A blood type auxiliary analysis method based on machine learning

Abstract

The present invention relates to a blood type auxiliary judgment method based on machine learning. The method analyzes the strong and weak agglutination areas of blood during the blood type detection process. The initial strong agglutination area upper limit, weak agglutination area interval, and non-agglutination area lower limit of the area enclosed by the blood strong agglutination standard circle, the blood weak agglutination standard circle, and the reaction hole boundary circle are used to analyze each blood sample to be detected in sequence. The complementary processing mode of area comparison and machine learning is executed. While continuously optimizing the agglutination detection model, the blanks between different types of agglutination area intervals are continuously filled and improved. Through the continuous mutual promotion between the two, the accuracy of blood type auxiliary judgment by the two methods of area comparison and machine learning is improved at the same time, the shortcomings of the prior art are overcome, and the working efficiency of blood type auxiliary judgment is effectively improved.

Description

Technical Field

[0001] The present invention relates to a blood type auxiliary analysis method based on machine learning, belonging to the technical field of blood type auxiliary detection. Background Art

[0002] At present, the blood typing system determines the blood type by taking a photo and observing the image to identify agglutination or non-agglutination. The principle is to identify agglutination or non-agglutination based on template or feature matching. Agglutination or non-agglutination can be determined if it meets the established template features. However, there are many weak agglutination images, and many that do not meet the template features are easily missed or misinterpreted, resulting in blood type identification errors or irregular unexpected antibody misdetection. At present, the blood typing instruments on the market analyze the results of the images in the above way, which can no longer meet the detection needs. The testers can only review the image results by naked eyes to avoid the issuance of wrong results. However, each person has personalized subjective bias in judging the image. In addition, after a long time and a large number of well position image reviews, people are prone to fatigue, resulting in the failure to review the problematic results. Therefore, the traditional rule-based interpretation method is adopted, which is cumbersome and prone to errors. Summary of the invention

[0003] The technical problem to be solved by the present invention is to provide a blood type auxiliary analysis method based on machine learning, which can correct and update the data intervals of each blood type classification image in real time, perform machine learning, update the agglutination detection model, and continuously improve the accuracy of blood type auxiliary detection.

[0004] In order to solve the above technical problems, the present invention adopts the following technical solutions: the present invention designs a blood type auxiliary judgment method based on machine learning, based on the blood sample amount, the blood typing reagent amount, and the reaction well cross-sectional size corresponding to the blood type detection process, the blood strong agglutination standard circle, the blood weak agglutination standard circle, and the reaction well boundary circle with the center of the reaction well cross-sectional center as the center of the circle, the area of ​​the area enclosed by the blood strong agglutination standard circle is initialized as the upper limit of the strong agglutination area, the area of ​​the area enclosed by the blood weak agglutination standard circle is initialized as the weak agglutination area lower limit and the weak agglutination area upper limit to form a weak agglutination area interval, the area of ​​the area enclosed by the reaction well boundary circle is initialized as the lower limit of the non-agglutination area, and the sample set is initialized as an empty set;

[0005] For each blood sample to be tested, for the target analysis sample after the blood sample to be tested reacts with the added corresponding blood typing reagent and precipitates for a preset time, obtain the target reaction well image corresponding to the target analysis sample, and perform the following steps:

[0006] Step A. Binarization processing is performed on the target reaction hole image with the pixel value of 0 corresponding to the red area and the pixel value of 255 corresponding to the non-red area to obtain a target binary image, and the area of ​​the target binary image in which the pixel value of the area to be analyzed is expanded outward from the center of the reaction hole cross section as the center of the circle, and the blood type analysis is performed in combination with the upper limit of the strong agglutination area, the interval of the weak agglutination area, and the lower limit of the non-agglutination area. If the blood type analysis is successful, the classification of the target analysis specimen corresponding to the target reaction hole image with respect to the strong agglutination classification, the weak agglutination classification, and the non-agglutination classification is determined, and then the blood type of the blood specimen to be tested is determined, and step D is entered; if the blood type analysis fails, step B is entered;

[0007] Step B. Apply the trained agglutination detection model to process the target reaction well image, obtain the probability that the target analysis specimen corresponding to the target reaction well image corresponds to the strong agglutination classification, the weak agglutination classification, and the non-agglutination classification, and take the classification with the maximum probability to form the classification of the target analysis specimen corresponding to the target reaction well image, and combine the added corresponding blood typing reagent to determine the blood type of the blood specimen to be tested. If the blood typing analysis is successful, then proceed to step C;

[0008] Step C. updating the upper limit of the strong agglutination area, the interval of the weak agglutination area or the lower limit of the non-agglutination area according to the classification of the target analysis specimen corresponding to the target reaction well image and the area of ​​the area to be analyzed, and then proceeding to step D;

[0009] Step D. Based on the preset attribute values ​​of the target reaction well image corresponding to the area to be analyzed, combined with the classification of the target analysis specimen corresponding to the target reaction well image, a blood type training sample is formed, added to and updated the sample set, and the number of samples belonging to the strong agglutination classification, the weak agglutination classification, and the non-agglutination classification in the sample set is determined to be greater than the preset training sample size. If yes, proceed to step E; otherwise, no further processing is performed;

[0010] Step E. Based on the sample set, the preset attribute values ​​of the reaction hole images in the blood type training samples corresponding to the areas to be analyzed are used as input, and the classification of the target analysis specimens corresponding to the reaction hole images in the blood type training samples is used as output, and the target classification network is trained to update the agglutination detection model.

[0011] As a preferred technical solution of the present invention: in step A, blood type analysis is performed as follows;

[0012] If the area of ​​the region to be analyzed is less than or equal to the upper limit of the strong agglutination area, the target analysis specimen is determined to correspond to the strong agglutination classification, and the blood type of the blood specimen to be tested is determined in combination with the corresponding blood typing reagent added, and the blood type analysis is successful;

[0013] If the area to be analyzed belongs to the weak agglutination area interval, the target analysis specimen is determined to correspond to the weak agglutination classification, and the blood type of the blood specimen to be tested is determined in combination with the corresponding blood typing reagent added, and the blood type analysis is successful;

[0014] If the area of ​​the region to be analyzed is greater than or equal to the lower limit of the non-agglutinated area, the target analysis specimen is determined to correspond to the non-agglutinated classification, and the blood type of the blood specimen to be tested is determined in combination with the corresponding blood typing reagent added, and the blood type analysis is successful;

[0015] Otherwise the blood type analysis fails.

[0016] As a preferred technical solution of the present invention: it also includes steps AB as follows, if the blood type analysis fails in step A, then go to step AB;

[0017] Step AB. If the area of ​​the region to be analyzed is greater than the upper limit of the strong agglutination area and less than the lower limit of the weak agglutination area, then define the strong agglutination classification and the weak agglutination classification as two to-be-classified classifications of the target analysis specimen corresponding to the target reaction well image; if the area of ​​the region to be analyzed is greater than the upper limit of the weak agglutination area and less than the lower limit of the non-agglutination area, then define the weak agglutination classification and the non-agglutination classification as two to-be-classified classifications of the target analysis specimen corresponding to the target reaction well image; then proceed to step B;

[0018] In the step B, the trained agglutination detection model is applied to process the target reaction well image to obtain the probabilities that the target analysis specimen corresponding to the target reaction well image corresponds to the strong agglutination classification, the weak agglutination classification, and the non-agglutination classification, and the classification with the maximum probability of the two to-be-classified classifications of the target analysis specimen corresponding to the target reaction well image is taken to form the classification of the target analysis specimen corresponding to the target reaction well image.

[0019] As a preferred technical solution of the present invention: in the step C, according to the classification of the target analysis specimen corresponding to the target reaction well image and the area of ​​the area to be analyzed, the upper limit of the strong agglutination area, the interval of the weak agglutination area or the lower limit of the non-agglutination area is updated in the following manner;

[0020] If the classification of the target analysis specimen corresponding to the target reaction well image is a strong agglutination classification, the upper limit of the strong agglutination area is updated according to the area of ​​the region to be analyzed;

[0021] If the classification of the target analysis specimen corresponding to the target reaction well image is weak agglutination classification, further analysis is performed to determine if the area of ​​the region to be analyzed is smaller than the lower limit of the weak agglutination area, then the lower limit of the weak agglutination area is updated with the area of ​​the region to be analyzed, and then the weak agglutination area interval is updated; if the area of ​​the region to be analyzed is larger than the upper limit of the weak agglutination area, then the upper limit of the weak agglutination area is updated with the area of ​​the region to be analyzed, and then the weak agglutination area interval is updated;

[0022] If the classification of the target analysis specimen corresponding to the target reaction well image is the non-agglutination classification, the non-agglutination area lower limit is updated according to the area of ​​the region to be analyzed.

[0023] As a preferred technical solution of the present invention: the preset attribute values ​​in step D include the area of ​​the region to be analyzed, and the average value of each pixel point in the region to be analyzed corresponding to the reaction well image with respect to the preset color space; then in step B, the average value of each pixel point in the region to be analyzed corresponding to the target reaction well image with respect to the preset color space is obtained, and combined with the area of ​​the region to be analyzed in the target reaction well image, the trained agglutination detection model is used for processing to obtain the probability that the target analysis specimen corresponding to the target reaction well image corresponds to the strong agglutination classification, the weak agglutination classification, and the non-agglutination classification.

[0024] As a preferred technical solution of the present invention: the target classification network in step E includes, from the input end to the output end, a front subnetwork, a first multi-scale feature extraction subnetwork, a first grid size reduction submodule, a second multi-scale feature extraction subnetwork, a second grid size reduction submodule, a third multi-scale feature extraction subnetwork, and an output subnetwork in sequence, wherein the input end of the front subnetwork constitutes the input end of the target classification network, and the output end of the output subnetwork constitutes the output end of the target classification network; wherein the first multi-scale feature extraction subnetwork includes, from the input end to the output end, a preset first number of first feature extraction submodules sequentially connected in series, the second multi-scale feature extraction subnetwork includes, from the input end to the output end, a preset second number of first feature extraction submodules sequentially connected in series, and the third multi-scale feature extraction subnetwork includes, from the input end to the output end, a preset third number of second feature extraction submodules sequentially connected in series;

[0025] The structures of the first feature extraction submodules are the same as each other. Each of the first feature extraction submodules includes an average pooling layer, a connection layer, and nine convolutional layers. In the structure of the first feature extraction submodule, one convolutional layer, three convolutional layers, and four convolutional layers are grouped, and each group connects the convolutional layers in series from the input end to the output end to form three branches. The output end of the average pooling layer is connected to the input end of the last convolutional layer to form the fourth branch. The input ends of the four branches are connected to each other to form the input end of the first feature extraction submodule. The output ends of the four branches are connected to each other, and the connected ends are connected to the input end of the connection layer. The output end of the connection layer constitutes the output end of the first feature extraction submodule.

[0026] The structures of the second feature extraction submodules are the same as each other. Each second feature extraction submodule includes an average pooling layer, three connection layers, and nine convolutional layers, among which one convolutional layer, three convolutional layers, and four convolutional layers are grouped; the input end and output end of the convolutional layer in the group where one convolutional layer is located constitute the input end and output end of the group; in the group where three convolutional layers are located, the input end of one convolutional layer constitutes the input end of the group, and the output end of the convolutional layer is connected to the input ends of the other two convolutional layers, and the output ends of the other two convolutional layers are connected to the input end of the first connection layer, and the output end of the first connection layer constitutes the output end of the group; in the group where four convolutional layers are located, two of the convolutional layers are connected from the input end to the output end. They are connected in series in sequence, the input end of the series constitutes the input end of the group, the output end of the series is connected to the input ends of the other two convolutional layers, the output ends of the other two convolutional layers are connected to the input end of the second connection layer, and the output end of the second connection layer constitutes the output end of the group; the output end of the average pooling layer is connected to the input end of the last convolutional layer to constitute a fourth group, the input end of the average pooling layer constitutes the input end of the fourth group, the output end of the last convolutional layer constitutes the output end of the fourth group, the input ends of the four groups are connected to each other to constitute the input end of the second feature extraction submodule, the output ends of the four groups are connected to each other, and the connected ends are connected to the input end of the third connection layer, and the output end of the third connection layer constitutes the output end of the second feature extraction submodule.

[0027] As a preferred technical solution of the present invention: the front subnetwork includes five convolutional layers and two maximum pooling layers, wherein the three convolutional layers are connected in series from the input end to the output end, the input end of the series connection constitutes the input end of the front subnetwork, the output end of the series connection is connected to the input end of one of the maximum pooling layers, the output end of the maximum pooling layer is connected in series with the remaining two convolutional layers and then connected to the input end of another maximum pooling layer, and the output end of the other maximum pooling layer constitutes the output end of the front subnetwork; the output subnetwork is connected in series with an average pooling layer, a random drop layer, a fully connected layer, and a Softmax layer from the input end to the output end, wherein the input layer of the average pooling layer constitutes the input end of the output subnetwork, and the output end of the Softmax layer constitutes the output end of the output subnetwork.

[0028] As a preferred technical solution of the present invention: the first grid size reduction submodule includes a maximum pooling layer, a connection layer, and four convolutional layers, wherein the three convolutional layers are connected in series from the input end to the output end, the series structure, another convolutional layer, and the maximum pooling layer are connected in parallel, and the parallel input end constitutes the input end of the first grid size reduction submodule, the parallel output end is connected to the input end of the connection layer, and the output end of the connection layer constitutes the output end of the first grid size reduction submodule; the second grid size reduction submodule includes a maximum pooling layer, a connection layer, and six convolutional layers, wherein two convolutional layers are connected in series from the input end to the output end to constitute a first series structure, and the other four convolutional layers are connected in series from the input end to the output end to constitute a second series structure, the first series structure, the second series structure, and the maximum pooling layer are connected in parallel, and the parallel input end constitutes the input end of the second grid size reduction submodule, the parallel output end is connected to the input end of the connection layer, and the output end of the connection layer constitutes the output end of the second grid size reduction submodule.

[0029] Corresponding to the above, the technical problem that the present invention also needs to solve is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, each step in a blood type assisted judgment method based on machine learning is implemented.

[0030] The corresponding technical problem to be solved by the present invention is to provide a computer-readable storage medium on which a computer program is stored, characterized in that when the computer program is executed by a processor, each step in a blood type auxiliary analysis method based on machine learning is implemented.

[0031] The blood type auxiliary judgment method based on machine learning described in the present invention adopts the above technical solution and has the following technical effects compared with the prior art:

[0032] The present invention designs a blood type auxiliary judgment method based on machine learning, which analyzes the strong and weak agglutination areas of blood during the blood type detection process, and uses the initial strong agglutination area upper limit, weak agglutination area interval, and non-agglutination area lower limit of the area enclosed by the blood strong agglutination standard circle, the blood weak agglutination standard circle, and the reaction hole boundary circle respectively, to sequentially analyze each blood sample to be detected, and executes a complementary processing method of area comparison and machine learning, while continuously optimizing the agglutination detection model, continuously filling and improving the gaps between different types of agglutination area intervals, and continuously promoting each other between the two, while improving the accuracy of blood type auxiliary judgment by both area comparison and machine learning, overcoming the shortcomings of the existing technology, and effectively improving the work efficiency of blood type auxiliary judgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1It is a flow chart of a batch authentication method combining designated time and privacy protection designed by the present invention;

[0034] Figure 2 It is a schematic diagram of the target classification network involved in the design of the present invention. DETAILED DESCRIPTION

[0035] The specific implementation modes of the present invention will be further described in detail below in conjunction with the accompanying drawings.

[0036] The present invention designs a blood type auxiliary judgment method based on machine learning. Based on the blood sample amount, blood typing reagent amount, and reaction hole cross-sectional dimensions corresponding to the blood type detection process, a blood strong agglutination standard circle, a blood weak agglutination standard circle, and a reaction hole boundary circle with the center of the reaction hole cross-sectional center as the center of the circle, the area of ​​the area enclosed by the blood strong agglutination standard circle is initialized as the upper limit of the strong agglutination area, the area of ​​the area enclosed by the blood weak agglutination standard circle is initialized as the weak agglutination area lower limit and the weak agglutination area upper limit to form a weak agglutination area interval, the area of ​​the area enclosed by the reaction hole boundary circle is initialized as the lower limit of the non-agglutination area, and the sample set is initialized as an empty set.

[0037] In actual application, the blood type auxiliary judgment method is designed to sequentially target each blood sample to be tested, and the target analysis sample after the blood sample to be tested reacts with the added corresponding blood type typing reagent and precipitates for a preset time, obtains the target reaction hole image corresponding to the target analysis sample, and Figure 1 As shown, perform the following steps A to E.

[0038] Step A. Binarize the target reaction hole image with the red area corresponding to the pixel value 0 and the non-red area corresponding to the pixel value 255 to obtain a target binary image, obtain the area of ​​the area to be analyzed with the pixel value of 0 expanding outward from the center of the reaction hole cross section in the target binary image, and perform blood type analysis in combination with the upper limit of the strong agglutination area, the interval of the weak agglutination area, and the lower limit of the non-agglutination area. If the blood type analysis is successful, the classification of the target analysis specimen corresponding to the target reaction hole image with respect to the strong agglutination classification, the weak agglutination classification, and the non-agglutination classification is determined, and then the blood type of the blood specimen to be tested is determined, and step D is entered; if the blood type analysis fails, step AB is entered.

[0039] In practical applications, in the above step A, blood type analysis is performed specifically in the following manner.

[0040] If the area of ​​the region to be analyzed is less than or equal to the upper limit of the strong agglutination area, the target analysis specimen is determined to correspond to the strong agglutination classification, and the blood type of the blood specimen to be tested is determined in combination with the corresponding blood typing reagent added, and the blood type analysis is successful;

[0041] If the area to be analyzed belongs to the weak agglutination area interval, the target analysis specimen is determined to correspond to the weak agglutination classification, and the blood type of the blood specimen to be tested is determined in combination with the corresponding blood typing reagent added, and the blood type analysis is successful;

[0042] If the area of ​​the region to be analyzed is greater than or equal to the lower limit of the non-agglutinated area, the target analysis specimen is determined to correspond to the non-agglutinated classification, and the blood type of the blood specimen to be tested is determined in combination with the corresponding blood typing reagent added, and the blood type analysis is successful;

[0043] Otherwise the blood type analysis fails.

[0044] Step AB. If the area of ​​the region to be analyzed is greater than the upper limit of the strong agglutination area and less than the lower limit of the weak agglutination area, then the strong agglutination classification and the weak agglutination classification are defined as two to-be-classified categories of the target analysis specimen corresponding to the target reaction well image; if the area of ​​the region to be analyzed is greater than the upper limit of the weak agglutination area and less than the lower limit of the non-agglutination area, then the weak agglutination classification and the non-agglutination classification are defined as two to-be-classified categories of the target analysis specimen corresponding to the target reaction well image; then proceed to step B.

[0045] Step B. Apply the trained agglutination detection model to process the target reaction well image, obtain the probabilities that the target analysis specimen corresponding to the target reaction well image corresponds to the strong agglutination classification, the weak agglutination classification, and the non-agglutination classification, respectively, and take the classification with the maximum probability of the two to-be-classified classifications of the target analysis specimen corresponding to the target reaction well image to form the classification of the target analysis specimen corresponding to the target reaction well image, and combine the added corresponding blood typing reagent to determine the blood type of the blood specimen to be tested, the blood typing analysis is successful, and then proceed to step C.

[0046] Step C. According to the classification of the target analysis specimen corresponding to the target reaction well image and the area of ​​the region to be analyzed, the upper limit of the strong agglutination area, the interval of the weak agglutination area or the lower limit of the non-agglutination area is updated as follows, and then the process proceeds to step D.

[0047] If the classification of the target analysis specimen corresponding to the target reaction well image is a strong agglutination classification, the upper limit of the strong agglutination area is updated according to the area of ​​the region to be analyzed;

[0048] If the classification of the target analysis specimen corresponding to the target reaction well image is weak agglutination classification, further analysis is performed to determine if the area of ​​the region to be analyzed is smaller than the lower limit of the weak agglutination area, then the lower limit of the weak agglutination area is updated with the area of ​​the region to be analyzed, and then the weak agglutination area interval is updated; if the area of ​​the region to be analyzed is larger than the upper limit of the weak agglutination area, then the upper limit of the weak agglutination area is updated with the area of ​​the region to be analyzed, and then the weak agglutination area interval is updated;

[0049] If the classification of the target analysis specimen corresponding to the target reaction well image is the non-agglutination classification, the non-agglutination area lower limit is updated according to the area of ​​the region to be analyzed.

[0050] Step D. Based on the preset attribute values ​​corresponding to the area to be analyzed in the target reaction well image and the classification of the target analysis specimen corresponding to the target reaction well image, a blood type training sample is formed, added to and updated the sample set, and it is determined whether the number of samples belonging to the strong agglutination classification, the weak agglutination classification, and the non-agglutination classification in the sample set is greater than the preset training sample size. If so, proceed to step E; otherwise, no further processing is performed.

[0051] In practical applications, the preset attribute values ​​in step D include the area of ​​the region to be analyzed and the average value of each pixel in the region to be analyzed corresponding to the reaction well image with respect to the preset color space; then in step B, the average value of each pixel in the region to be analyzed corresponding to the target reaction well image with respect to the preset color space is obtained, and combined with the area of ​​the region to be analyzed in the target reaction well image, the trained agglutination detection model is used for processing to obtain the probability that the target analysis specimen corresponding to the target reaction well image corresponds to the strong agglutination classification, the weak agglutination classification, and the non-agglutination classification. The preset color space here is such as the RGB color space or the HSV color space.

[0052] Step E. Based on the sample set, the preset attribute values ​​of the reaction hole images in the blood type training samples corresponding to the areas to be analyzed are used as input, and the classification of the target analysis specimens corresponding to the reaction hole images in the blood type training samples is used as output, and the target classification network is trained to update the agglutination detection model.

[0053] like Figure 2 As shown, the target classification network involved in step E includes, from the input end to the output end, a front subnetwork, a first multi-scale feature extraction subnetwork, a first grid size reduction submodule, a second multi-scale feature extraction subnetwork, a second grid size reduction submodule, a third multi-scale feature extraction subnetwork, and an output subnetwork, wherein the input end of the front subnetwork constitutes the input end of the target classification network, and the output end of the output subnetwork constitutes the output end of the target classification network; wherein, the first multi-scale feature extraction subnetwork includes, from the input end to the output end, a preset first number of first feature extraction submodules connected in series in sequence, the second multi-scale feature extraction subnetwork includes, from the input end to the output end, a preset second number of first feature extraction submodules connected in series in sequence, and the third multi-scale feature extraction subnetwork includes, from the input end to the output end, a preset third number of second feature extraction submodules connected in series in sequence.

[0054] The structures of the first feature extraction submodules are the same as each other, such as Figure 2As shown, each first feature extraction submodule includes an average pooling layer, a connection layer, and nine convolutional layers. In the structure of the first feature extraction submodule, one convolutional layer, three convolutional layers, and four convolutional layers are grouped, and each group connects the convolutional layers in series from the input end to the output end to form three branches. The output end of the average pooling layer is connected to the input end of the last convolutional layer to form the fourth branch. The input ends of the four branches are connected to each other to form the input end of the first feature extraction submodule. The output ends of the four branches are connected to each other, and the connected ends are connected to the input end of the connection layer. The output end of the connection layer constitutes the output end of the first feature extraction submodule.

[0055] The structures of the second feature extraction submodules are the same as each other, such as Figure 2 As shown, each second feature extraction submodule includes an average pooling layer, three connection layers, and nine convolutional layers, among which one convolutional layer, three convolutional layers, and four convolutional layers are grouped; the input end and output end of the convolutional layer in the group where one convolutional layer is located constitute the input end and output end of the group; in the group where three convolutional layers are located, the input end of one convolutional layer constitutes the input end of the group, the output end of the convolutional layer is connected to the input ends of the other two convolutional layers, the output ends of the other two convolutional layers are connected to the input end of the first connection layer, and the output end of the first connection layer constitutes the output end of the group; in the group where four convolutional layers are located, two of the convolutional layers are connected in series from the input end to the output end, and the series connection The input end of the series connection constitutes the input end of the group, the output end of the series connection is connected to the input end of the other two convolutional layers, the output ends of the other two convolutional layers are connected to the input end of the second connection layer, and the output end of the second connection layer constitutes the output end of the group; the output end of the average pooling layer is connected to the input end of the last convolutional layer to constitute the fourth group, the input end of the average pooling layer constitutes the input end of the fourth group, the output end of the last convolutional layer constitutes the output end of the fourth group, the input ends of the four groups are connected to each other to constitute the input end of the second feature extraction submodule, the output ends of the four groups are connected to each other, and the connected ends are connected to the input end of the third connection layer, and the output end of the third connection layer constitutes the output end of the second feature extraction submodule.

[0056] In the design of the above target classification network, Figure 2As shown, the specific design of the front subnetwork includes five convolutional layers and two maximum pooling layers, wherein the three convolutional layers are connected in series from the input end to the output end, the input end of the series constitutes the input end of the front subnetwork, the output end of the series is connected to the input end of one of the maximum pooling layers, the output end of the maximum pooling layer is connected in series with the remaining two convolutional layers and then connected to the input end of another maximum pooling layer, and the output end of the other maximum pooling layer constitutes the output end of the front subnetwork; the output subnetwork is connected in series with an average pooling layer, a random drop layer, a fully connected layer, and a Softmax layer from the input end to the output end, wherein the input layer of the average pooling layer constitutes the input end of the output subnetwork, and the output end of the Softmax layer constitutes the output end of the output subnetwork.

[0057] And in applications such as Figure 2 As shown, the first grid size reduction submodule includes a maximum pooling layer, a connection layer, and four convolutional layers, wherein the three convolutional layers are sequentially connected in series from the input end to the output end, the series structure, another convolutional layer, and the maximum pooling layer are connected in parallel, and the parallel input end constitutes the input end of the first grid size reduction submodule, the parallel output end is connected to the input end of the connection layer, and the output end of the connection layer constitutes the output end of the first grid size reduction submodule; the second grid size reduction submodule includes a maximum pooling layer, a connection layer, and six convolutional layers, wherein two convolutional layers are sequentially connected in series from the input end to the output end to constitute a first series structure, and the other four convolutional layers are sequentially connected in series from the input end to the output end to constitute a second series structure, the first series structure, the second series structure, and the maximum pooling layer are connected in parallel, and the parallel input end constitutes the input end of the second grid size reduction submodule, the parallel output end is connected to the input end of the connection layer, and the output end of the connection layer constitutes the output end of the second grid size reduction submodule.

[0058] The blood type assisted identification method designed by the above invention is applied in practice, and a corresponding computer device is designed, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements each step of the blood type assisted identification method based on machine learning designed by the present invention when executing the computer program; accordingly, a computer-readable storage medium is designed, on which a computer program is stored, and the computer program implements each step of the blood type assisted identification method based on machine learning when executed by the processor.

[0059] The present invention is designed to analyze the strong and weak agglutination areas of blood during the blood typing process, and uses the initial strong agglutination area upper limit, weak agglutination area interval, and non-agglutination area lower limit of the areas enclosed by the blood strong agglutination standard circle, the blood weak agglutination standard circle, and the reaction hole boundary circle respectively, to sequentially analyze each blood specimen to be tested, and implements a complementary processing method of area comparison and machine learning. While continuously optimizing the agglutination detection model, the blanks between different types of agglutination area intervals are continuously filled and improved. Through continuous mutual promotion between the two, the accuracy of blood typing auxiliary judgment by both area comparison and machine learning is improved, the shortcomings of the prior art are overcome, and the work efficiency of blood typing auxiliary judgment is effectively improved.

[0060] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A blood type auxiliary judgment method based on machine learning, characterized in that: Based on the blood sample volume, the blood typing reagent volume, and the cross-sectional size of the reaction well, a blood strong agglutination standard circle, a blood weak agglutination standard circle, and a reaction well boundary circle with the center of the reaction well cross-sectional center as the center of the circle in the blood typing detection process, the area of ​​the region enclosed by the blood strong agglutination standard circle is initialized as the upper limit of the strong agglutination area, the area of ​​the region enclosed by the blood weak agglutination standard circle is initialized as the weak agglutination area lower limit and the weak agglutination area upper limit to form a weak agglutination area interval, the area of ​​the region enclosed by the reaction well boundary circle is initialized as the lower limit of the non-agglutination area, and the sample set is initialized as an empty set; For each blood sample to be tested, for the target analysis sample after the blood sample to be tested reacts with the added corresponding blood typing reagent and precipitates for a preset time, obtain the target reaction well image corresponding to the target analysis sample, and perform the following steps: Step A. Binarization processing is performed on the target reaction well image with the pixel value of 0 corresponding to the red area and the pixel value of 255 corresponding to the non-red area to obtain a target binary image, and the area of ​​the area to be analyzed with the pixel value of 0 extending outward from the center of the reaction well cross section in the target binary image is obtained, and the blood type analysis is performed in combination with the upper limit of the strong agglutination area, the interval of the weak agglutination area, and the lower limit of the non-agglutination area. If the blood type analysis is successful, the classification of the target analysis specimen corresponding to the target reaction well image with respect to the strong agglutination classification, the weak agglutination classification, and the non-agglutination classification is determined, and then the blood type of the blood specimen to be tested is determined, and step D is entered; If the blood type analysis fails, proceed to step B; Step B. Apply the trained agglutination detection model to process the target reaction well image, obtain the probability that the target analysis specimen corresponding to the target reaction well image corresponds to the strong agglutination classification, the weak agglutination classification, and the non-agglutination classification, and take the classification with the maximum probability to form the classification of the target analysis specimen corresponding to the target reaction well image, and combine the added corresponding blood typing reagent to determine the blood type of the blood specimen to be tested. If the blood typing analysis is successful, then proceed to step C; Step C. updating the upper limit of the strong agglutination area, the interval of the weak agglutination area or the lower limit of the non-agglutination area according to the classification of the target analysis specimen corresponding to the target reaction well image and the area of ​​the area to be analyzed, and then proceeding to step D; Step D. Based on the preset attribute values ​​of the target reaction well image corresponding to the area to be analyzed, combined with the classification of the target analysis specimen corresponding to the target reaction well image, a blood type training sample is formed, added to and updated the sample set, and the number of samples belonging to the strong agglutination classification, the weak agglutination classification, and the non-agglutination classification in the sample set is determined to be greater than the preset training sample size. If yes, proceed to step E; otherwise, no further processing is performed; Step E. Based on the sample set, the preset attribute values ​​of the reaction hole images in the blood type training samples corresponding to the areas to be analyzed are used as input, and the classification of the target analysis specimens corresponding to the reaction hole images in the blood type training samples is used as output, and the target classification network is trained to update the agglutination detection model.

2. According to claim 1, a blood type auxiliary judgment method based on machine learning is characterized in that: In step A, blood type analysis is performed as follows; If the area of ​​the region to be analyzed is less than or equal to the upper limit of the strong agglutination area, the target analysis specimen is determined to correspond to the strong agglutination classification, and the blood type of the blood specimen to be tested is determined in combination with the corresponding blood typing reagent added, and the blood type analysis is successful; If the area to be analyzed belongs to the weak agglutination area interval, the target analysis specimen is determined to correspond to the weak agglutination classification, and the blood type of the blood specimen to be tested is determined in combination with the corresponding blood typing reagent added, and the blood type analysis is successful; If the area of ​​the region to be analyzed is greater than or equal to the lower limit of the non-agglutinated area, the target analysis specimen is determined to correspond to the non-agglutinated classification, and the blood type of the blood specimen to be tested is determined in combination with the corresponding blood typing reagent added, and the blood type analysis is successful; Otherwise the blood typing analysis fails.

3. According to claim 1, a blood type auxiliary judgment method based on machine learning is characterized in that: The method further includes steps AB as follows: if the blood type analysis fails in step A, the method proceeds to step AB; Step AB. If the area of ​​the region to be analyzed is greater than the upper limit of the strong agglutination area and less than the lower limit of the weak agglutination area, then define the strong agglutination classification and the weak agglutination classification as two to-be-classified classifications of the target analysis specimen corresponding to the target reaction well image; if the area of ​​the region to be analyzed is greater than the upper limit of the weak agglutination area and less than the lower limit of the non-agglutination area, then define the weak agglutination classification and the non-agglutination classification as two to-be-classified classifications of the target analysis specimen corresponding to the target reaction well image; then proceed to step B; In the step B, the trained agglutination detection model is applied to process the target reaction well image to obtain the probabilities that the target analysis specimen corresponding to the target reaction well image corresponds to the strong agglutination classification, the weak agglutination classification, and the non-agglutination classification, and the classification with the maximum probability of the two to-be-classified classifications of the target analysis specimen corresponding to the target reaction well image is taken to form the classification of the target analysis specimen corresponding to the target reaction well image.

4. According to claim 1, a blood type auxiliary judgment method based on machine learning is characterized in that: In the step C, according to the classification of the target analysis specimen corresponding to the target reaction well image and the area of ​​the area to be analyzed, the upper limit of the corresponding strong agglutination area, the interval of the weak agglutination area or the lower limit of the non-agglutination area is updated in the following manner; If the classification of the target analysis specimen corresponding to the target reaction well image is a strong agglutination classification, the upper limit of the strong agglutination area is updated according to the area of ​​the region to be analyzed; If the classification of the target analysis specimen corresponding to the target reaction well image is weak agglutination classification, further analysis is performed to determine if the area of ​​the region to be analyzed is smaller than the lower limit of the weak agglutination area, then the lower limit of the weak agglutination area is updated with the area of ​​the region to be analyzed, and then the weak agglutination area interval is updated; if the area of ​​the region to be analyzed is larger than the upper limit of the weak agglutination area, then the upper limit of the weak agglutination area is updated with the area of ​​the region to be analyzed, and then the weak agglutination area interval is updated; If the classification of the target analysis specimen corresponding to the target reaction well image is the non-agglutination classification, the non-agglutination area lower limit is updated according to the area of ​​the region to be analyzed.

5. According to claim 1, a blood type auxiliary judgment method based on machine learning is characterized in that: The preset attribute values ​​in step D include the area of ​​the region to be analyzed and the average value of each pixel point in the region to be analyzed corresponding to the reaction well image with respect to the preset color space; then in step B, the average value of each pixel point in the region to be analyzed corresponding to the target reaction well image with respect to the preset color space is obtained, and combined with the area of ​​the region to be analyzed in the target reaction well image, the trained agglutination detection model is used for processing to obtain the probability that the target analysis specimen corresponding to the target reaction well image corresponds to the strong agglutination classification, the weak agglutination classification, and the non-agglutination classification.

6. According to claim 1, a blood type auxiliary identification method based on machine learning is characterized in that: The target classification network in step E includes, from the input end to the output end, a pre-subnetwork, a first multi-scale feature extraction subnetwork, a first grid size reduction submodule, a second multi-scale feature extraction subnetwork, a second grid size reduction submodule, a third multi-scale feature extraction subnetwork, and an output subnetwork in order. The input end of the pre-subnetwork constitutes the input end of the target classification network, and the output end of the output subnetwork constitutes the output end of the target classification network; wherein, the first multi-scale feature extraction subnetwork includes, from the input end to the output end, a preset first number of first feature extraction submodules connected in series in sequence, the second multi-scale feature extraction subnetwork includes, from the input end to the output end, a preset second number of first feature extraction submodules connected in series in sequence, and the third multi-scale feature extraction subnetwork includes, from the input end to the output end, a preset third number of second feature extraction submodules connected in series in sequence; The structures of the first feature extraction submodules are the same as each other. Each of the first feature extraction submodules includes an average pooling layer, a connection layer, and nine convolutional layers. In the structure of the first feature extraction submodule, one convolutional layer, three convolutional layers, and four convolutional layers are grouped, and each group connects the convolutional layers in series from the input end to the output end to form three branches. The output end of the average pooling layer is connected to the input end of the last convolutional layer to form the fourth branch. The input ends of the four branches are connected to each other to form the input end of the first feature extraction submodule. The output ends of the four branches are connected to each other, and the connected ends are connected to the input end of the connection layer. The output end of the connection layer constitutes the output end of the first feature extraction submodule. The structures of the second feature extraction submodules are the same as each other. Each second feature extraction submodule includes an average pooling layer, three connection layers, and nine convolutional layers, among which one convolutional layer, three convolutional layers, and four convolutional layers are grouped; the input end and output end of the convolutional layer in the group where one convolutional layer is located constitute the input end and output end of the group; in the group where three convolutional layers are located, the input end of one convolutional layer constitutes the input end of the group, and the output end of the convolutional layer is connected to the input ends of the other two convolutional layers, and the output ends of the other two convolutional layers are connected to the input end of the first connection layer, and the output end of the first connection layer constitutes the output end of the group; in the group where four convolutional layers are located, two of the convolutional layers are connected from the input end to the output end. They are connected in series in sequence, the input end of the series constitutes the input end of the group, the output end of the series is connected to the input ends of the other two convolutional layers, the output ends of the other two convolutional layers are connected to the input end of the second connection layer, and the output end of the second connection layer constitutes the output end of the group; the output end of the average pooling layer is connected to the input end of the last convolutional layer to constitute a fourth group, the input end of the average pooling layer constitutes the input end of the fourth group, the output end of the last convolutional layer constitutes the output end of the fourth group, the input ends of the four groups are connected to each other to constitute the input end of the second feature extraction submodule, the output ends of the four groups are connected to each other, and the connected ends are connected to the input end of the third connection layer, and the output end of the third connection layer constitutes the output end of the second feature extraction submodule.

7. The blood type auxiliary identification method based on machine learning according to claim 6 is characterized by: The front sub-network includes five convolutional layers and two maximum pooling layers, wherein the three convolutional layers are sequentially connected in series from the input end to the output end, the input end of the series connection constitutes the input end of the front sub-network, the output end of the series connection is connected to the input end of one of the maximum pooling layers, the output end of the maximum pooling layer is sequentially connected in series with the remaining two convolutional layers and then connected to the input end of another maximum pooling layer, and the output end of the other maximum pooling layer constitutes the output end of the front sub-network; the output sub-network is sequentially connected in series with an average pooling layer, a random drop layer, a fully connected layer, and a Softmax layer from the input end to the output end, wherein the input layer of the average pooling layer constitutes the input end of the output sub-network, and the output end of the Softmax layer constitutes the output end of the output sub-network.

8. The blood type auxiliary identification method based on machine learning according to claim 6 is characterized by: The first grid size reduction submodule includes a maximum pooling layer, a connection layer, and four convolutional layers, wherein the three convolutional layers are sequentially connected in series from the input end to the output end, the series structure, another convolutional layer, and the maximum pooling layer are connected in parallel, and the parallel input end constitutes the input end of the first grid size reduction submodule, the parallel output end is connected to the input end of the connection layer, and the output end of the connection layer constitutes the output end of the first grid size reduction submodule; The second grid size reduction submodule includes a maximum pooling layer, a connection layer, and six convolutional layers, wherein two convolutional layers are connected in series from the input end to the output end to form a first series structure, and the other four convolutional layers are connected in series from the input end to the output end to form a second series structure. The first series structure, the second series structure, and the maximum pooling layer are connected in parallel, and the parallel input end constitutes the input end of the second grid size reduction submodule, the parallel output end is connected to the input end of the connection layer, and the output end of the connection layer constitutes the output end of the second grid size reduction submodule.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images