurality of fifth images and fifth labels;

[0046] Performing a slider segmentation process on the N fifth images based on a preset image size to obtain a plurality of segmented sample image blocks;

[0047] The sample image blocks are stacked vertically according to the channel dimension of the image channel, and the stacked sample image blocks and the fifth label are input into a deep neural network with a channel attention mechanism for model training until the model converges to obtain a pre-trained ANA recognition model. The loss function of the model training is the cross entropy of the fifth label and the predicted label. The pre-trained ANA recognition model is used to input training sample data that conforms to the target user's ANA picture interpretation habits into the pre-trained ANA recognition model for transfer learning to obtain the ANA recognition model as described in the first aspect and any feasible implementation method.

[0048] To achieve the above-mentioned object, the third aspect of the present invention provides a device for identifying a fluorescent image of an antinuclear antibody, the device comprising:

[0049] Data acquisition module: used to obtain N fluorescence images of the target user's sample to be identified, where N is a positive integer;

[0050] Segmentation processing module: used for performing slider segmentation on the N fluorescent images based on a preset image size to obtain a plurality of segmented image blocks;

[0051] Result prediction module: used to stack the image blocks vertically according to the channel dimension of the image channel, and input the stacked image blocks into a preset ANA recognition model for prediction processing to determine the predicted label of each fluorescent image, wherein the label includes karyotype and titer. The ANA recognition model is obtained by inputting training sample data that conforms to the target user's ANA image interpretation habits into a pre-trained ANA recognition model for transfer learning;

[0052] Result determination module: used to perform result verification processing based on the predicted label of each of the fluorescent images to determine the target interpretation result of the sample to be identified.

[0053] To achieve the above-mentioned objectives, the fourth aspect of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor performs the steps of the method for interpreting fluorescent images of antinuclear antibodies as described in the first aspect and any feasible implementation, or the steps of the method for training an ANA recognition model as described in the second aspect.

[0054] To achieve the above-mentioned objectives, the fifth aspect of the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the computer program is executed by the processor to perform the steps of the method for interpreting fluorescent images of antinuclear antibodies as described in the first aspect and any feasible implementation, or the steps of the method for training an ANA recognition model as described in the second aspect.

[0055] The embodiments of the present invention have the following beneficial effects:

[0056] The present invention provides a method for interpreting fluorescent images of antinuclear antibodies, the method comprising: obtaining N fluorescent images of ANA tests of a target user's sample to be identified, where N is a positive integer; performing slider segmentation on the N fluorescent images based on a preset image size to obtain a plurality of segmented image blocks; stacking the image blocks vertically according to the channel dimensions of the image channels, and inputting the stacked image blocks into a preset ANA recognition model for prediction processing to determine the predicted label of each fluorescent image, the label including karyotype and titer, the ANA recognition model being obtained by inputting training sample data that conforms to the target user's ANA image interpretation habits into a pre-trained ANA recognition model for transfer learning; performing result verification processing based on the predicted label of each fluorescent image to determine the target interpretation result of the sample to be identified. In this way, the ANA recognition model that conforms to the target user's interpretation habits can be used to interpret ANA fluorescent images, with higher interpretation efficiency and more accurate results. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0058] in:

[0059] Figure 1 This is a flow chart of a method for interpreting a fluorescent image of an antinuclear antibody according to an embodiment of the present invention;

[0060] FIG2( a ) is a schematic diagram of a green fluorescence image according to an embodiment of the present invention;

[0061] FIG2( b ) is a schematic diagram of a label according to an embodiment of the present invention;

[0062] Figure 3 Flowchart of a training method for an ANA recognition model according to an embodiment of the present invention;

[0063] FIG4( a ) is a schematic diagram of a partial hierarchical structure of an ANA recognition model according to an embodiment of the present invention;

[0064] FIG4( b ) is a schematic diagram of another part of the hierarchical structure of an ANA recognition model according to an embodiment of the present invention;

[0065] FIG4( c ) is a schematic diagram of another part of the hierarchical structure of an ANA recognition model according to an embodiment of the present invention;

[0066] Figure 5 is another flow chart of a method for interpreting a fluorescent image of an antinuclear antibody according to an embodiment of the present invention;

[0067] Figure 6 This is a structural block diagram of a device for interpreting a fluorescent image of an antinuclear antibody according to an embodiment of the present invention;

[0068] Figure 7 1 is a structural block diagram of a training device for an ANA recognition model according to an embodiment of the present invention;

[0069] Figure 8 4 is a structural block diagram of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0070] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0071] The method described in this application can be applied to both terminals and servers. This embodiment uses the terminal as an example. The terminal can be a desktop terminal or a mobile terminal. The mobile terminal can be at least one of a mobile phone, a tablet computer, and a laptop computer. The server can be implemented as a standalone server or a server cluster consisting of multiple servers. This is not limited here.

[0072] See also Figure 1 , Figure 1 FIG. 1 is a flow chart of a method for interpreting a fluorescent image of an antinuclear antibody according to an embodiment of the present invention, Figure 1 The method comprises the following steps:

[0073] 101. Obtain N fluorescence images of the target user's sample to be identified, obtained through ANA testing;

[0074] It should be noted that the present application provides a method for interpreting fluorescent images of antinuclear antibodies. First, the interpretation object, that is, the fluorescent image of ANA detection, must be determined. Figure 2(a) is a schematic diagram of a green fluorescent image in an embodiment of the present invention, wherein the fluorescent image in Figure 2(a) is green. Specifically, N fluorescent images of ANA detection of the target user's sample to be identified are obtained. For example, the sample to be identified for ANA detection, such as a blood sample or specimen, is photographed by the Helios automatic photography system, and the fluorescent image of ANA detection corresponding to the sample to be identified for ANA detection is collected and obtained. Multiple images can be taken, so the fluorescent images of ANA detection can be N, wherein N is a positive integer, and the value of N can be 3, that is, 3 fluorescent images of ANA detection are taken for the same sample to be identified. The above is only an example and is not a specific limitation.

[0075] 102. Based on a preset image size, perform slider segmentation on the N fluorescent images to obtain a plurality of segmented image blocks;

[0076] 103. Stacking the image blocks vertically according to the channel dimension of the image channel, and inputting the stacked image blocks into a preset ANA recognition model for prediction processing to determine a predicted label for each of the fluorescence images;

[0077] Furthermore, the N fluorescent images obtained are subjected to slider segmentation processing. Specifically, based on a preset image size, such as 416*416, the N fluorescent images are subjected to slider segmentation to obtain several image blocks after segmentation. For example, the size of the image taken by the Helios automatic camera system is 2560*1920, and the image channel is RGB. Currently, the video memory of commonly used graphics cards is usually unable to support 2560*1920 size images for direct model training or subsequent karyotype and titer identification, so the fluorescent images will be preprocessed in this application to reduce the size of the images while maintaining the image features as much as possible. For example, by performing slider cutting on the fluorescent image, a 2560*1920 image is cut into 35 416*416 image blocks. Since the green fluorescent image of Figure 2(a) is identified, the G channel of the image can be selected and stacked, and finally 416*416*35 data is input to the neural network. It should be noted that dividing the fluorescence image into 35 image blocks at a size of 416*416 ensures that all features of the original fluorescence image are captured. Furthermore, when N is 3, 105 image blocks can be obtained. It should be noted that since slider cutting can cause repetitions between image blocks, vector calculations can be performed on the image blocks after they are obtained. The image blocks can then be stacked based on the vector calculation results, achieving vertical stacking with overlapping repetitive regions, which can then be input into the deep learning network for model training.

[0078] Among them, the model can be a deep artificial neural network built based on ResNet, which can recognize and output labels of fluorescent images, including nuclear type and titer. Exemplary, this application will classify 19 nuclear types, such as nuclear homogeneity type, nuclear granule type, nucleolus type, nuclear dot type, centromere type, nuclear membrane type, cytoplasmic granule type, cytoplasmic fiber type, nuclear dense spot type, PCNA type, centromere protein-F, Golgi body type, spindle fiber type, NuMA type, centrosome type, intermediate body, chromosome type, rod-loop Golgi body, and cell bridge type. Each karyotype has five titer classifications, such as 1:80, 1:160, 1:320, 1:640, and 1:1280. This means the model can identify 19 karyotypes and five titers, including a special titer that is negative. Ultimately, the model can identify 19 karyotypes, five titers, and a negative titer, for a total of 19*5+1=96 combinations. It is understood that each fluorescence image can have multiple karyotypes, meaning that each image label can include multiple karyotype-titer correspondences. The ANA recognition model is derived by transfer learning by inputting training sample data that aligns with the target user's ANA image interpretation habits into a pre-trained ANA recognition model. Because different target users have different habits, the training samples also vary. Therefore, the ANA recognition model in this implementation is unique for different target users, better suited to each user's interpretation habits.

[0079] Furthermore, the text of the label identified by the model can be converted into a one-hot encoding with a length of 114. This means that there are 19 karyotypes, each with 1 + 5 encoding bits, for a final encoding length of 19 * 6 = 114. The encoding rule is as follows: Each karyotype is sequentially assigned 6 bits [000000], with different bits corresponding to different titers. Then, based on the titer of the karyotype corresponding to the image, the encoding of the karyotype is assigned 1. If the image does not contain the karyotype, bit 1 (the first of the 6 bits of the karyotype) is assigned 1. For example, a negative karyotype can be represented as [100000]. If the karyotype is present, the corresponding position of the karyotype is assigned 1 based on the titer level. Exemplarily, the relationship between the 6 bits of each karyotype and the titer is as follows: if the titer is 1:80, assign 1 to position 2, such as [010000]; if the titer is 1:160, assign 1 to position 3, such as [001000]; if the titer is 1:320, assign 1 to position 4, such as [000100]; if the titer is 1:640, assign 1 to position 5, such as [000010]; if the titer is 1:1280, assign 1 to position 6, such as [000001]. Please refer to Figure 2(b), which is a schematic diagram of a label in an embodiment of the present invention. Figure 2(b) shows examples of four labels, wherein if the label is negative or a certain karyotype is not identified, then as shown in 201, position 1 of the karyotype is 1. Further, if the positions 1 of the 19 karyotypes in a row where 201 is located are all 1, the label of the fluorescence image is negative; if the label is nuclear homogeneity and the titer of nuclear homogeneity is 1:80, then as shown in 202, position 2 of nuclear homogeneity is 1; if the label is nuclear homogeneity and the titer of nuclear homogeneity is 1:160, then as shown in 203, position 3 of nuclear homogeneity is 1; if the label is nuclear homogeneity and the titer of nuclear homogeneity is 1:160, and the nuclear particle type and the titer of the nuclear particle type is 1:160, then as shown in 204, position 3 of nuclear homogeneity is 1, and as shown in 205, position 3 of the nuclear particle type is 1. It should be noted that the above is only an example and does not limit this embodiment. As the types of karyotypes and titers change, the length of the data or the form of label expression can be adaptively increased. Therefore, any variation without departing from the technical concept of this embodiment falls within the scope of this application.

[0080] 104. Perform result verification processing based on the predicted label of each of the fluorescent images to determine the target interpretation result of the sample to be identified.

[0081] Finally, to ensure the accuracy of the model's predictions, this embodiment verifies the predicted labels for each fluorescence image to determine the target interpretation of the sample to be identified. For example, this involves manually comparing the predicted labels with the manual results to determine if the predicted labels are correct, ultimately obtaining the target interpretation of the sample to be identified. This verification process further verifies and confirms the authenticity and accuracy of the predicted labels, ensuring that the final target interpretation results are more reliable and effective.

[0082] The present invention provides a method for interpreting fluorescent images of antinuclear antibodies, the method comprising: obtaining N fluorescent images of ANA tests of a target user's sample to be identified, where N is a positive integer; performing slider segmentation on the N fluorescent images based on a preset image size to obtain a plurality of segmented image blocks; stacking the image blocks vertically according to the channel dimensions of the image channels, and inputting the stacked image blocks into a preset ANA recognition model for prediction processing to determine the predicted label of each fluorescent image, the label including karyotype and titer, the ANA recognition model being obtained by inputting training sample data that conforms to the target user's ANA image interpretation habits into a pre-trained ANA recognition model for transfer learning; performing result verification processing based on the predicted label of each fluorescent image to determine the target interpretation result of the sample to be identified. In this way, the ANA recognition model that conforms to the target user's interpretation habits can be used to interpret ANA fluorescent images, with higher interpretation efficiency and more accurate results.

[0083] See also Figure 3 , Figure 3 FIG. 1 is a flow chart of a training method for an ANA recognition model according to an embodiment of the present invention, Figure 3 The method comprises the following steps:

[0084] 301. Obtain an ANA training sample set;

[0085] It is understandable that in order to realize the automatic recognition of fluorescence images, it is necessary to pre-train the ANA recognition model to realize the interpretation of fluorescence images. Model training requires training samples. Therefore, the ANA training sample set is first obtained. The ANA training sample set includes the correspondence between several fifth images and fifth labels; the image data of the training samples are derived from the ANA fluorescence images detected by the laboratory on a daily basis, and the labels of the images are derived from the interpretation of the images by qualified testers. There are 19 types of these labels in total, and the images are divided into 5 titers such as 1:80, 1:160, 1:320, 1:640, and 1:1280 according to the fluorescence intensity. Among them, the data volume of the ANA training sample set of this application can reach 200,000, all of which are obtained based on the laboratory's own large database, and 200,000 clinical data are used to train the AI ​​model (ANA recognition model). Among them, the first, second, third, fourth and fifth are only used for distinction, and do not specifically limit the technical features unless otherwise specified.

[0086] 302. Performing a slider segmentation process on the N fifth images based on a preset image size to obtain a plurality of segmented sample image blocks;

[0087] 303. Stack the sample image blocks vertically according to the channel dimension of the image channel, and input the stacked sample image blocks and the fifth label into a deep neural network with a channel attention mechanism for model training until the model converges to obtain a pre-trained ANA recognition model.

[0088] Furthermore, as in step 102, in order to meet the data processing requirements, slider segmentation processing is performed on the N pieces of the fifth picture based on the preset picture size to obtain a number of sample image blocks after segmentation, and the model is trained based on the sample image blocks after segmentation. The content of step 302 can refer to the content of step 102 and will not be repeated here. Furthermore, the picture channel includes RGB, and the sample image blocks are stacked vertically according to the channel dimension of the picture channel, and the stacked sample image blocks and the fifth label are input into a deep neural network with a channel attention mechanism for model training until the model converges to obtain a pre-trained ANA recognition model. The loss function of the model training is the cross entropy of the fifth label and the predicted label, that is, the loss function is the cross entropy of the predicted result and the actual result. The pre-trained ANA recognition model is used to input the training sample data that meets the target user's ANA picture interpretation habits into the pre-trained ANA recognition model for transfer learning to obtain the following. Figure 1 ANA recognition model. Based on the aforementioned images and their corresponding labels, a deep learning artificial neural network is built. This neural network is based on ResNet and incorporates a channel attention mechanism. The pre-trained ANA recognition model is used as a pre-trained model and serves as the baseline model for target user calibration.

[0089] For example, in ResNet, the original data is added to the original data after the convolution operation, thereby alleviating the degradation problem of the neural network as the number of layers increases to a certain extent.

[0090] Two ResNet structures are used in this neural network: the downsized DownSampleRes and the regular-sized NormalRes. Their structures are shown in Figures 4(a) and 4(b). Figure 4(a) is a schematic diagram of a partial hierarchical structure of an ANA recognition model according to an embodiment of the present invention; Figure 4(b) is a schematic diagram of another partial hierarchical structure of an ANA recognition model according to an embodiment of the present invention. In Figures 4(a) and 4(b), Conv2D is a convolutional layer, filter is the number of output channels, kernel_size is the convolution kernel size, strides is the stride, BatchNormalization is batch normalization, ReLU is the activation function, ADD is addition, Input is input, Output is output, h is height, w is width, and c is depth.

[0091] Exemplarily, the channel attention mechanism means that the corresponding neural layer is used to link the channels of the stacked images using a fully connected layer to learn the features between the image channels. For details, please refer to Figure 4(c). Figure 4(c) is a schematic diagram of the hierarchical structure of another part of an ANA recognition model in an embodiment of the present invention, where FC is a fully connected layer and Scale is a normalization function.

[0092] Furthermore, the neural network output structure is as follows: After the ResNet is completed, the data is normalized again to form the final output data. The normalization process is as follows: the final data is fully connected 19 times (with 6 outputs and a Sigmoid activation function), and then these 19 data points are concatenated to form a 114-dimensional data set. As mentioned above, the loss function for model training is the cross entropy between the actual and predicted results.

[0093] Ultimately, based on the above neural network structure and laboratory data, an AI model capable of classifying ANA images was trained. The laboratory data was based on KingMed's historical ANA test records.

[0094] The present invention provides a training method for an ANA recognition model, the method comprising: obtaining an ANA training sample set, the ANA training sample set including a correspondence between a plurality of fifth images and a fifth label; performing a slider segmentation process on N fifth images based on a preset image size to obtain a plurality of segmented sample image blocks; stacking the sample image blocks vertically according to the channel dimension of the image channel, and inputting the stacked sample image blocks and the fifth label into a deep neural network with a channel attention mechanism for model training until the model converges, thereby obtaining a pre-trained ANA recognition model, wherein the loss function of the model training is the cross entropy of the fifth label and the predicted label, and the pre-trained ANA recognition model is used to input training sample data that meets the target user's ANA image interpretation habits into the pre-trained ANA recognition model for transfer learning, so as to obtain the following: Figure 1 ANA recognition model. Through the above method, we can use 200,000 clinical data to train the AI ​​model based on the laboratory's large database to obtain a pre-trained ANA recognition model. This pre-trained ANA recognition model is used as the pre-trained model, that is, as the benchmark model for target user calibration.

[0095] See also Figure 5 , Figure 5 FIG. 1 is another flow chart of a method for interpreting a fluorescent image of an antinuclear antibody according to an embodiment of the present invention. Figure 5 The method shown includes the following steps:

[0096] 501. Obtain N fluorescence images of the target user's sample to be identified, obtained through ANA testing;

[0097] Wherein, N is a positive integer.

[0098] 502. Based on a preset image size, perform slider segmentation on the N fluorescence images to obtain a plurality of segmented image blocks;

[0099] 503. Stack the image blocks vertically according to the channel dimension of the image channel, and input the stacked image blocks into a preset ANA recognition model for prediction processing to determine a predicted label for each of the fluorescence images;

[0100] Among them, the labels include karyotype and titer. The ANA recognition model is obtained by inputting training sample data that conforms to the target user's ANA image interpretation habits into the pre-trained ANA recognition model for transfer learning.

[0101] It should be noted that steps 501, 502 and 503 are Figure 1 The contents shown in steps 101, 102 and 103 are similar and will not be described here in detail to avoid repetition. For details, please refer to the aforementioned steps 101, 102 and 103.

[0102] It is understandable that most AI models currently on the market are pre-trained and cannot be dynamically fine-tuned based on the user's actual prediction results. In order to solve this problem, this system has added a transfer learning module, which can use the different data between the actual detection results and the AI ​​prediction results and some control data as materials for transfer training to fine-tune the AI ​​model. For example, in this embodiment, the pre-trained model will be trained twice based on the data of the different target users' own interpretation habits, that is, the training sample data that conforms to the target user's ANA picture interpretation habits is input into the pre-trained ANA recognition model for transfer learning to obtain different ANA recognition models corresponding to different target users. In a feasible implementation method, transfer learning is performed as follows, including steps P01, P02 and P03:

[0103] P01. Obtain the target error database of the target user;

[0104] Among them, the target error database includes the correspondence between the first number of first images and the first labels; the target error database is a training sample database that conforms to the target user's ANA image interpretation habits. Furthermore, the image with the actual label that conforms to the target user's interpretation habits can be input into a pre-trained model. When the predicted label output by the model is different from the actual label, the correspondence between the image and the actual label is recorded in the error database. Furthermore, the first image in the target error database is an ANA fluorescence image with a different predicted label and an actual label, and the first label can be a main label selected using the actual label of the first image. When the amount of data in the error database reaches the first number, the target error database is obtained.

[0105] In one feasible implementation, this embodiment configures a first label for the first image in the error database. The first label is a main label determined based on the actual label of the first image. Therefore, step P01 may include steps P11, P21, P31, P41, and P51:

[0106] P11, determining the karyotype type of the actual label of each first image in a preset error database;

[0107] It should be noted that the preset error database is the error database before the first label is configured. Furthermore, the preset error database includes the correspondence between first images and actual labels. Furthermore, the karyotype type of the actual label of each first image in the preset error database is determined. Karyotype types include but are not limited to single karyotype, composite karyotype, or negative. Furthermore, in this embodiment, a first label is configured for each first image based on the karyotype type of the actual label of each first image.

[0108] P21. If the karyotype type is a single karyotype or negative, use the actual label of the first image as the first label of the first image;

[0109] Among them, if the karyotype type of the actual label of the first image is a single karyotype or negative (refer to 201, 202 and 203 in Figure 2(b)), the actual label of the first image is used as the first label of the first image to obtain the first label corresponding to the first image.

[0110] P31. If the karyotype type is a composite karyotype, use the label with the highest titer included in the actual labels of the first image as the first label of the first image;

[0111] If the karyotype is a composite karyotype (see 204 and 205 in FIG. 2( b )), the label with the highest titer included in the actual labels of the first image is used as the first label of the first image to obtain the first label corresponding to the first image.

[0112] P41. If the karyotype type is a composite karyotype, and the actual labels of the first image of the composite karyotype include two or more labels with the highest titers, randomly select one of the two or more labels with the highest titers as the first label of the first image;

[0113] If the karyotype is a composite karyotype, and the actual labels for the first image of the composite karyotype include two or more labels with the highest titers (referring to 204 and 205 in Figure 2(b)), then one of the two or more labels with the highest titers is randomly selected as the first label for the first image, obtaining the first label corresponding to the first image. For example, if both karyotypes 204 and 205 in Figure 2(b) have a titer of 1:160, both can be considered the highest titer, thus meeting the condition that the actual labels include two or more labels with the highest titers. Therefore, either the homogeneous nuclear type 1:160 or the granular nuclear type 1:160 is randomly selected as the first label corresponding to the first image.

[0114] P51. Obtain the target error database based on the first image and the first label.

[0115] Furthermore, first labels of a first number of first images are obtained to obtain a target error database.

[0116] Exemplarily, steps P11, P21, P31, P41 and P51:

[0117] (1) Select the data that are different from the recent actual detection results (actual labels) and the AI ​​prediction results (predicted labels) as the preset error database (Dataset miss), Dataset miss The amount of data must reach 250 cases (the first quantity).

[0118] (2) Give Dataset miss The first image in the image is given the first label (Label main ), used for random sampling of data later. Label main Assignment rule: If the actual result of the first image is a single karyotype or negative, the label of the image main is the actual test result; if the actual result of the first image is a composite karyotype, the label of the image main The karyotype with the highest titer among the composite karyotypes. If there are two or more karyotypes with the highest titer, a combination of the karyotype and the highest titer is randomly selected as the label for the image. main .

[0119] P02. Determine a transfer training sample data set based on the target error database and a preset transfer training sample selection rule;

[0120] Furthermore, after obtaining the above-mentioned target error database, a transfer training sample data set can be obtained through the target error database. The transfer training sample data set is used for transfer learning of the pre-trained ANA recognition model to obtain different ANA recognition models for different target users. Specifically, the transfer training sample data set is determined based on the target error database and the preset transfer training sample selection rules, wherein the transfer training sample selection rules are used to select training samples for transfer learning, including but not limited to data expansion rules, etc., wherein the transfer training sample data set includes a second number of correspondences between second images and second labels, and the second number is greater than the first number.

[0121] In one feasible implementation, step P02 may include steps P12, P22, P32, P42, P52, P62, P72, P82, P92, and P102:

[0122] P12. Determine the proportion of each first label in the error database;

[0123] It should be noted that the members of the migration training sample data set can be obtained by random sampling, and this embodiment performs random sampling according to the proportion, so the proportion of each first label in the error database is first determined, such as: core homogeneity 1:80, core homogeneity 1:80, core homogeneity 1:80, core homogeneity 1:160, core homogeneity 1:160, core particle type 1:160, core homogeneity 1:320, etc. Further, the proportion relationship is [(core homogeneity 1:80): (core homogeneity 1:160): (core particle type 1:160): (core homogeneity 1:320)] = [3:2:1:1].

[0124] P22. Randomly extract the first image and the first label from the error database according to the proportion to obtain candidate first images and first labels;

[0125] Furthermore, after determining the proportion, random extraction can be performed, wherein the first image and the first label in the error database are randomly extracted according to the proportion to obtain candidate first images and first labels. That is, if a first image with a first label of "nuclear homogeneity 1:320" is randomly extracted, then according to the above proportion, it is necessary to randomly extract 3 first images with a first label of nuclear homogeneity 1:80, 2 first images with a first label of nuclear homogeneity 1:160, and 1 first image with a first label of nuclear particle type 1:160. That is, the candidate first image and first label are the first image and first label randomly extracted from the error database. It is further necessary to determine whether the candidate first image and first label are repeated with the image and label in the training sample dataset. If so, execute step P32; if not, execute step P42.

[0126] P32. If the image and label in the transfer training sample dataset overlap with the candidate first image and first label, perform random data augmentation on the candidate first image and first label, and add the processed candidate first image and first label to the transfer training sample dataset;

[0127] If the candidate first image and first label are already in the transferred training sample dataset, this indicates that the image and label in the transferred training sample dataset overlap with the candidate first image and first label, i.e., the samples are duplicated. Therefore, the candidate first image and first label need to be randomly augmented before being added to the transferred training sample dataset to obtain the second image and second label. Random data augmentation includes, but is not limited to, image processing methods such as vertical mirroring, horizontal mirroring, and 180° rotation.

[0128] P42. If the image and label in the transfer training sample dataset do not overlap with the candidate first image and first label, directly add the candidate first image and first label to the transfer training sample dataset;

[0129] It is understandable that if there is no repetition, the candidate first picture and the first label can be directly put into the migration training sample dataset to obtain the second picture and the second label.

[0130] P52. Repeat the step of randomly selecting the first images and first labels from the error database according to the proportions to obtain candidate first images and first labels until the number of samples in the migration training sample dataset reaches a third number, which is less than the second number.

[0131] Furthermore, step P22 is repeated to continuously expand the number of samples until the number of samples in the migrated training sample dataset reaches a third number, which is less than the second number. For example, the third number may be 4800, and the second number may be 9600. After the number of samples in the migrated training sample dataset reaches the third number, another method is used to expand the data volume, as follows.

[0132] P62. When the number of samples in the migration training sample dataset reaches a third number, traverse the second labels in the migration training sample dataset to determine label category data for each label in the migration training sample dataset;

[0133] Furthermore, after the number of samples in the migration training sample data set reaches the third number, the second label in the migration training sample data set is traversed to determine the label category data of each label in the migration training sample data set. Exemplarily, the label category data is used to indicate the number of label categories for each label, such as 99 second images with the second label of nuclear homogeneity 1:80, 101 second images with the second label of nuclear homogeneity 1:160, 100 second images with the second label of nuclear homogeneity 1:160, and 110 second images with the second label of nuclear homogeneity 1:320; then the number of label categories for nuclear homogeneity 1:80 is 99, the number of label categories for nuclear homogeneity 1:160 is 101, the number of label categories for nuclear particle type 1:160 is 100, and the number of label categories for nuclear homogeneity 1:320 is 100.

[0134] P72. Determine the relationship between the total number of each type of label category data and the first numerical threshold;

[0135] Furthermore, the total number of label category data, that is, the total number of second images of the target label category, is compared with the first threshold value. If the total number is less than the first threshold value, step P82 is executed; if the total number is greater than the first threshold value, step P92 is executed.

[0136] P82. When there is data of the first label category whose total number is less than a first threshold, randomly select a third label and a third image of the same label category as the first label category data from a preset sample database and add them to the migration training sample dataset until the total number of data of the first label category equals the first threshold;

[0137] Among them, when there is first label category data with a total number less than the first numerical threshold, such as 99 second pictures with the second label of nuclear homogeneity 1:80 in the above example, and the first numerical threshold is 100, then 99<100, and the first label category data is "99 second pictures with the second label of nuclear homogeneity 1:80", then the number of nuclear homogeneity 1:80 needs to be supplemented to 100, so the third label and the third picture of the same label category as the first label category data are randomly selected from the preset sample database and added to the migration training sample dataset. The preset sample database is a public sample database, such as a laboratory inspection database, that is, a third picture with the third label of "nuclear homogeneity 1:80" is randomly selected from the laboratory inspection database and added to the migration training sample dataset.

[0138] P92. When there is data of the second label category whose total number is greater than the first threshold, randomly delete the correspondence between the fourth image and the fourth label in the second label category data until the total number of data of the second label category equals the first threshold;

[0139] When there is data of the second label category whose total number is greater than the first threshold, for example, in the above example, there are 110 second images with the second label of nuclear homogeneity 1:320, and the first threshold is 100, then 110>100, and the second label category data is "110 second images with the second label of nuclear homogeneity 1:320." Therefore, the number of nuclear homogeneity 1:320 images needs to be reduced to 100. Therefore, the correspondence between the fourth image and the fourth label in the second label category data is randomly deleted until the total number of second label category data equals the first threshold. In other words, 10 fourth images with the fourth label of nuclear homogeneity 1:320 are deleted from the second label category data, so that the number of second label category data is 100.

[0140] P102. Repeat the step of determining the relationship between the total number of the label category data and the first number threshold until the number of samples in the migration training sample dataset reaches a second number, thereby obtaining a final migration training sample dataset.

[0141] The above process is repeated continuously, i.e., the step of determining the relationship between the total number of label category data and the first threshold value is repeated until the number of samples in the migrated training sample dataset reaches the second number, thereby obtaining the final migrated training sample dataset. The final number of samples in the migrated training sample dataset reaches the required second number, which may be 9600.

[0142] Illustratively, steps P12, P22, P32, P42, P52, P62, P72, P82, P92, and P102 include the following:

[0143] (1) Obtain the target error database (Datase tmiss )After that, define a dataset (Dataset retrain ), which is the final dataset used for training, namely the migration training sample dataset, Dataset retrain The data is obtained as follows: According to the Datase tmiss The first label in each main The proportion of randomly selected Datase tmiss Data to Dataset retrain If the randomly selected data has been included in the Dataset retrain If the image is repeated, the image will be randomly amplified (including vertical mirroring, horizontal mirroring, and 180° rotation) and then included in the Dataset. retrain Until Dataset retrain The amount of data reaches 4800 (the third number).

[0144] Then iterate over each Label main Type, determine the Dataset retrain Whether the amount of data of this type (label category data) exceeds 100 cases (the first numerical threshold). If it exceeds 100 cases, the data of this type will be randomly removed until the amount of data of this type reaches 100 cases. If it is less than 100 cases, the corresponding category of data will be randomly selected from the laboratory's (Golden Domain) big data pool and included in the Dataset. retrain Final Dataset retrain There should be 9600 pieces of data (the second quantity).

[0145] P03. Input the migration training sample data set into the pre-trained ANA recognition model for migration training to obtain a trained ANA recognition model.

[0146] Finally, the model is trained using the data in the migration training sample dataset (Datasetretrain) according to the training method of the aforementioned pre-trained ANA recognition model, that is, the migration training sample dataset is input into the pre-trained ANA recognition model for migration training, and then the original AI model is replaced to obtain the trained ANA recognition model.

[0147] Furthermore, different regular expression templates are provided based on the project, converting existing results in the target user database into corresponding standard result data. For example, for antinuclear antibodies, this would convert "granularity type 1:80" in the user database into "nuclear granularity type, 80." Currently, many common result data are covered, and new regular expressions can be added based on actual conditions to achieve standardization.

[0148] Furthermore, in this embodiment, in order to ensure the accuracy of the results, the same sample to be identified or multiple photos are taken, so N is greater than 1, and then the result verification processing is performed based on the predicted label of each fluorescent image to determine the target interpretation result of the sample to be identified, including the following steps 504, 505 and 506:

[0149] 504. Classify the predicted labels of the N fluorescence images to obtain a category set for each label category;

[0150] Among them, after using the ANA recognition model to identify the fluorescence image of the sample to be identified, the predicted label of the fluorescence image can be obtained. The predicted label can be the sum of the predicted labels of the image blocks corresponding to each fluorescence image. Then, the predicted labels of N fluorescence images are classified to obtain a category set for each label category, that is, the same label categories are clustered to obtain a category set for each label category. The category set includes the correspondence between the label category and the number of images.

[0151] 505. Determine the maximum number of images in all the category sets;

[0152] Furthermore, this embodiment verifies the accuracy and authenticity of the results by the majority rule. Therefore, the maximum number of images in the category set is determined. This maximum number is also the most labeled result in the prediction result, which represents the majority.

[0153] 506. Determine a target quantity threshold of the N fluorescence images using N and a preset quantity threshold algorithm;

[0154] Among them, since the minority obeys the majority when there are two or more maximum values, simply obeying the majority will lead to errors in the results. Therefore, the present application uses N and a preset quantity threshold algorithm to determine the target quantity threshold of N fluorescence images, wherein the quantity threshold algorithm is to add 1 to the product of N and one-half to obtain the target quantity threshold. Exemplarily, when N is 3, the target quantity threshold is 2, that is, (3*0.5)+1=1.5, wherein the number of images is an integer, so the calculated result is rounded to the target quantity threshold of 2. Then, if the maximum value is greater than or equal to the target quantity threshold, step 507 is executed; if the maximum value is less than the target quantity threshold, step 508 is executed.

[0155] 507. If the maximum value is greater than or equal to the target quantity threshold, the predicted label corresponding to the maximum value is used as the target interpretation result of the sample to be identified;

[0156] Furthermore, if the maximum value is greater than or equal to the target number threshold, it means that the maximum value is more than half of the actual total number. Therefore, the result of the minority obeys the majority is error-free. Therefore, the predicted label corresponding to the maximum value of the label category is used as the target judgment result of the sample to be identified.

[0157] For example, the Helios system will take three pictures of a sample to be identified. Generally, these three pictures should all belong to the same type, but in actual AI interpretation, the three pictures may be given different types, for example: nuclear granular type 1:80, nuclear granular type 1:80, nuclear granular type 1:160. In this warning module, the following three situations will be handled accordingly: (1) If the predicted labels of the three pictures are consistent, the corresponding results will be output directly; (2) If the predicted labels of two pictures are consistent but the predicted label of the remaining picture is inconsistent, the results of the two consistent pictures will be output; (3) If the predicted labels of the three pictures are different, a warning prompt will be issued and the case will be handed over to manual interpretation.

[0158] Furthermore, in the actual testing process, negative data accounts for about 50%. Generally, the testing physician expects that the negative images interpreted by the AI ​​are true negatives. In order to prevent the AI ​​model from interpreting positive images as negative, after the AI ​​model completes the interpretation, some conventional methods will be used to warn the AI ​​interpretation results. That is, the predicted label corresponding to the maximum value will be used as the interpretation label of the target interpretation result of the sample to be identified, and then it also includes L01, L02, L03, L04, L05 and L06:

[0159] L01. Determine whether the interpretation label of the target interpretation result is negative;

[0160] It can be understood that, referring to Figure 2(b), the determination of whether the label is negative is based on the label. Therefore, the determination of whether the target interpretation result is negative is based on the interpretation label, which is the predicted label corresponding to the maximum value. If it is negative, step L02 is executed to determine whether it is a true negative to avoid false positives. If it is not negative, the final result can be directly output, that is, the system will finally output the predicted result of the image. The inspection personnel can manually review the results of the AI ​​interpretation or trust the results of the AI ​​interpretation and issue a report.

[0161] L02. When the judgment label is negative, obtaining the pixel value of the R channel of each fluorescent image of the ANA test;

[0162] L03. Determine the average fluorescence intensity of each of the fluorescence images based on the pixel values ​​and a preset average value algorithm;

[0163] L04. Perform a single-sample T test using the average fluorescence intensity of the N fluorescence images and the preset average fluorescence intensity of negative data to determine the test value of each fluorescence image;

[0164] L05. If the test value of each of the fluorescent images is less than the preset confidence interval percentage, the interpretation label is determined to be a true negative;

[0165] L06. If the inspection value of any of the fluorescent images is greater than or equal to the preset confidence interval percentage, the output warning indication is executed, and the warning indication indicates that the identification label has failed and manual judgment is recommended.

[0166] In this system, the mean fluorescence intensity refers to the average of the G channel pixel values ​​in an image. Generally speaking, the lower the titer of the image, the lower the mean fluorescence intensity, and there is a positive correlation between the two. Therefore, when the label is judged as negative, the pixel value of the R channel of each fluorescence image tested by ANA is obtained for subsequent calculation of the mean fluorescence intensity. In addition, this system calculates the mean fluorescence intensity of 200,000 images (sample data used for training the pre-trained ANA recognition model) and calculates the mean and standard deviation of the mean fluorescence intensity of the negative data. If three images pass the majority warning and are predicted as negative by the model, the mean fluorescence intensity of the three images is calculated based on the pixel value of the G channel. The single-sample T test is then performed on the mean fluorescence intensity of the negative data. If all three images are within the 95% of the one-sided confidence interval, a negative conclusion is output; otherwise, manual interpretation is performed. The 95% confidence interval is the preset confidence interval percentage.

[0167] After the above-mentioned AI prediction and early warning mechanism, the system will finally output the prediction results of the image. The inspectors can conduct manual review based on the results of AI interpretation or publish a report based on the results of AI interpretation.

[0168] 508. If the maximum value is less than the target quantity threshold, a warning indication is output, where the warning indication indicates that the prediction result has failed and manual interpretation is recommended.

[0169] That is, if the maximum value is less than the target quantity threshold, it means that the maximum value is less than half, and the result will be in error, so a warning indication is output. The warning indication means that the prediction result has not passed, and manual interpretation is recommended.

[0170] The present invention provides a method for interpreting fluorescence images of antinuclear antibodies, the method comprising: obtaining N fluorescence images of ANA detection of a target user's sample to be identified, where N is a positive integer; performing slider segmentation on the N fluorescence images based on a preset image size to obtain a plurality of segmented image blocks; stacking the image blocks vertically according to the channel dimension of the image channel, and inputting the stacked image blocks into a preset ANA recognition model for prediction processing to determine the predicted label of each fluorescence image; classifying the predicted labels of the N fluorescence images to obtain a category set of each label category; determining the maximum number of images in all category sets; determining a target number threshold of the N fluorescence images using N and a preset number threshold algorithm; if the maximum value is greater than or equal to the target number threshold, using the predicted label corresponding to the maximum value as the target interpretation result of the sample to be identified; if the maximum value is less than the target number threshold, outputting a warning indication, the warning indication being that the prediction result has failed and manual interpretation is recommended. In this way, the ANA recognition model that conforms to the target user's interpretation habit can be used to interpret ANA fluorescence images, with higher interpretation efficiency and more accurate results. The predicted results will also be verified to further determine the true negative of the results and prevent misjudgment.

[0171] See also Figure 6 , Figure 6 FIG. 1 is a structural block diagram of a device for interpreting a fluorescence image of an antinuclear antibody according to an embodiment of the present invention. Figure 6 The device comprises:

[0172] Data acquisition module 601: used to obtain N fluorescence images of the target user's sample to be identified, where N is a positive integer;

[0173] Segmentation processing module 602: configured to perform slider segmentation on the N fluorescence images based on a preset image size to obtain a plurality of segmented image blocks;

[0174] Result prediction module 603: configured to stack the image blocks vertically according to the channel dimension of the image channel, and input the stacked image blocks into a preset ANA recognition model for prediction processing to determine a predicted label for each of the fluorescence images, wherein the label includes a karyotype and titer. The ANA recognition model is obtained by inputting training sample data that conforms to the target user's ANA image interpretation habits into a pre-trained ANA recognition model for transfer learning;

[0175] The result determination module 604 is configured to perform result verification processing based on the predicted label of each of the fluorescent images to determine the target interpretation result of the sample to be identified.

[0176] It should be noted that Figure 6 The functions of each module in the device shown are similar to Figure 1 The contents of each step in the method are similar, and are not described here to avoid repetition. Figure 1 The content of each step in the method.

[0177] The present invention provides a device for identifying fluorescent images of antinuclear antibodies, the device comprising: a data acquisition module for acquiring N fluorescent images of ANA tests of a target user's sample to be identified, where N is a positive integer; a segmentation processing module for performing slider segmentation on the N fluorescent images based on a preset image size to obtain a plurality of segmented image blocks; a result prediction module for stacking the image blocks vertically according to the channel dimensions of the image channels, and inputting the stacked image blocks into a preset ANA recognition model for prediction processing to determine a predicted label for each fluorescent image, the label including a karyotype and titer, the ANA recognition model being obtained by inputting training sample data that conforms to the target user's ANA image interpretation habits into a pre-trained ANA recognition model for transfer learning; and a result determination module for performing result verification processing based on the predicted label of each fluorescent image to determine a target interpretation result of the sample to be identified. In this manner, the ANA recognition model that conforms to the target user's interpretation habits can be used to interpret ANA fluorescent images, resulting in higher interpretation efficiency and more accurate results.

[0178] See also Figure 7 , Figure 7 FIG. 1 is a structural block diagram of a training device for an ANA recognition model according to an embodiment of the present invention, Figure 7 The apparatus shown comprises:

[0179] The sample acquisition module 701 is configured to acquire an ANA training sample set, wherein the ANA training sample set includes a correspondence between a plurality of fifth images and fifth labels;

[0180] The image segmentation module 702 is configured to segment the N fifth images based on a preset image size to obtain a plurality of segmented sample image blocks;

[0181] Model training module 703: used to stack the sample image blocks vertically according to the channel dimension of the image channel, and input the stacked sample image blocks and the fifth label into a deep neural network with a channel attention mechanism for model training until the model converges to obtain a pre-trained ANA recognition model;

[0182] Among them, the loss function of the model training is the cross entropy of the fifth label and the predicted label, and the pre-trained ANA recognition model is used to input the training sample data that conforms to the target user's ANA picture interpretation habits into the pre-trained ANA recognition model for transfer learning to obtain the ANA recognition model as described above.

[0183] It should be noted that Figure 7 The functions of each module in the device shown are similar to Figure 3 The contents of each step in the method are similar, and are not described here to avoid repetition. Figure 3 The content of each step in the method.

[0184] The present invention provides a training device for an ANA recognition model, comprising: a sample acquisition module for acquiring an ANA training sample set, the ANA training sample set including a correspondence between a plurality of fifth images and a fifth label; an image segmentation module for segmenting N fifth images based on a preset image size to obtain a plurality of segmented sample image blocks; a model training module for vertically stacking the sample image blocks according to the channel dimension of the image channels, and inputting the stacked sample image blocks and the fifth labels into a deep neural network with a channel attention mechanism for model training until the model converges, thereby obtaining a pre-trained ANA recognition model; wherein the loss function for model training is the cross entropy between the fifth label and the predicted label, and the pre-trained ANA recognition model is used to input training sample data that conforms to the target user's ANA image interpretation habits into the pre-trained ANA recognition model for transfer learning to obtain the above-mentioned ANA recognition model. By using the above-mentioned method, an AI model can be trained based on a large laboratory database using 200,000 clinical data to obtain a pre-trained ANA recognition model, and the pre-trained ANA recognition model can be used as a pre-trained model, i.e., as a benchmark model for calibration of the target user.

[0185] Figure 8 FIG1 shows an internal structure diagram of a computer device in an embodiment. The computer device can be a terminal or a server. Figure 8As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system and may also store a computer program. When the computer program is executed by the processor, the processor can implement the above method. The internal memory may also store a computer program. When the computer program is executed by the processor, the processor can implement the above method. It will be understood by those skilled in the art that Figure 8 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0186] In one embodiment, a computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the following Figure 1 、 Figure 3 or Figure 5 The steps of the method.

[0187] In one embodiment, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the processor is caused to execute the following Figure 1 、 Figure 3 or Figure 5 The steps of the method.

[0188] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0189] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0190] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A method for interpreting fluorescent images of antinuclear antibodies, characterized in that: The method comprises: Obtain N fluorescence images of the target user's sample to be identified, obtained through ANA testing, where N is a positive integer; Based on a preset image size, performing slider segmentation on the N fluorescence images to obtain a plurality of segmented image blocks; The image blocks are stacked vertically according to the channel dimension of the image channel, and the stacked image blocks are input into a preset ANA recognition model for prediction processing to determine a predicted label for each of the fluorescence images, wherein the label includes a karyotype and a titer. The ANA recognition model is obtained by inputting training sample data that conforms to the target user's ANA image interpretation habits into a pre-trained ANA recognition model for transfer learning; Performing result verification processing based on the predicted label of each of the fluorescent images to determine the target interpretation result of the sample to be identified; The transfer learning at least includes: obtaining a target error database of the target user, the target error database including a correspondence between a first number of first images and first labels; the target error database is a training sample database that conforms to the target user's ANA image interpretation habits, the first image is an ANA fluorescence image with a predicted label different from an actual label, and the first label is a primary label selected using the actual label of the first image; Wherein, obtaining the target error database of the target user includes: Determining the karyotype type of the actual label of each first image in a preset error database; If the karyotype type is a single karyotype or negative, the actual label of the first image is used as the first label of the first image; If the karyotype type is a composite karyotype, the label with the highest titer included in the actual labels of the first image is used as the first label of the first image; If the karyotype type is a composite karyotype, and the actual labels of the first image of the composite karyotype include two or more labels with the highest titers, then randomly select one label with the highest titer from the two or more labels with the highest titer as the first label of the first image; The target error database is obtained based on the first image and the first label.

2. The method according to claim 1, characterized in that Transfer learning is performed as follows, including: Obtaining a target error database for the target user, the target error database including correspondences between a first number of first images and first labels; the target error database being a training sample database consistent with the target user's ANA image interpretation habits, the first images being ANA fluorescence images whose predicted labels differ from their actual labels, and the first labels being primary labels selected using the actual labels of the first images; Determining a migration training sample dataset according to the error database and a preset migration training sample selection rule, wherein the migration training sample dataset includes a second number of correspondences between second images and second labels, where the second number is greater than the first number; The migration training sample data set is input into the pre-trained ANA recognition model for migration training to obtain a trained ANA recognition model.

3. The method according to claim 2, characterized in that The step of determining the migration training sample data set according to the error database and a preset migration training sample selection rule includes: Determining the proportion of each first label in the error database; Randomly extracting the first image and the first label from the error database according to the proportion to obtain candidate first images and first labels; If the image and label in the transfer training sample dataset are the same as the candidate first image and first label, performing random data augmentation processing on the candidate first image and first label, and adding the processed candidate first image and first label to the transfer training sample dataset; If the image and label in the transfer training sample dataset do not overlap with the candidate first image and first label, directly put the candidate first image and first label into the transfer training sample dataset; Repeating the step of randomly extracting the first image and the first label from the error database according to the proportion to obtain candidate first images and first labels until the number of samples in the migration training sample dataset reaches a third number, which is less than the second number; When the number of samples in the migration training sample data set reaches a third number, traversing the second labels in the migration training sample data set to determine label category data for each label in the migration training sample data set; Determine the relationship between the total number of each type of label category data and the first numerical threshold; When there is first label category data whose total number is less than a first threshold value, a third label and a third image of the same label category as the first label category data are randomly selected from a preset sample database and added to the migration training sample dataset until the total number of the first label category data equals the first threshold value; When there is second label category data whose total number is greater than the first threshold, randomly deleting the correspondence between the fourth image and the fourth label in the second label category data until the total number of the second label category data is equal to the first threshold; Repeat the step of determining the relationship between the total number of the label category data and the first number threshold until the number of samples in the migration training sample data set reaches a second number, thereby obtaining a final migration training sample data set.

4. The method according to claim 1, characterized in that If N is greater than 1, the result verification process is performed based on the predicted label of each of the fluorescent images to determine the target interpretation result of the sample to be identified, including: Classifying the predicted labels of the N fluorescence images to obtain a category set for each label category, wherein the category set includes a correspondence between the label category and the number of images; Determine the maximum number of images in all the category sets; Determine a target quantity threshold for the N fluorescence images using N and a preset quantity threshold algorithm, wherein the quantity threshold algorithm is to obtain the target quantity threshold by adding 1 to the product of N and one-half; If the maximum value is greater than or equal to the target quantity threshold, the predicted label corresponding to the maximum value is used as the target interpretation result of the sample to be identified; If the maximum value is less than the target quantity threshold, a warning indication is output, where the warning indication indicates that the prediction result has failed and manual interpretation is recommended.

5. The method according to claim 4, characterized in that: The method further comprises: using the predicted label corresponding to the maximum value as the interpretation label of the target interpretation result of the sample to be identified; Determining whether the interpretation label of the target interpretation result is negative; When the judgment label is negative, the pixel value of the R channel of each fluorescence image of the ANA test is obtained; Determining the average fluorescence intensity of each of the fluorescence images according to the pixel values ​​and a preset average value algorithm; Performing a single-sample T test using the average fluorescence intensity of the N fluorescence images and the preset average fluorescence intensity of negative data to determine the test value of each fluorescence image; If the test value of each of the fluorescent images is less than the preset confidence interval percentage, the judgment label is determined to be a true negative; If the inspection value of any of the fluorescent images is greater than or equal to the preset confidence interval percentage, the output warning indication is executed, and the warning indication indicates that the identification label has failed and manual judgment is recommended.

6. A training method for an ANA recognition model, characterized in that: The method comprises: Obtaining an ANA training sample set, where the ANA training sample set includes a correspondence between a plurality of fifth images and fifth labels; Segmenting the N fifth images based on a preset image size to obtain a plurality of segmented sample image blocks; The sample image blocks are stacked vertically according to the channel dimension of the image channel, and the stacked sample image blocks and the fifth label are input into a deep neural network with a channel attention mechanism for model training until the model converges to obtain a pre-trained ANA recognition model; The loss function of the model training is the cross entropy of the fifth label and the predicted label. The pre-trained ANA recognition model is used to input training sample data that conforms to the target user's ANA picture interpretation habits into the pre-trained ANA recognition model for transfer learning to obtain the ANA recognition model as described in any one of claims 1-5.

7. A device for interpreting fluorescence images of antinuclear antibodies, characterized in that: The device comprises: Data acquisition module: used to obtain N fluorescence images of the target user's sample to be identified, where N is a positive integer; Segmentation processing module: used for performing slider segmentation on the N fluorescent images based on a preset image size to obtain a plurality of segmented image blocks; Result prediction module: used to stack the image blocks vertically according to the channel dimension of the image channel, and input the stacked image blocks into a preset ANA recognition model for prediction processing to determine the predicted label of each fluorescent image, wherein the label includes karyotype and titer. The ANA recognition model is obtained by inputting training sample data that conforms to the target user's ANA image interpretation habits into a pre-trained ANA recognition model for transfer learning; Result determination module: used to perform result verification processing based on the predicted label of each fluorescent image to determine the target interpretation result of the sample to be identified; The transfer learning at least includes: obtaining a target error database of the target user, the target error database including a correspondence between a first number of first images and first labels; the target error database is a training sample database that conforms to the target user's ANA image interpretation habits, the first image is an ANA fluorescence image with a predicted label different from an actual label, and the first label is a primary label selected using the actual label of the first image; Wherein, obtaining the target error database of the target user includes: Determining the karyotype type of the actual label of each first image in a preset error database; If the karyotype type is a single karyotype or negative, the actual label of the first image is used as the first label of the first image; If the karyotype type is a composite karyotype, the label with the highest titer included in the actual labels of the first image is used as the first label of the first image; If the karyotype type is a composite karyotype, and the actual labels of the first image of the composite karyotype include two or more labels with the highest titers, then randomly select one label with the highest titer from the two or more labels with the highest titer as the first label of the first image; The target error database is obtained based on the first image and the first label.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 5, or the steps of the method according to claim 6.

9. A computer device comprising a memory and a processor, characterized in that: The memory stores a computer program, and when the computer program is executed by the processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 5, or the steps of the method according to claim 6.

Images