CN115311216A

CN115311216A - Method, device, equipment and storage medium for interpreting fluorescent picture of antinuclear antibody

Info

Publication number: CN115311216A
Application number: CN202210869963.XA
Authority: CN
Inventors: 任茂源; 刘斯; 余霆嵩; 冯蕾; 林小乔; 李玲; 张玲; 陈彬; 陈丙一; 李慧源
Original assignee: Guangzhou Kingmed Diagnostics Central Co Ltd
Current assignee: Guangzhou Kingmed Diagnostics Central Co Ltd
Priority date: 2022-07-22
Filing date: 2022-07-22
Publication date: 2022-11-08

Abstract

The embodiment of the invention discloses an interpretation method, device, equipment and storage medium of a fluorescent picture of an antinuclear antibody, wherein the method comprises the following steps: based on a preset picture size, carrying out slider segmentation on the N fluorescent pictures to obtain a plurality of segmented image blocks; longitudinally stacking image blocks according to the channel dimension of a picture channel, inputting the stacked image blocks into a preset ANA recognition model for prediction processing, and determining a prediction label of each fluorescent picture, wherein the label comprises a karyotype and a titer, and the ANA recognition model is obtained by inputting training sample data which accords with the ANA picture interpretation habit of a target user into a pre-trained ANA recognition model for transfer learning; and (4) carrying out result inspection processing according to the prediction label of each fluorescence picture, and determining the target interpretation result of the sample to be identified. By means of the method, the ANA identification model which accords with the interpretation habit of the target user can be utilized to interpret the ANA fluorescent picture, interpretation efficiency is high, and the result is accurate.

Description

Method, device, equipment and storage medium for interpreting fluorescent picture of antinuclear antibody

Technical Field

The invention relates to the technical field of antinuclear antibody detection, in particular to a method, a device, equipment and a storage medium for interpreting a fluorescence picture of an antinuclear antibody.

Background

At present, instruments of part of manufacturers have the function of automatically interpreting fluorescent pictures, such as: helios full-automatic immunofluorescent apparatus in AESKU, europattern full-automatic immunofluorescent apparatus in Europe. The instruments of these instrument manufacturers can automatically complete the whole process of the indirect immunofluorescence method, and can interpret the fluorescence picture of the antinuclear antibody (ANA). However, these instruments can only interpret partial karyotypes and cannot interpret mixed karyotypes and the like, and the interpretation criteria of these instruments are not modifiable and the detection models cannot be corrected according to the actual interpretation results or interpretation habits of the examining physician in daily work. These manufacturers' equipment can only identify some common karyotypes and cannot self-calibrate the model based on historical interpretation data.

Because the model interpretation is too rigid and does not accord with the habit of the target user, most of the inspection doctors still adopt manual interpretation at present, and the interpretation efficiency is low, so that a means capable of improving the ANA interpretation efficiency is urgently needed.

Disclosure of Invention

The invention mainly aims to provide a method, a device, equipment and a storage medium for interpreting a fluorescent picture of an antinuclear antibody, which can solve the problem of low interpretation efficiency in the prior art.

In order to achieve the above object, a first aspect of the present invention provides a method for interpreting a fluorescent picture of an antinuclear antibody, the method comprising:

acquiring N fluorescence pictures of ANA detection of a sample to be identified of a target user, wherein N is a positive integer;

based on a preset picture size, carrying out sliding block segmentation on the N fluorescent pictures to obtain a plurality of segmented image blocks;

longitudinally stacking the image blocks according to the channel dimension of a picture channel, inputting the stacked image blocks into a preset ANA recognition model for prediction processing, and determining a prediction label of each fluorescent picture, wherein the label comprises a karyotype and a titer, and the ANA recognition model is obtained by inputting training sample data which accords with the ANA picture interpretation habit of the target user into a pre-trained ANA recognition model for transfer learning;

and performing result inspection processing according to the prediction label of each fluorescent picture, and determining the target interpretation result of the sample to be identified.

In one possible implementation, the transfer learning is performed as follows, including:

acquiring a target error database of the target user, wherein the target error database comprises a corresponding relation between a first number of first pictures and a first label; the target error database is a training sample database which accords with the ANA picture interpretation habit of a target user, the first picture is an ANA fluorescent picture with a prediction label different from an actual label, and the first label is a main label selected by using the actual label of the first picture;

determining a migration training sample data set according to the error database and a preset migration training sample selection rule, wherein the migration training sample data set comprises a corresponding relation between a second number of second pictures and second labels, and the second number is greater than the first number;

and inputting the migration training sample data set into the pre-trained ANA recognition model for migration training to obtain the trained ANA recognition model.

In one possible implementation, the obtaining a target error database of the target user includes:

determining the kernel type of the actual label of each first picture in a preset error database;

if the karyotype type is single karyotype or negative, taking the actual tag of the first picture as the first tag of the first picture;

if the karyotype type is a composite karyotype, taking a label with the highest titer in the actual labels of the first picture as a first label of the first picture;

if the karyotype type is a composite karyotype and the actual labels of the first picture of the composite karyotype include two or more labels with the highest titer, randomly selecting one label with the highest titer from the two or more labels with the highest titer as the first label of the first picture;

and obtaining the target error database based on the first picture and the first label.

In a feasible implementation manner, the determining a 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 picture and the first label in the error database according to the ratio to obtain a candidate first picture and a candidate first label;

if the pictures and the labels in the migration training sample data set are repeated with the candidate first pictures and the candidate first labels, performing random data amplification processing on the candidate first pictures and the candidate first labels, and putting the processed candidate first pictures and the processed candidate first labels into the migration training sample data set;

if the pictures and the labels in the migration training sample data set are not repeated with the candidate first pictures and the candidate first labels, directly putting the candidate first pictures and the candidate first labels into the migration training sample data set;

repeatedly executing the step of randomly extracting the first picture and the first label in the error database according to the ratio to obtain a candidate first picture and a candidate first label until the number of samples of the transfer training sample data set reaches a third number, wherein the third number is smaller than the second number;

traversing a second label in the migration training sample data set after the number of samples in the migration training sample data set reaches a third number, and determining label category data of each label in the migration training sample data set;

determining a magnitude relationship between a total amount of each of said tag category data and a first count threshold;

when first label category data with the total number smaller than a first number threshold exists, randomly selecting a third label and a third picture of the same label category as the first label category data from a preset sample database, and adding the third label and the third picture to the migration training sample data set until the total number of the first label category data is equal to the first number threshold;

when second label category data with the total number larger than a first number threshold exists, randomly deleting the corresponding relation between a fourth picture and a fourth label in the second label category data until the total number of the second label category data is equal to the first number threshold;

and repeating the step of determining the size relation between the total amount of the label category data and the first number threshold value until the number of the samples of the migration training sample data set reaches a second number, so as to obtain a final migration training sample data set.

In a possible implementation manner, if N is greater than 1, the performing a result checking process according to the prediction label of each fluorescence picture to determine a target interpretation result of the sample to be identified includes:

classifying the predicted labels of the N fluorescent pictures to obtain a category set of each label category, wherein the category set comprises the corresponding relation between the label category and the number of the pictures;

determining the maximum value of the number of pictures in all the category sets;

determining a target quantity threshold of N fluorescent pictures by using the N and a preset quantity threshold algorithm, wherein the quantity threshold algorithm is to add 1 to the product of the N and one half to obtain the target quantity threshold;

if the maximum value is larger than or equal to the target quantity threshold value, taking the prediction label corresponding to the maximum value as a target interpretation result of the sample to be identified;

if the maximum value is smaller than the target quantity threshold value, outputting an early warning indication, wherein the early warning indication indicates that the prediction result fails, and recommending manual interpretation.

In a possible implementation manner, the taking the prediction tag corresponding to the maximum value as an interpretation tag of a target interpretation result of the sample to be identified then further includes:

determining whether the interpretation tag of the target interpretation result is negative;

when the interpretation label is negative, acquiring the pixel value of the R channel of each fluorescence picture detected by the ANA;

determining the average fluorescence intensity of each fluorescence picture according to the pixel value and a preset average value algorithm;

carrying out single sample T test by using the average fluorescence intensity of the N fluorescence pictures and the average fluorescence intensity of preset negative data, and determining the test value of each fluorescence picture;

if the check value of each fluorescent picture is smaller than the preset confidence interval percentage, determining that the interpretation label is true negative;

and if the inspection value of any one of the fluorescent pictures is greater than or equal to a preset confidence interval percentage, executing the output early warning indication, wherein the early warning indication indicates that the interpretation label fails, and suggesting manual judgment.

In order to achieve the above object, a second aspect of the present invention provides a method for training an ANA recognition model, the method comprising:

obtaining an ANA training sample set, wherein the ANA training sample set comprises a plurality of corresponding relations between fifth pictures and fifth labels;

performing sliding block segmentation processing on the N fifth pictures based on preset picture sizes to obtain a plurality of segmented sample image blocks;

the method comprises the steps of longitudinally stacking sample image blocks according to the channel dimension of a picture channel, inputting the stacked sample image blocks and a fifth label into a deep neural network with a channel attention mechanism for model training until a model converges to obtain a pre-trained ANA recognition model, wherein a loss function of the model training is the cross entropy of the fifth label and a prediction label, and the pre-trained ANA recognition model is used for inputting training sample data which accords with an ANA picture interpretation habit of a target user into the pre-trained ANA recognition model for transfer learning so as to obtain the ANA recognition model according to the first aspect and any feasible implementation mode.

In order to achieve the above object, a third aspect of the present invention provides an apparatus for recognizing a fluorescent picture of an antinuclear antibody, the apparatus comprising:

a data acquisition module: the method comprises the steps of obtaining N fluorescence pictures of ANA detection of a sample to be identified of a target user, wherein N is a positive integer;

a segmentation processing module: the image segmentation method comprises the steps of performing sliding block segmentation on N fluorescent images based on a preset image size to obtain a plurality of segmented image blocks;

a result prediction module: the image blocks are longitudinally stacked according to the channel dimension of the image channel, the stacked image blocks are input into a preset ANA recognition model for prediction processing, a prediction label of each fluorescent image is determined, the label comprises a karyotype and a titer, and the ANA recognition model is obtained by inputting training sample data which accords with the ANA image interpretation habit of the target user into a pre-trained ANA recognition model for transfer learning;

a result determination module: and the device is used for carrying out result inspection processing according to the prediction label of each fluorescent picture and determining the target interpretation result of the sample to be identified.

In order to achieve the above object, a fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to execute the steps of the method for interpreting a fluorescent picture of an antinuclear antibody according to the first aspect and any one of the possible implementations, or the steps of the method for training an ANA recognition model according to the second aspect.

In order to achieve the above object, a fifth aspect of the present invention provides a computer device, which includes 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 the fluorescent picture of the antinuclear antibody according to the first aspect and any possible implementation manner, or the steps of the method for training the ANA recognition model according to the second aspect.

The embodiment of the invention has the following beneficial effects:

the invention provides an interpretation method of a fluorescent picture of an antinuclear antibody, which comprises the following steps: acquiring N fluorescence pictures detected by ANA of a sample to be identified of a target user, wherein N is a positive integer; based on a preset picture size, carrying out sliding block segmentation on the N fluorescent pictures to obtain a plurality of segmented image blocks; longitudinally stacking image blocks according to the channel dimension of the image channel, inputting the stacked image blocks into a preset ANA recognition model for prediction processing, and determining a prediction label of each fluorescent image, wherein the label comprises a karyotype and a titer, and the ANA recognition model is obtained by inputting training sample data which accords with the ANA image interpretation habit of a target user into a pre-trained ANA recognition model for transfer learning; and (4) carrying out result inspection processing according to the prediction label of each fluorescence picture, and determining the target interpretation result of the sample to be identified. By means of the method, the ANA identification model which accords with the interpretation habit of the target user can be used for interpreting the ANA fluorescent picture, interpretation efficiency is high, and the result is accurate.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the embodiments or the prior art descriptions will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without creative efforts.

Wherein:

FIG. 1 is a flowchart of a method for interpreting a fluorescent picture of an antinuclear antibody according to an embodiment of the present invention;

FIG. 2 (a) is a schematic view of a green fluorescent picture according to an embodiment of the present invention;

FIG. 2 (b) is a schematic diagram of a tag in an embodiment of the present invention;

FIG. 3 is a flowchart of a method for training an ANA recognition model according to an embodiment of the present invention;

FIG. 4 (a) is a partial hierarchy of an ANA identification model according to an embodiment of the present invention;

FIG. 4 (b) is a schematic diagram of another part of the hierarchical structure of an ANA identification model according to an embodiment of the present invention;

FIG. 4 (c) is a schematic diagram of a further partial hierarchy of an ANA recognition model according to an embodiment of the present invention;

FIG. 5 is another flowchart of a method for interpreting a fluorescent picture of an antinuclear antibody according to an embodiment of the present invention;

FIG. 6 is a block diagram showing the structure of an apparatus for interpreting a fluorescent picture of an antinuclear antibody according to an embodiment of the present invention;

FIG. 7 is a block diagram of an embodiment of a device for training an ANA recognition model;

fig. 8 is a block diagram of a computer device according to an embodiment of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The method described in the present application may be applied to both a terminal and a server, and this embodiment is exemplified by the terminal, where the terminal may specifically be a desktop terminal or a mobile terminal, and the mobile terminal may specifically be at least one of a mobile phone, a tablet computer, a notebook computer, and the like. The server may be implemented as a stand-alone server or as a server cluster consisting of a plurality of servers. And are not limited herein.

Referring to fig. 1, fig. 1 is a flowchart of a method for interpreting a fluorescent picture of an antinuclear antibody according to an embodiment of the present invention, where the method shown in fig. 1 includes the following steps:

101. acquiring N fluorescence pictures of ANA detection of a sample to be identified of a target user;

it should be noted that, this application provides an interpretation method of a fluorescence picture of an antinuclear antibody, first, an interpretation object, that is, a fluorescence picture detected by ANA needs to be determined, and fig. 2 (a) is a schematic diagram of a green fluorescence picture in an embodiment of the present invention, where the fluorescence picture in fig. 2 (a) is green, specifically, N fluorescence pictures detected by ANA of a sample to be identified of a target user are obtained, for example, a sample to be identified, such as a blood sample or a specimen, which is subjected to ANA detection is photographed by an automatic photographing system of Helios, and a plurality of fluorescence pictures corresponding to the sample to be identified, which is subjected to ANA detection, which are collected and obtained, may be photographed, so that the fluorescence pictures detected by ANA may be N, where N is a positive integer, and N is a value of 3, that the same sample to be identified is photographed by 3 fluorescence pictures detected by ANA, which is only by way of example and not limited specifically.

102. Based on a preset picture size, carrying out sliding block segmentation on the N fluorescent pictures to obtain a plurality of segmented image blocks;

103. longitudinally stacking the image blocks according to the channel dimension of the image channel, inputting the stacked image blocks into a preset ANA identification model for prediction processing, and determining a prediction label of each fluorescent image;

further, the obtained N fluorescent pictures are subjected to slider segmentation, specifically, based on a preset picture size, for example, 416 × 416, the N fluorescent pictures are subjected to slider segmentation to obtain a plurality of segmented image blocks, for example, the size of a picture taken by the automatic photographing system of Helios is 2560 × 1920, and a picture channel is RGB. The video memory of the current common video card cannot support 2560 × 1920 pictures to directly perform model training or subsequent identification of karyotypes and titers, so that the fluorescent pictures are preprocessed in the application, and the sizes of the pictures are reduced under the condition that the features of the pictures are kept as much as possible. Illustratively, by performing slider cuts on the fluorescence pictures, cutting one 2560 × 1920 picture into 35 416 × 416 picture blocks, since the green fluorescence picture of fig. 2 (a) is identified, G channels of the pictures can be selected and stacked, and finally, data of 416 × 35 is inputted to the neural network. It should be noted that, the fluorescent picture is divided into 35 image blocks by the size of 416 × 416, so that the characteristics of the original whole fluorescent picture can be ensured to be obtained. Further, when N is 3, 105 image blocks can be obtained. It should be noted that, due to the fact that repeated portions exist among the image blocks due to the sliding block cutting, vector calculation can be performed on the basis of the image blocks after the image blocks are obtained, stacking of the image blocks is performed according to vector calculation results, longitudinal stacking can be achieved in a repeated area overlapping mode, and then the image blocks are input into a deep learning network to perform model training.

The model can be a deep artificial neural network constructed based on ResNet, and can identify and output labels of the fluorescence picture, wherein the labels comprise a karyotype and a titer. Illustratively, the present application will classify 19 karyotypes, such as karyotype, nucleolar type, centromere type, karyotype, cytoplasmic granular type, cytoplasmic fibroid type, nuclear compact spot type, PCNA type, centromere-F, golgi type, spindle fibroid type, nuMA type, centromere type, intermediate, chromosome type, rod loop Golgi type, and cytobridge type. And each karyotype has 5 titer classifications, such as 1:80,1:160,1:320,1:640,1:1280 and other 5 titers, namely 19 karyotypes and 5 titers can be identified by the model, a special titer, namely the identification result is negative, and the final model can identify 19 karyotypes, 5 titers and negative, namely 19 × 5+1=96 combinations. It can be understood that each fluorescence image may have a plurality of karyotypes, that is, the label of each image may include a corresponding relationship between a plurality of karyotypes and a titer, wherein the ANA recognition model is obtained by inputting training sample data that meets the ANA image interpretation habit of the target user into a pre-trained ANA recognition model for migration learning. Different target users have different habits, and training samples are different, so that the ANA recognition model in the implementation is different for different target users and is more suitable for the interpretation habits of each user.

Further, the text result of the label identified by the model can be converted into one-hot codes with a length of 114, that is, 19 types of karyotypes, each karyotype is configured with 1+5 coding bits, and finally the coding length is 19 + 6=114. The transcoding rules are as follows: each karyotype is sequentially assigned with 6 bits [ 000000 ] and different bits correspond to different titers, then 1 is assigned to the code of the karyotype according to the result of the titer of the karyotype corresponding to the picture, if no karyotype is present in the picture, the number 1 bit (the first bit in the 6 bits of the karyotype) of the karyotype is assigned with 1, for example, the karyotype is negative, which can be represented as [ 100000 ], and if the karyotype is present, the corresponding position of the karyotype is assigned with 1 according to the grade of the titer. Illustratively, the relationship between 6 positions per karyotype and titer is as follows: if the titer is 1:80, number 2 is assigned with 1, such as [ 010000 ]; if the titer is 1:160, number 3 is assigned with 1, such as [ 001000 ]; if the titer is 1:320, number 4 is assigned with 1, such as [ 000100 ]; if the titer is 1:640, number 5 is assigned 1, such as [ 000010 ]; if the titer is 1:1280, assigning 1 to the 6 th position, such as [ 000001 ]. Referring to fig. 2 (b), fig. 2 (b) is a schematic diagram of a label according to an embodiment of the present invention, and fig. 2 (b) shows an example of four kinds of labels, where if a label is negative or a certain karyotype is not identified, bit 1 of the karyotype is 1 as shown in 201, and further, if bit 1 of 19 karyotypes in a row in which 201 is located is all 1, a label of the fluorescent picture is negative; if the tag is nuclear homogeneity and the titer of the nuclear homogeneity is 1; if the tag is nuclear homogeneity and the titer of nuclear homogeneity is 1; if the tag is nuclear homogeneity and the titer of the nuclear homogeneity is 1. It should be noted that the above description is only an example, and not a limitation to the present embodiment, and the length of data or the expression form of the tag may be adaptively increased according to the type of the karyotype and the titer, so that modifications without departing from the technical concept of the present embodiment are within the scope of the present application.

104. And carrying out result inspection treatment according to the prediction label of each fluorescent picture, and determining the target interpretation result of the sample to be identified.

Finally, in order to ensure the accuracy of the model prediction result, the embodiment performs result inspection processing according to the prediction label of each fluorescence image, determines the target interpretation result of the sample to be recognized, for example, compares the artificial result with the prediction label through manual inspection, determines whether the prediction label is correct, and obtains the target interpretation result of the sample to be recognized. The result checking process is to further check and confirm whether the predicted tag is real or accurate. The final target interpretation result is more real and effective.

The invention provides an interpretation method of a fluorescent picture of an antinuclear antibody, which comprises the following steps: acquiring N fluorescence pictures of ANA detection of a sample to be identified of a target user, wherein N is a positive integer; based on a preset picture size, carrying out sliding block segmentation on the N fluorescent pictures to obtain a plurality of segmented image blocks; longitudinally stacking image blocks according to the channel dimension of a picture channel, inputting the stacked image blocks into a preset ANA recognition model for prediction processing, and determining a prediction label of each fluorescent picture, wherein the label comprises a karyotype and a titer, and the ANA recognition model is obtained by inputting training sample data which accords with the ANA picture interpretation habit of a target user into a pre-trained ANA recognition model for transfer learning; and (4) carrying out result inspection processing according to the prediction label of each fluorescence picture, and determining the target interpretation result of the sample to be identified. By means of the method, the ANA identification model which accords with the interpretation habit of the target user can be utilized to interpret the ANA fluorescent picture, interpretation efficiency is high, and the result is accurate.

Referring to fig. 3, fig. 3 is a flowchart illustrating a method for training an ANA recognition model according to an embodiment of the present invention, where the method shown in fig. 3 includes the following steps:

301. obtaining an ANA training sample set;

it can be understood that, in order to realize automatic fluorescent picture identification, an ANA identification model needs to be trained in advance to realize interpretation of the fluorescent picture, and model training needs to train a sample, so that an ANA training sample set is obtained first, and the ANA training sample set comprises a plurality of corresponding relations between fifth pictures and fifth labels; the picture data of the training sample is derived from ANA fluorescence pictures detected in daily life in a laboratory, and the labels of the pictures are derived from the interpretation of the pictures by qualified detection personnel. These tags were 19 in total, and the pictures were divided into 1:80,1:160,1:320,1:640,1:1280, etc. 5 titers. The data volume of the ANA training sample set can reach 20 thousands, the data volume is obtained based on a large database of a laboratory, and an AI model (ANA identification model) is trained by using 20 thousands of clinical data. The first, second, third, fourth and fifth are used for distinction only, and no specific limitation is imposed on technical features unless otherwise specified.

302. Performing sliding block segmentation processing on the N fifth pictures based on a preset picture size to obtain a plurality of segmented sample image blocks;

303. and longitudinally stacking the sample image blocks 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 is converged to obtain a pre-trained ANA recognition model.

Further, in order to meet the processing requirement of the data in step 102, the N fifth pictures are subjected to slider segmentation processing based on a preset picture size to obtain a plurality of segmented sample image blocks, and model training is performed based on the segmented sample image blocks. The content of step 302 may refer to the content of step 102, which is not described herein again. Further, the picture channels comprise RGB, the sample image blocks are longitudinally stacked according to the channel dimensions of the picture channels, the stacked sample image blocks and the fifth label are input into a deep neural network with a channel attention mechanism to carry out model training until the model converges, and a pre-trained ANA recognition model is obtained. The loss function of the model training is the cross entropy of the fifth label and the prediction label, that is, the loss function is the cross entropy of the prediction result and the actual result, and the pre-trained ANA recognition model is used for inputting training sample data which accords with the ANA picture interpretation habit of the target user into the pre-trained ANA recognition model for transfer learning, so as to obtain the ANA recognition model shown in FIG. 1. Namely, based on the pictures and the labels corresponding to the pictures, the deep learning artificial neural network is established. The neural network is constructed based on ResNet, and a channel attention mechanism is added. And the pre-trained ANA recognition model is used as a pre-training model and is used as a reference model for target user calibration.

Illustratively, in ResNet, the original data is convolved and then added to the original data, thereby alleviating the degradation of the neural network with increasing layer number to some extent.

Two ResNet structures are used in the present neural network: the structures of the reduced size dowsampleres and the normal size NormalRes are shown in fig. 4 (a) and fig. 4 (b): FIG. 4 (a) is a partial hierarchy of an ANA recognition model according to an embodiment of the present invention; fig. 4 (b) is a schematic diagram of another partial hierarchy of an ANA identification model according to an embodiment of the present invention. In fig. 4 (a) and 4 (b), conv2D is the convolution layer, filter is the number of output channels, kernel _ size is the convolution kernel size, strings is the step size, and batch normalization is batch normalization; reLU is an activation function; ADD is addition; inptu is input; output is Output; h is height, w is width, and c is depth.

For example, the Channel Attention mechanism (Channel Attention) refers to that a corresponding neural layer is used to use a full connection layer link for channels of stacked pictures to learn features between the picture channels, and specifically, refer to fig. 4 (c), where fig. 4 (c) is a schematic diagram of a further partial hierarchy structure of an ANA recognition model in an embodiment of the present invention, where FC is a full connection layer, and Scale is a normalization function.

Further, the output structure of the neural network is as follows: after the ResNet is completed, the data is normalized again to form the final output data. The normalization procedure is as follows: the final data is fully concatenated 19 times (output number 6, activation function Sigmoid), and then the 19 data are concatenated to form 114-dimensional data. The pre-loss function is a loss function trained by calculating the cross entropy of the actual result and the predicted result as a model.

And finally training an AI model capable of classifying ANA pictures based on the neural network structure and the data basis of the laboratory. The data base of the laboratory can be the data record of the past ANA test of the gold domain.

The invention provides a training method of an ANA recognition model, which comprises the following steps: obtaining an ANA training sample set, wherein the ANA training sample set comprises a plurality of corresponding relations between fifth pictures and fifth labels; performing sliding block segmentation processing on the N fifth pictures based on a preset picture size to obtain a plurality of segmented sample image blocks; the method comprises the steps of longitudinally stacking sample image blocks according to the channel dimension of a picture channel, inputting the stacked sample image blocks and a fifth label into a deep neural network with a channel attention mechanism to carry out model training until the model converges to obtain a pre-trained ANA recognition model, wherein a loss function of model training is the cross entropy of the fifth label and a prediction label, and the pre-trained ANA recognition model is used for inputting training sample data which accords with the ANA picture interpretation habit of a target user into the pre-trained ANA recognition model to carry out transfer learning so as to obtain the ANA recognition model shown in figure 1. By the method, an AI model can be trained by using clinically 20 million data based on a large laboratory database to obtain a pre-trained ANA recognition model, and the pre-trained ANA recognition model is used as a pre-trained model, namely, a reference model for calibration of a target user.

Referring to fig. 5, fig. 5 is another flow chart of a method for interpreting a fluorescence image of an antinuclear antibody according to an embodiment of the present invention, and the method shown in fig. 5 includes the following steps:

501. acquiring N fluorescence pictures detected by ANA of a sample to be identified of a target user;

wherein N is a positive integer.

502. Based on a preset picture size, carrying out sliding block segmentation on the N fluorescent pictures to obtain a plurality of segmented image blocks;

503. longitudinally stacking the image blocks according to the channel dimension of a picture channel, inputting the stacked image blocks into a preset ANA identification model for prediction processing, and determining a prediction label of each fluorescent picture;

the ANA recognition model is obtained by inputting training sample data which accords with the ANA picture interpretation habit of a target user into a pre-trained ANA recognition model for transfer learning.

It should be noted that

steps

501, 502, and 503 are similar to those shown in

steps

101, 102, and 103 in fig. 1, and for avoiding repetition, no detailed description is provided here, and reference may be specifically made to

steps

101, 102, and 103.

It can be understood that most AI models on the market are pre-trained, and cannot be dynamically fine-tuned according to the actual prediction result of the user. In order to solve the problem, a transfer learning module is added to the system, different data and partial comparison data of an actual detection result and an AI prediction result can be used as materials for transfer training to fine tune an AI model, for example, in the embodiment, the pre-trained model is secondarily trained based on data of self interpretation habits of different target users, that is, training sample data according with the ANA picture interpretation habits of the target users is input into the pre-trained ANA recognition model for transfer learning, so as to obtain different ANA recognition models corresponding to the different target users, and in a feasible implementation manner, transfer learning is performed according to the following manner, including steps P01, P02 and P03:

p01, acquiring a target error database of the target user;

the target error database comprises a corresponding relation between a first number of first pictures and a first label; the target error database is a training sample database which accords with the ANA picture interpretation habit of the target user, furthermore, a picture of an actual label which accords with the interpretation habit of the target user can be input into a pre-trained model, when a predicted label output by the model is different from the actual label, the corresponding relation of the picture and the actual label is recorded into the error database, further, a first picture in the target error database is an ANA fluorescent picture with the predicted label different from the actual label, and the first label can be a main label obtained by selecting the actual label of the first picture, and when the data volume of the error database reaches a first number, the target error database is obtained.

In a possible implementation manner, the present embodiment configures a first tag for the first picture in the error database, where the first tag is a main tag determined based on the actual tag of the first picture, so that step P01 may include steps P11, P21, P31, P41, and P51:

p11, determining the karyotype type of the actual label of each first picture in a preset error database;

it should be noted that the preset error database is an error database before the first tag is not configured, and further the preset error database includes a corresponding relationship between the first picture and the actual tag, and further, a karyotype type of the actual tag of each first picture in the preset error database is determined, where the karyotype type includes, but is not limited to, a single karyotype, a composite karyotype, or a negative karyotype. Furthermore, in this embodiment, the first tag is configured for each first picture based on the core type of the actual tag of each first picture.

P21, if the karyotype type is single karyotype or negative, taking the actual tag of the first picture as the first tag of the first picture;

if the core type of the actual tag of the first picture is a single core type or negative (refer to 201, 202, and 203 in fig. 2 (b)), the actual tag of the first picture is used as the first tag of the first picture, and the first tag corresponding to the first picture is obtained.

P31, if the karyotype type is a composite karyotype, taking a label with the highest titer in the actual labels of the first picture as a first label of the first picture;

if the karyotype type is a composite karyotype (see 204 and 205 in fig. 2 (b)), the tag with the highest titer included in the actual tags of the first picture is used as the first tag of the first picture, so as to obtain the first tag corresponding to the first picture.

P41, if the karyotype type is a composite karyotype and the actual labels of the first picture of the composite karyotype include two or more labels with the highest titer, randomly selecting one label with the highest titer from the two or more labels with the highest titer as the first label of the first picture;

if the karyotype type is a composite karyotype, and the actual tags of the first picture of the composite karyotype include two or more tags with the highest titer (see 204 and 205 in fig. 2 (b)), one tag with the highest titer is randomly selected from the two or more tags with the highest titer as the first tag of the first picture, so as to obtain the first tag corresponding to the first picture. For example, in 204 and 205 of fig. 2 (b), both karyotypes are titer of 1 to 160, and both can be regarded as the highest titer, so that the actual tags include two or more tags with the highest titers, and therefore, the first tag corresponding to the first picture is selected as one of the karyotype 1.

And P51, obtaining the target error database based on the first picture and the first label.

Further, first labels of a first number of first pictures are obtained, and a target error database is obtained.

Exemplarily, steps P11, P21, P31, P41, and P51:

(1) Selecting data with a recent actual detection result (actual label) different from an AI prediction result (prediction label) as a preset error database (Dataset) _miss ) Member of (1), dataset _miss The amount of data of (a) is 250 cases (a first amount).

(2) To Dataset _miss The first picture in (1) is given a first Label (Label) _main ) For subsequent random extraction of data. Label (Label) _main The assigning rule of (1): label for a first picture if the actual result for that picture is a single karyotype or negative _main The actual detection result is obtained; label of the first picture if the actual result of the picture is a composite karyotype _main The karyotype with the highest titer in the compound karyotypes, and if 2 or more karyotypes with the highest titer exist, randomly selecting a combination of one karyotype and the highest titer as the Label of the picture _main 。

P02, determining a migration training sample data set according to the target error database and a preset migration training sample selection rule;

further, after the target error database is obtained, a migration training sample data set can be obtained through the target error database, and the migration training sample data set is used for migration learning of a pre-trained ANA recognition model to obtain different ANA recognition models of different target users, specifically, the migration training sample data set is determined according to the target error database and a preset migration training sample selection rule, wherein the migration training sample selection rule is used for selecting a training sample for the migration learning, and includes but is not limited to a data expansion rule and the like, the migration training sample data set includes a corresponding relation between a second number of second pictures and second labels, and the second number is greater than the first number.

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

p12, determining the proportion of each first label in the error database;

it should be noted that, members of the migration training sample data set can be obtained through random extraction, and in this embodiment, random extraction is performed according to the proportion, so that the proportion of each first label in the error database is first determined, for example: nuclear homogeneity 1:320, and the like, and further, the ratio relationship is [ core homogeneity 1: (nuclear homogeneity 1: (core particle type 1: (nuclear homogeneity 1: 2:1:1 ] is used.

P22, randomly extracting the first picture and the first label in the error database according to the ratio to obtain a candidate first picture and a candidate first label;

further, after the ratio is determined, random extraction may be performed, wherein the first picture and the first label in the error database are randomly extracted according to the ratio to obtain a candidate first picture and a candidate first label, that is, if 1 first label is randomly extracted as "nuclear homogeneity 1: the 320 ″ first picture needs to randomly extract 3 first pictures whose first labels are kernel homogeneity 1. It is further required to determine whether the candidate first picture and the candidate first label are overlapped with the picture and the label in the training sample data set, if so, execute step P32, and if not, execute step P42.

P32, if the pictures and the labels in the migration training sample data set are repeated with the candidate first picture and the candidate first label, performing random data amplification processing on the candidate first picture and the candidate first label, and putting the processed candidate first picture and the processed candidate first label into the migration training sample data set;

if the candidate first picture and the candidate first label are already in the transfer training sample data set, it is indicated that the picture and the label in the transfer training sample data set are repeated with the candidate first picture and the candidate first label, that is, the sample is repeated, and then the candidate first picture and the candidate first label need to be subjected to random data amplification processing and then added to the transfer training sample data set, so as to obtain a second picture and a second label. The random data amplification processing includes but is not limited to vertical mirror image, horizontal mirror image, 180 ° rotation and other image processing means.

P42, if the pictures and the labels in the migration training sample data set are not repeated with the candidate first pictures and the candidate first labels, directly putting the candidate first pictures and the candidate first labels into the migration training sample data set;

it can be understood that, if not repeated, the candidate first picture and the candidate first label may be directly put into the migration training sample data set to obtain the second picture and the second label.

P52, repeating the step of randomly extracting the first picture and the first label in the error database according to the proportion to obtain a candidate first picture and a candidate first label until the number of the samples of the migration training sample data set reaches a third number, wherein the third number is smaller than the second number;

further, the step P22 is continuously repeated, and the number of samples is continuously expanded until the number of samples in the migration training sample data set reaches a third number, where the third number is smaller than the second number. Illustratively, the third number may be 4800 and the second number may be 9600. And after the number of the samples of the migration training sample data set reaches the third number, expanding the data size in another way, which is specifically as follows.

P62, traversing a second label in the migration training sample data set after the number of samples in the migration training sample data set reaches a third number, and determining label category data of each label in the migration training sample data set;

further, after the number of samples in the migration training sample data set reaches a third number, traversing a second label in the migration training sample data set, and determining label category data of each label in the migration training sample data set, where the label category data is used to indicate the number of label categories of each label, for example, 99 second pictures with the second label being 1 of nuclear homogeneity 1:320 second picture 110; then the number of tag classes for core homogeneity 1:320 has a tag category number of 100.

P72, determining the size relation between the total amount of each type of label category data and a first number threshold;

further, the total number of the tag category data, i.e. the total number of the second pictures of the target tag category. Wherein, the first number threshold is 100, and then the magnitude relation of the total number and the first threshold is compared, if the total number is smaller than the first number threshold, the step P82 is executed; if the total number is greater than the first number threshold, step P92 is performed.

P82, when first label category data with the total number smaller than a first number threshold exists, randomly selecting a third label and a third picture of a label category which is the same as the first label category data from a preset sample database, and adding the third label and the third picture to the migration training sample data set until the total number of the first label category data is equal to the first number threshold;

when there is first tag category data whose total number is less than a first number threshold, for example, 99 second pictures with a second tag of nuclear homogeneity 1 of 80 in the above example, and the first number threshold is 100, then 99 "is 100, and then the first tag category data is" 99 second pictures with a second tag of nuclear homogeneity 1 of 80", it is necessary to supplement the number of the nuclear homogeneity 1 to 100, so that a third tag of the same tag category as the first tag category data is randomly selected from a preset sample database and added to a migration training sample data set together with the third picture, and the preset sample database is a common sample database, for example, a laboratory test library, that is, a third picture of a third tag randomly selected from the laboratory test library" nuclear homogeneity 1 "is added to the migration training sample data set.

P92, when second label category data with the total number larger than a first number threshold exists, randomly deleting the corresponding relation between a fourth picture and a fourth label in the second label category data until the total number of the second label category data is equal to the first number threshold;

when there is second tag category data with a total number greater than the first number threshold, such as the second tag is nuclear homogeneity 1 in the above example: 110 second pictures of 320, if the first number threshold is 100, then 110>100, and the second label category data is "the second label is a nuclear homogeneity 1:320 second picture 110 ", then the core homogeneity 1: the number of 320 pictures is reduced to 100, so that the corresponding relation between the fourth picture and the fourth label in the second label type data is randomly deleted until the total number of the second label type data is equal to the first number threshold. That is, deleting 10 fourth tags in the second tag category data as core homogeneity 1:320 such that the second tag category data is 100.

And P102, repeating the step of determining the size relation between the total amount of the label category data and the first number threshold value until the number of the samples of the migration training sample data set reaches a second number, and obtaining a final migration training sample data set.

And continuously repeating the process, namely repeating the step of determining the size relation between the total number of the label category data and the first number threshold value until the number of the samples of the migration training sample data set reaches the second number, so as to obtain the final migration training sample data set. And until the number of the samples of the training sample data set is finally migrated to meet the requirement, the second number is met, and the second number can be 9600.

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

(1) Obtain a target error database (Datase) _tmiss ) Thereafter, a data set (Dataset) is defined _retrain ) The data set is the data set finally used for training, namely the migration training sample data set, dataset _retrain The data acquisition mode is as follows: according to Datase _tmiss In (2) the first Label _main In proportion to randomly extract Datase _tmiss Data to Dataset _retrain If the randomly drawn data has been included in the Dataset _retrain That is, if the method is repeated, the picture is subjected to random data amplification (including vertical mirror image, horizontal mirror image, 180 ° rotation) and then included in Dataset _retrain To Dataset _retrain The amount of data of (b) reaches 4800 (third amount).

Then traverse each Label _main Type of (2), judging Dataset _retrain If the data amount (label category data) of the type exceeds 100 cases (first number threshold), if the data amount (label category data) of the type exceeds 100 cases, the data of the type is randomly removed until the data amount of the type reaches 100 cases, and if the data amount (label category data) of the type is less than 100 cases, the data of the corresponding category is randomly selected from a big data pool of a laboratory (gold field) and is included in a Dataset _retrain In (1). Final Dataset _retrain There should be 9600 pieces of data (second number).

And P03, inputting the migration training sample data set into the pre-trained ANA recognition model for migration training to obtain the trained ANA recognition model.

And finally, training a model by using data in the migration training sample data set (Datasetretrain) according to the training mode of the pre-trained ANA recognition model, namely inputting the migration training sample data set into the pre-trained ANA recognition model for migration training, and then replacing the original AI model, namely obtaining the trained ANA recognition model.

Further, different regular expression templates are provided according to the items, and existing results in the target user database are converted into corresponding standard result data. Taking an antinuclear antibody as an example, this would convert "particle type 1. At present, a lot of common result data are covered, and new regular expressions can be continuously added according to actual conditions, so that standardization is realized.

Further, in this embodiment, in order to ensure that the result is accurate for the same sample to be identified or multiple photographs are taken, so N is greater than 1, and further, the result is checked according to the prediction tag of each fluorescent picture to determine the target interpretation result of the sample to be identified, including the following

steps

504, 505, and 506:

504. classifying the predictive labels of the N fluorescent pictures to obtain a class set of each label class;

after the fluorescent pictures of the sample to be identified are identified by using the ANA identification model, the prediction labels of the fluorescent pictures can be obtained, the prediction labels can be the sum of the prediction labels of the image blocks corresponding to each fluorescent picture, and then the prediction labels of the N fluorescent pictures are classified to obtain the category set of each label category, that is, the same label categories are clustered to obtain the category set of each label category, wherein the category set comprises the corresponding relation between the label category and the number of the pictures.

505. Determining the maximum value of the number of pictures in all the category sets;

further, the present embodiment verifies the result in a few majority-compliant ways to obtain accuracy and authenticity. Determining the maximum value of the number of pictures in the category set, wherein the maximum value is the most label result in the prediction results, and the maximum value represents the majority;

506. determining a target quantity threshold value of the N fluorescent pictures by utilizing the N and a preset quantity threshold value algorithm;

the method comprises the steps of obtaining a target quantity threshold value of N fluorescence pictures, wherein the target quantity threshold value is obtained by adding 1 to the product of N and one half of N, and the target quantity threshold value is obtained by utilizing a preset quantity threshold value algorithm. Exemplarily, when N is 3, the target number threshold is 2, that is, (3 × 0.5) +1=1.5, where the number of pictures is an integer, and thus the calculation result is rounded to 2. Further, if the maximum value is equal to or greater than the target number threshold, step 507 is performed, and if the maximum value is less than the target number threshold, step 508 is performed.

507. If the maximum value is larger than or equal to the target quantity threshold value, taking the prediction label corresponding to the maximum value as a target interpretation result of the sample to be identified;

further, if the maximum value is greater than or equal to the target number threshold, it may be indicated that the maximum value is more than half of the actual total number, so that the result of the minority majority is error-free, and therefore, the predicted label corresponding to the maximum value of the label category is used as the target interpretation result of the sample to be identified.

Illustratively, the Helios system takes three pictures of a sample to be identified, which should generally all be of the same type, but in the actual AI interpretation, the three pictures may be given different types, for example: nuclear particle type 1:80, nuclear particle type 1:80, nuclear particle type 1:160. in the early warning module, the following three conditions are correspondingly processed: (1) If the prediction labels of the three pictures are consistent, directly outputting corresponding results; (2) If the prediction labels of two pictures are consistent but the prediction label of the remaining picture is not consistent, outputting the results of the two consistent pictures; (3) The prediction labels of the three pictures are different, and early warning prompt is carried out and manual interpretation is carried out.

Further, in the actual test procedure, negative data is about 50%, and the examining physician would expect negative pictures read by the AI judgment to be true negative. In order to prevent the AI model from interpreting positive pictures as negative pictures, after the interpretation of the AI model is completed, the AI interpretation result is warned by using some conventional methods, that is, the prediction tag corresponding to the maximum value is used as the interpretation tag of the target interpretation result of the sample to be recognized, and then the method further comprises the following steps of L01, L02, L03, L04, L05 and L06:

l01, determining whether the interpretation label of the target interpretation result is negative;

it is understood that, referring to fig. 2 (b), whether or not it is negative is determined from the tags, and thus, whether or not the interpretation tag of the target interpretation result, which is the prediction tag corresponding to the maximum value, is negative is determined. If the negative result is positive, executing the step L02 to determine whether the negative result is true negative so as to avoid false detection; if the result is not negative, the final result can be directly output, namely the system can finally output the prediction result of the picture, and the detection personnel can perform manual examination or believe that the result of the AI interpretation issues a report according to the result of the AI interpretation.

L02, when the interpretation label is negative, acquiring the pixel value of the R channel of each fluorescence picture detected by the ANA;

l03, determining the average fluorescence intensity of each fluorescence picture according to the pixel value and a preset average value algorithm;

l04, performing single sample T test by using the average fluorescence intensity of the N fluorescence pictures and the average fluorescence intensity of preset negative data, and determining the test value of each fluorescence picture;

l05, if the test value of each fluorescence picture is less than the preset confidence interval percentage, determining that the interpretation label is true negative;

and L06, if the check value of any one of the fluorescent pictures is greater than or equal to a preset confidence interval percentage, executing the output early warning indication, wherein the early warning indication indicates that the interpretation label fails, and suggesting manual judgment.

Since the average fluorescence intensity in the present system refers to the average value of G channel pixel values in one picture, generally, the lower the titer of the picture, the lower the average fluorescence intensity, and a positive correlation exists between the two, when the interpretation label is negative, the pixel value of the R channel of each fluorescence picture detected by ANA is obtained for use in the subsequent calculation of the average fluorescence intensity. And the system calculates the average fluorescence intensity of 20 ten thousand pictures (the training sample data of the pre-trained ANA recognition model), calculates the average value and the standard deviation of the average fluorescence intensity of negative data in the pictures, calculates the average fluorescence intensity of three pictures based on the pixel value of a G channel after 3 pictures pass through a few pictures and are subjected to majority early warning and are predicted to be negative by the model, performs single-sample T test with the average fluorescence intensity of the negative data, outputs a negative conclusion if the three pictures are within 95% of one side, and otherwise, manually interprets the negative conclusion. Wherein 95% is a preset confidence interval percentage.

After the AI prediction and early warning mechanism, the system finally outputs the prediction result of the picture, and detection personnel can perform manual auditing or believe that the result of the AI interpretation is published according to the result of the AI interpretation.

508. If the maximum value is smaller than the target quantity threshold value, outputting an early warning indication, wherein the early warning indication indicates that the prediction result fails, and recommending manual interpretation.

That is, if the maximum value is smaller than the target number threshold value, it is indicated that the maximum value is not over half, and the result has an error, so that an early warning indication is output, and the early warning indication indicates that the prediction result fails, and manual interpretation is recommended.

The invention provides an interpretation method of a fluorescence picture of an antinuclear antibody, which comprises the following steps: acquiring N fluorescence pictures of ANA detection of a sample to be identified of a target user, wherein N is a positive integer; based on a preset picture size, carrying out sliding block segmentation on the N fluorescent pictures to obtain a plurality of segmented image blocks; longitudinally stacking image blocks according to the channel dimension of a picture channel, inputting the stacked image blocks into a preset ANA identification model for prediction processing, and determining a prediction label of each fluorescent picture; classifying the prediction labels of the N fluorescent pictures to obtain a class set of each label class; determining the maximum value of the number of pictures in all the category sets; determining a target quantity threshold of the N fluorescence pictures by using N and a preset quantity threshold algorithm; if the maximum value is larger than or equal to the target quantity threshold value, taking the prediction label corresponding to the maximum value as a target interpretation result of the sample to be identified; if the maximum value is smaller than the target quantity threshold value, outputting an early warning instruction, wherein the early warning instruction indicates that the prediction result fails, and recommending manual interpretation. By means of the method, the ANA identification model which accords with the interpretation habit of the target user can be used for interpreting the ANA fluorescent picture, interpretation efficiency is high, and the result is accurate. And the prediction result is also verified, so that the true negative of the result can be further judged, and the misjudgment is prevented.

Referring to fig. 6, fig. 6 is a block diagram of a reading apparatus for a fluorescent picture of an antinuclear antibody according to an embodiment of the present invention, where the apparatus shown in fig. 6 includes:

the data acquisition module 601: the method comprises the steps of obtaining N fluorescence pictures detected by ANA of a sample to be identified of a target user, wherein N is a positive integer;

the segmentation processing module 602: the image segmentation method comprises the steps of performing sliding block segmentation on N fluorescent pictures based on a preset picture size to obtain a plurality of segmented image blocks;

the result prediction module 603: the image blocks are longitudinally stacked according to the channel dimension of the image channel, the stacked image blocks are input into a preset ANA recognition model for prediction processing, a prediction label of each fluorescent image is determined, the label comprises a karyotype and a titer, and the ANA recognition model is obtained by inputting training sample data which accords with the ANA image interpretation habit of the target user into a pre-trained ANA recognition model for transfer learning;

the result determination module 604: and the device is used for carrying out result inspection processing according to the prediction label of each fluorescent picture and determining the target interpretation result of the sample to be identified.

It should be noted that functions of each module in the apparatus shown in fig. 6 are similar to contents of each step in the method described in fig. 1, and for avoiding repetition of this description, details of each step in the method described in fig. 1 may be specifically referred to.

The invention provides a device for identifying a fluorescent picture of an antinuclear antibody, which comprises: a data acquisition module: the method comprises the steps of obtaining N fluorescence pictures detected by ANA of a sample to be identified of a target user, wherein N is a positive integer; a segmentation processing module: the image segmentation method comprises the steps of performing sliding block segmentation on N fluorescent images based on a preset image size to obtain a plurality of segmented image blocks; a result prediction module: the image blocks are longitudinally stacked according to the channel dimension of the image channel, the stacked image blocks are input into a preset ANA recognition model for prediction processing, a prediction label of each fluorescent image is determined, the label comprises a karyotype and a titer, and the ANA recognition model is obtained by inputting training sample data which accords with the ANA image interpretation habit of the target user into a pre-trained ANA recognition model for migration learning; a result determination module: and the device is used for carrying out result inspection processing according to the prediction label of each fluorescent picture and determining the target interpretation result of the sample to be identified. By means of the method, the ANA identification model which accords with the interpretation habit of the target user can be utilized to interpret the ANA fluorescent picture, interpretation efficiency is high, and the result is accurate.

Referring to fig. 7, fig. 7 is a block diagram illustrating a structure of an apparatus for training an ANA recognition model according to an embodiment of the present invention, where the apparatus shown in fig. 7 includes:

the sample acquisition module 701: the ANA training sample set is used for obtaining the ANA training sample set, and the ANA training sample set comprises a plurality of corresponding relations between fifth pictures and fifth labels;

the picture segmentation module 702: the image segmentation device is used for segmenting the N fifth pictures based on preset picture sizes to obtain a plurality of segmented sample image blocks;

model training module 703: the image processing device is used for longitudinally stacking the sample image blocks according to the channel dimension of the image channel, 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, and obtaining a pre-trained ANA recognition model;

the loss function of the model training is the cross entropy of a fifth label and a prediction label, and the pre-trained ANA recognition model is used for inputting training sample data which accords with the ANA picture interpretation habit of a target user into the pre-trained ANA recognition model for transfer learning, so that the ANA recognition model is obtained.

It should be noted that the functions of each module in the apparatus shown in fig. 7 are similar to the contents of each step in the method described in fig. 3, and for avoiding repetition, no detailed description is provided here, and the contents of each step in the method described in fig. 3 may be referred to specifically.

The invention provides a training device of an ANA recognition model, which comprises: a sample acquisition module: the ANA training sample set is used for obtaining the ANA training sample set, and the ANA training sample set comprises a plurality of corresponding relations between fifth pictures and fifth labels; the picture segmentation module: the image segmentation method comprises the steps of segmenting N fifth images based on preset image sizes to obtain a plurality of segmented sample image blocks; a model training module: the image processing system is used for longitudinally stacking the sample image blocks according to the channel dimension of the image channel, 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, and obtaining a pre-trained ANA recognition model; and the pre-trained ANA recognition model is used for inputting training sample data which accords with the ANA picture interpretation habit of the target user into the pre-trained ANA recognition model for transfer learning so as to obtain the ANA recognition model. By the method, an AI model can be trained by using clinically 20 million data based on a large laboratory database to obtain a pre-trained ANA recognition model, and the pre-trained ANA recognition model is used as a pre-trained model, namely, a reference model for calibration of a target user.

FIG. 8 is a diagram illustrating an internal structure of a computer device in one embodiment. The computer device may specifically be a terminal, and may also be a server. As shown in fig. 8, the computer device includes a processor, a memory, and a network interface connected by a system bus. Wherein 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 which, when executed by the processor, causes the processor to carry out the above-mentioned method. The internal memory may also have stored therein a computer program which, when executed by the processor, causes the processor to perform the method described above. Those skilled in the art will appreciate that the architecture shown in fig. 8 is merely a block diagram of some of the structures associated with the disclosed aspects and is not intended to limit the computing devices to which the disclosed aspects apply, as particular computing devices may include more or less components than those shown, or may combine certain components, or have a different arrangement of components.

In an embodiment, a computer device is proposed, comprising a memory and a processor, the memory storing a computer program which, when executed by the processor, causes the processor to perform the steps of the method as described in fig. 1, fig. 3 or fig. 5.

In an embodiment, a computer-readable storage medium is proposed, in which a computer program is stored which, when being executed by a processor, causes the processor to carry out the steps of the method as described in fig. 1, 3 or 5.

It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above may be implemented by a computer program, which may be stored in a non-volatile computer readable storage medium, and when executed, may include the processes of the embodiments of the methods described above. Any reference to memory, storage, database or other medium used in the embodiments provided herein 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 a variety of forms such as Static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double Data Rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous Link DRAM (SLDRAM), rambus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

The technical features of the above embodiments can be arbitrarily combined, and for the sake of brevity, all possible combinations of the technical features in the above embodiments are not described, but should be considered as the scope of the present specification as long as there is no contradiction between the combinations of the technical features.

The above-mentioned embodiments only express several embodiments of the present application, and the description thereof is specific and detailed, but not construed as limiting the scope of the present application. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the concept of the present application, which falls within the scope of protection of the present application. Therefore, the protection scope of the present patent application shall be subject to the appended claims.

Claims

1. A method for interpreting a fluorescent picture of an antinuclear antibody, said method comprising:

2. The method of claim 1, wherein the transfer learning is performed as follows:

3. The method as claimed in claim 2, wherein said obtaining a target error database of the target user comprises:

4. The method according to claim 2, wherein determining a migration training sample data set according to the error database and a preset migration training sample selection rule comprises:

determining the proportion of each first label in the error database;

determining a magnitude relationship of a total amount of each of the tag category data to a first number threshold;

5. The method according to claim 1, wherein if N is greater than 1, the performing a result verification process according to the prediction label of each of the fluorescence pictures to determine the target interpretation result of the sample to be identified comprises:

classifying the prediction labels of the N fluorescent pictures to obtain a category set of each label category, wherein the category set comprises the corresponding relation between the label category and the number of the pictures;

determining a target quantity threshold of the N fluorescent pictures by using the N and a preset quantity threshold algorithm, wherein the quantity threshold algorithm is to add 1 to the product of the N and one half to obtain the target quantity threshold;

6. The method according to claim 5, wherein the step of using the prediction tag corresponding to the maximum value as an interpretation tag of the target interpretation result of the sample to be identified further comprises:

and if the check value of any one of the fluorescent pictures is greater than or equal to the preset confidence interval percentage, executing the output early warning indication, wherein the early warning indication indicates that the interpretation label fails, and suggesting manual judgment.

7. A method for training an ANA recognition model, the method comprising:

dividing the N fifth pictures based on a preset picture size to obtain a plurality of divided sample image blocks;

longitudinally stacking the sample image blocks 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 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 prediction label, and the pre-trained ANA recognition model is used for inputting training sample data which accords with the ANA picture interpretation habit of the target user into the pre-trained ANA recognition model for transfer learning so as to obtain the ANA recognition model according to any one of claims 1-6.

8. An apparatus for recognizing a fluorescent picture of an antinuclear antibody, said apparatus comprising:

a data acquisition module: the method comprises the steps of obtaining N fluorescence pictures detected by ANA of a sample to be identified of a target user, wherein N is a positive integer;

a segmentation processing module: the image segmentation method comprises the steps of performing sliding block segmentation on N fluorescent pictures based on a preset picture size to obtain a plurality of segmented image blocks;

9. A computer-readable storage medium, storing a computer program, wherein the computer program, when executed by a processor, causes the processor to perform the steps of the method as claimed in any one of claims 1 to 6, or to perform the steps of the method as claimed in claim 7.

10. A computer device comprising a memory and a processor, characterized in that the memory stores a computer program which, when executed by the processor, causes the processor to carry out the steps of the method as claimed in any one of claims 1 to 6 or to carry out the steps of the method as claimed in claim 7.