Blood leukocyte recognition AI training method, calculation processing device and storage device

Abstract

A blood leukocyte recognition AI training method comprises the following steps: diluting blood to obtain a blood sample; flatly spreading the blood sample to enable white blood cells to suspend in the liquid and to present a natural state; shooting the blood sample to obtain a picture; carrying out AI training by using the obtained pictures to obtain a leukocyte AI feature data set A; aI training is carried out on the leukocyte picture in the suspension state to obtain an AI feature data set A, and AI detection of blood in the suspension state can be effectively supported and recognized. The AI feature data set A is obtained by carrying out AI training on the leukocyte picture in the suspension state, AI detection of blood in the suspension state can be effectively supported and recognized, and the AI recognition function in a new-mode blood detection technology is effectively supported by tiling and imaging blood cells.

Description

Technical Field

[0001] This application belongs to the technical field of analysis of formed elements based on microscopic magnified images, and particularly relates to a method for training AI for blood leukocyte recognition, a computing and processing device, and a storage device. Background Art

[0002] White blood cells, commonly known as leukocytes, are the "guards" in the human body's fight against diseases. When germs invade the human body, white blood cells can pass through the capillary wall by deforming, concentrate in the invaded part by the germs, and surround and phagocytize the germs.

[0003] Adult range: (4.0 - 10.0)×10^9 cells / L. Various reasons can cause the white blood cell count to be high or low. Reasons for high white blood cell count: 1. Trauma or tissue necrosis, such as extensive burns. 2. Systemic or local infection, such as acute tonsillitis, acute appendicitis, etc. 3. A significant increase should alert to leukemia. 4. Certain tumors can also cause an increase in white blood cells.

[0004] Reasons for low white blood cell count: 1. Autoimmune system diseases: such as rheumatoid arthritis, systemic lupus erythematosus, etc. 2. Hypersplenism: such as cirrhosis, portal vein embolism, heart failure, etc. 3. Bone marrow lesions: such as leukemia, myeloma, bone marrow metastatic cancer, etc. 4. Taking certain drugs: such as cyclophosphamide, antithyroid drugs, antitumor drugs, etc. 5. Chemical poisons: such as benzene, xylene, chemical paint, etc. 6. X-ray ionization radiation.

[0005] Blood routine examination is one of the most basic laboratory tests in clinical practice. The items of blood routine examination include red blood cells, white blood cells, hemoglobin, and platelet count, etc. And white blood cells are an important item in blood routine examination.

[0006] Blood routine examination usually adopts blood cell counting: a hematology analyzer will automatically calculate the collected blood specimen, including indicators such as red blood cell count, white blood cell count, platelet count, hematocrit, mean corpuscular volume, and mean corpuscular hemoglobin content.

[0007] The detection principles of blood routine examination include impedance method, radiofrequency conductivity method, spectrophotometry, and flow cytometry method.

[0008] The applicant has proposed a series of Chinese patents, such as

[0009] 1. CN2020112669290, "Cell Analysis Method and System and Quantitative Method and System";

[0010] 2. CN2020112669182, "Imaging Method and System for Cell Suspension Samples and Kit";

[0011] 3. CN2020112669182, "Imaging Method and System for Cell Suspension Samples and Kit"

[0012] 4. CN2022104799126, "Fast Focusing Method for Microscopic Image Acquisition Device and Microscopic Image Acquisition Method"

[0013] Use a brand - new technical solution to measure the content of cells in blood, including the analysis, identification, and counting of white blood cells in blood.

[0014] Artificial intelligence image recognition technology is an image - processing technology based on artificial intelligence algorithms. By analyzing and learning images, it realizes the recognition and understanding of image content.

[0015] Data assets refer to data resources that are owned or controlled by individuals or enterprises, can bring future economic benefits to enterprises, and are recorded in physical or electronic ways. Data assets are data sets in the cyberspace that have data ownership rights (exploration rights, usage rights, ownership rights), are valuable, measurable, and readable.

[0016] The content of white blood cells in blood is relatively low. White blood cells are further classified into types such as neutrophils, eosinophils, basophils, monocytes, and lymphocytes. When using AI technology to identify and classify white blood cells, it is extremely labor - intensive to collect, present, and label white blood cells. The finally trained AI feature data set is an important digital asset, which can be traded as a general accessory in the industry and can improve the overall level of the industry.

[0017] Using direct imaging of blood suspension, compared with dry - smear microscopy, the AI training is different. The AI feature data sets trained by various different detection methods cannot be applied to the identification of white blood cells in direct imaging of blood suspension. This application proposes an AI training method for "blood analysis technology for blood suspension microscopy", especially for white blood cell analysis technology. Summary of the Invention

[0018] In this application, the inventor proposes that by performing AI training on pictures of white blood cells in a suspended state, the obtained AI feature data set A can effectively support the AI detection of suspended - state blood. By flattening the blood cells for imaging, it effectively supports the AI recognition function in the new - mode blood detection technology.

[0019] The technical solution of this application to solve the above - mentioned technical problems is an AI training method for blood white blood cell recognition, including diluting blood to obtain a blood sample; flattening the blood sample so that white blood cells are suspended in the liquid and presented in a natural state; taking pictures of the blood sample to obtain pictures; and performing AI training on the obtained pictures to obtain the white blood cell AI feature data set A.

[0020] The blood sample is laid flat so that the red blood cells are in a single-layer flat state.

[0021] The stacked height of the red blood cells when the blood sample is laid flat does not exceed the diameter of the white blood cells.

[0022] The AI training includes image preprocessing and AI training processing. The image preprocessing outputs the labeled white blood cells, and based on the labeled white blood cells, the AI training processing outputs the white blood cell AI feature dataset.

[0023] A staining agent is added to the blood to stain the white blood cells.

[0024] A liquid staining agent is added to the blood. The liquid staining agent dilutes the blood and stains the nuclei of the white blood cells.

[0025] Using the white blood cell AI feature dataset A as the feature dataset, one or more pictures input to the AI recognition of the said picture are recognized by the AI software to obtain marked pictures. The marked pictures include the labeled white blood cells. The marked pictures are manually reviewed to form manually reviewed marked pictures, and the manually reviewed marked pictures are used for AI training to obtain the white blood cell AI feature dataset B.

[0026] The described AI training method for blood white blood cell recognition includes any one of the following technical features: TA10: The white blood cells are neutrophilic granulocytes, and the white blood cell feature dataset A or B includes the features of stained neutrophilic granulocytes; TA20: The white blood cells are lymphocytes, and the white blood cell feature dataset A or B includes the features of stained lymphocytes; TA30: The white blood cells are eosinophilic granulocytes, and the white blood cell feature dataset A or B includes the features of stained eosinophilic granulocytes; TA40: The white blood cells are basophilic granulocytes, and the white blood cell feature dataset A or B includes the features of stained basophilic granulocytes; TA50: The white blood cells are neutrophilic stab granulocytes, and the white blood cell feature dataset A or B includes the features of stained neutrophilic stab granulocytes; TA60: The white blood cells are neutrophilic segmented granulocytes, and the white blood cell feature dataset A or B includes the features of stained neutrophilic segmented granulocytes; TA70: The white blood cells are monocytes, and the white blood cell feature dataset A or B includes the features of stained monocytes.

[0027] The described blood white blood cell recognition AI training method includes any one of the following technical features: TE10: The white blood cells are neutrophilic granulocytes, and the white blood cell feature dataset A or B includes unstained neutrophilic granulocyte features; TE20: The white blood cells are lymphocytes, and the white blood cell feature dataset A or B includes unstained lymphocyte features; TE30: The white blood cells are eosinophilic granulocytes, and the white blood cell feature dataset A or B includes unstained eosinophilic granulocyte features; TE40: The white blood cells are basophilic granulocytes, and the white blood cell feature dataset A or B includes unstained basophilic granulocyte features; TE50: The white blood cells are neutrophilic stab granulocytes, and the white blood cell feature dataset A or B includes unstained neutrophilic stab granulocyte features; TE60: The white blood cells are neutrophilic segmented granulocytes, and the white blood cell feature dataset A or B includes unstained neutrophilic segmented granulocyte features; TE70: The white blood cells are monocytes, and the white blood cell feature dataset A or B includes unstained monocyte features.

[0028] The white blood cell feature dataset A is a white blood cell feature dataset, and the white blood cell feature dataset B is a sub-item white blood cell feature dataset.

[0029] The technical solution of the present application to solve the above technical problem can also be a computing processing device, and the computing processing device includes the above-mentioned method for blood white blood cell recognition AI training.

[0030] The technical solution of the present application to solve the above technical problem can also be a computing processing device, and the memory of the computing processing device includes the above-mentioned white blood cell AI feature dataset A or white blood cell AI feature dataset B.

[0031] The technical solution of the present application to solve the above technical problem can also be a data storage device, and the data storage includes the above-mentioned white blood cell AI feature dataset A or white blood cell AI feature dataset B.

[0032] The technical effects of the above technical solutions include: By performing AI training on the pictures of white blood cells in a suspended state, the obtained AI feature dataset A can effectively support the recognition of suspended state blood AI detection.

[0033] The technical effects of the above technical solutions include: By flattening the blood cells into an image, making the white blood cells protrude from the red blood cells, it effectively supports the AI recognition function in the new mode blood detection technology.

[0034] The technical effects of the above technical solutions include: By flattening the blood red blood cells into a single layer image, the recognition rate of white blood cells is improved.

[0035] The technical effects of the above technical solution include: by laying multiple layers of blood red blood cells flat, with the laying height not exceeding the diameter of white blood cells for imaging, the distribution density of white blood cells is increased, and at the same time, an interference pattern between white blood cells and red blood cells is provided, improving the recognition rate of white blood cells in the actual state.

[0036] The technical effects of the above technical solution include: through staining, the inner core characteristics of white blood cells are highlighted, improving the recognition accuracy of white blood cell categories.

[0037] The technical effects of the above technical solution include: through secondary training, on the basis of identifying white blood cells, white blood cell classification recognition is carried out, turning the impossible into engineering achievable, making it possible to classify and recognize suspended white blood cells, and enabling the industrial application of the suspended white blood cell classification technology.

[0038] The technical effects of the above technical solution include: encapsulating the trained feature data into white blood cell feature datasets A and B, which respectively correspond to white blood cell recognition and white blood cell type differentiation, expanding the application scenarios.

[0039] The technical effects of the above technical solution include: encapsulating the trained feature data into white blood cell feature datasets A and B as independent data, which can be sold separately, reducing the engineering difficulty of the entire industry, enabling new technologies to be promoted and implemented as soon as possible, and benefiting patients. Description of the Drawings

[0040] Figure 1 is a schematic diagram of the steps of an AI training method for blood white blood cell recognition;

[0041] Figure 2 is a schematic diagram of the steps of an AI training method for blood white blood cell recognition;

[0042] Figure 3 is a schematic diagram of blood cells in a single-layer suspended state;

[0043] Figure 4 is a schematic diagram of blood red blood cells in a multi-layer suspended state;

[0044] Figure 5 is a schematic diagram of the steps of an AI training method for blood white blood cell recognition;

[0045] Figure 6 is a schematic diagram of the steps of an AI training method for blood white blood cell recognition;

[0046] Figure 7 is a schematic diagram of the steps of an AI training method for blood white blood cell recognition;

[0047] Figure 8 is a type of data file generated during the AI training process for blood white blood cell recognition;

[0048] Figure 9 It is a picture of suspended blood red blood cells

[0049] Figure 10 It is a partial picture of the suspended picture of blood red blood cells decomposed into small pictures

[0050] Figure 11 It is a scene of manually interpreting the identified white blood cells

[0051] Figure 12 It is an intermediate file output during the training process

[0052] Figure 13 It is a part of the intermediate file generated during the recognition training process

[0053] Figure 14 It is a small part of the identified white blood cells, and classification recognition is carried out on this basis Specific implementation manner

[0054] The following further details the content of the present application in conjunction with each attached drawing. It should be noted that the following is a description of the preferred embodiments of the present invention and does not constitute any limitation to the present invention. The description of the preferred embodiments of the present invention is only for the description of the general principles of the present invention. The numbers such as "first", "second" and "A", "B" involved in the present invention are only for the convenience of description and do not represent the order relationship in time or space. The combinations of letters and numbers "TA", "TB", "H" involved in the present invention are only for the convenience of description, and the specific meanings are determined by the specific words they represent.

[0055] Such as Figure 3 , the blood contains a large number of red blood cells and white blood cells. In the natural suspension state, waiting for the cells to sink to the bottom, the diameter of white blood cells is larger than that of red blood cells. At the same time, the diameter of white blood cells is larger than that of red blood cells, and there are relatively obvious characteristics.

[0056] Such as Figure 1 , a method for training an AI for blood white blood cell recognition. By preprocessing the blood sample, that is, diluting it with water to obtain a blood sample; the dilution reduces the concentration of red blood cells, and in the photographed picture, white blood cells can appear from the pile of red blood cells. Spread the blood sample flat to make the white blood cells suspended in the liquid and present a natural state; in order to obtain a picture by photographing, spreading flat can make the red blood cells and white blood cells arranged flat, especially making the white blood cells have the same height. For the obtained photo, the white blood cells have the same size characteristics. Photograph the blood sample to obtain a picture; use the obtained picture for AI training to obtain the white blood cell AI feature dataset A.

[0057] Such as Figure 2 、Such as Figure 3, the blood sample is laid flat so that the red blood cells are in a single-layer flat state. In the blood, after the red blood cells are laid flat, it will not affect the imaging of white blood cells.

[0058] The stacked height of the red blood cells when the blood sample is laid flat does not exceed the diameter of the white blood cells.

[0059] Such as Figure 4 , the concentration of red blood cells is relatively high. If the red blood cells are laid flat in a single layer, the white blood cells will be very sparse in the photo. In actual measurement, there will also be a situation of partial stacking of red blood cells. Therefore, in order to improve the recognition rate, it is necessary to train the stacked state of red blood cells. As long as the stacked height does not exceed the diameter of the white blood cells, the white blood cells can obtain better imaging.

[0060] Such as Figure 5 , the AI training includes image preprocessing and AI training processing. The image preprocessing outputs the labeled white blood cells, and the AI training processing outputs the white blood cell AI feature dataset according to the labeled white blood cells.

[0061] Directly trained, the accuracy rate of AI recognition can be achieved above 95%, which can meet the conventional applications in veterinary places. However, for human medical tests, the error needs to be controlled below 3%. Therefore, it is necessary to improve the accuracy rate through a large number of sample recognition trainings. For medical use, it is also necessary to truly conduct AI recognition verification. For the white blood cells or white blood cell classifications that are not accurately recognized, manual recheck and marking are carried out for the second AI training.

[0062] A staining agent is added to the blood to stain the white blood cells. In the quantitative detection of white blood cell classification, it is necessary to classify and identify different white blood cells according to the characteristics of the cell nucleus. After staining, the characteristics of the white blood cell nucleus can be highlighted, and the white blood cell nucleus also needs to be stained during the training process.

[0063] A liquid staining agent is added to the blood. The liquid staining agent dilutes the blood and stains the white blood cell nucleus. During the staining process, the dispersion of the liquid staining agent is good, which can stain the cells evenly. At the same time, it can dilute the blood, facilitate the laying flat of red blood cells, and reduce the interference of red blood cells on the AI recognition training of white blood cells.

[0064] Such as Figure 8 , using the white blood cell AI feature dataset A as the feature dataset, using the AI software to recognize one or more pictures input into the AI recognition of the picture, obtaining the marked pictures, the marked pictures include the labeled white blood cells, performing manual recheck on the marked pictures to form the manually rechecked marked pictures, and using the manually rechecked marked pictures for AI training to obtain the white blood cell AI feature dataset B.

[0065] For the AI feature dataset A trained at one time, the recognition rate is generally relatively low. Using the AI feature dataset A as features for AI recognition verification, that is, testing the accuracy of the feature dataset A.

[0066] The classification and recognition of white blood cells include any of the following white blood cells: the white blood cell is a neutrophil, and the white blood cell feature dataset A or B includes the features of stained neutrophils; the white blood cell is a lymphocyte, and the white blood cell feature dataset A or B includes the features of stained lymphocytes; the white blood cell is an eosinophil, and the white blood cell feature dataset A or B includes the features of stained eosinophils; the white blood cell is a basophil, and the white blood cell feature dataset A or B includes the features of stained basophils; the white blood cell is a neutrophilic band cell, and the white blood cell feature dataset A or B includes the features of stained neutrophilic band cells; the white blood cell is a neutrophilic segmented cell, and the white blood cell feature dataset A or B includes the features of stained neutrophilic segmented cells;

[0067] The white blood cell is a monocyte, and the white blood cell feature dataset A or B includes the features of stained monocytes.

[0068] For different white blood cells, the differential features are very small, and experienced medical technicians are required for annotation, which consumes a large amount of engineering time. For example, Figure 7 , it is possible to, on the basis of the first AI training, after identifying the white blood cells, invite professionals for the second annotation. That is, in the first AI training, the white blood cells are identified, and during the second AI training process, the specific types of white blood cells are trained, greatly reducing the workload of training.

[0069] The AI training method for blood white blood cell recognition can also be training without staining, which requires high requirements for the trainers.

[0070] The white blood cell is a neutrophil, and the white blood cell feature dataset A or B includes the features of unstained neutrophils; the white blood cell is a lymphocyte, and the white blood cell feature dataset A or B includes the features of unstained lymphocytes; the white blood cell is an eosinophil, and the white blood cell feature dataset A or B includes the features of unstained eosinophils; the white blood cell is a basophil, and the white blood cell feature dataset A or B includes the features of unstained basophils; the white blood cell is a neutrophilic band cell, and the white blood cell feature dataset A or B includes the features of unstained neutrophilic band cells; the white blood cell is a neutrophilic segmented cell, and the white blood cell feature dataset A or B includes the features of unstained neutrophilic segmented cells; the white blood cell is a monocyte, and the white blood cell feature dataset A or B includes the features of unstained monocytes.

[0071] Figure 8It is part of the data and file types generated by the AI ​​training process of blood leukocyte recognition. If the paddledetection model is used to train leukocyte recognition, a large number of three types of files will be generated, namely .pdmodel files, pdparams files and .pdopt files. The number of files is very large, and the data volume exceeds 100GB. pdmodel saves the training model, including the network and optimization iterations; pdparams saves the updated parameters in the network; pdopt saves all input and output variables. The above second AI training can be repeated. The more rounds of the second AI training, the larger the entire folder.

[0072] Figure 9 This is a picture of blood red blood cells suspended in suspension. The stained picture is colorful. It can be seen that the red blood cells are basically distributed in a single layer, and the number of white blood cells is relatively small. For example, in the sample used, basophils may be one ten-thousandth of the total number of white blood cells. Without the second AI training, it is almost impossible to manually obtain the interpretation of basophils. Therefore, the white blood cells are identified in the first round and obtained. Figure 14 After the atlas of white blood cells was created, it took a huge amount of manpower to select basophils from tens of thousands of photos. Of course, this turned the impossible into at least something that can be achieved.

[0073] Figure 10 It is a part of the picture of red blood cell suspension decomposed into small pictures; after the captured picture is segmented, it becomes a smaller picture, which is convenient for manual interpretation and annotation.

[0074] Figure 11 This is a scenario where the identified white blood cells are manually interpreted; by manually labeling the white blood cells, especially the types with small content, the identified white blood cells need to be preserved because the samples are scarce and not easy to find.

[0075] Figure 12 It is the intermediate file output by the training process; the various images of recognized white blood cell types should be classified and saved. Especially for those with low content, such as the content of one ten-thousandth, it requires a lot of manpower and cost to obtain them again.

[0076] Figure 13 It is part of the intermediate files generated during the recognition training process; Figure 14 It is a small portion of the identified white blood cells on the basis of which the classification is performed.

[0077] A computing and processing device, comprising the above-mentioned AI training method for blood leukocyte identification.

[0078] A computing and processing device, the computing and processing device includes the above-mentioned white blood cell AI feature dataset A or white blood cell AI feature dataset B.

[0079] A data storage device, the data storage includes the above-mentioned white blood cell AI feature dataset A or white blood cell AI feature dataset B.

[0080] Although the present invention is described and illustrated according to preferred embodiments and several alternative solutions, the invention is not limited by the specific descriptions in this specification. Other additional alternatives or equivalent components can also be used to practice the present invention.

Claims

1. A method for training AI for blood white blood cell recognition, characterized in that, it includes the following technical features: Dilute the blood to obtain a blood sample; Lay the blood sample flat so that white blood cells are suspended in the liquid, presenting a natural state; Take a picture of the blood sample to obtain a picture; Use the obtained picture for AI training to obtain the white blood cell AI feature dataset A.

2. The method for training AI for blood white blood cell recognition according to claim 1, characterized in that, it includes the following technical features: The blood sample is laid flat until the red blood cells present a single-layer flat state.

3. The method for training AI for blood white blood cell recognition according to claim 1 includes the following technical features: The stacked height of the red blood cells when the blood sample is laid flat does not exceed the diameter of the white blood cells.

4. The method for training AI for blood white blood cell recognition according to claim 1, characterized in that, The AI training includes image preprocessing and AI training processing. The image preprocessing outputs the labeled white blood cells, and the AI training processing outputs the white blood cell AI feature dataset according to the labeled white blood cells.

5. The method for training AI for blood white blood cell recognition according to claim 1, characterized in that, Add a staining agent to the blood to stain the white blood cells.

6. The method for training AI for blood white blood cell recognition according to claim 1, characterized in that, Add a liquid staining agent to the blood. The liquid staining agent dilutes the blood and stains the white blood cell nuclei.

7. The method for training AI for blood white blood cell recognition according to claim 1 or 5, characterized in that, Use the white blood cell AI feature dataset A as the feature dataset, use AI software to recognize one or more pictures input into the AI to recognize the said picture, obtain the marked pictures. The marked pictures include the labeled white blood cells. Perform manual review on the marked pictures to form the manually reviewed marked pictures, and use the manually reviewed marked pictures for AI training to obtain the white blood cell AI feature dataset B.

8. The method for training AI for blood white blood cell recognition according to claim 7 includes any one of the following technical features: TA10: The white blood cells are neutrophilic granulocytes, and the white blood cell feature dataset A or B includes the features of stained neutrophilic granulocytes; TA20: The white blood cells are lymphocytes, and the white blood cell feature dataset A or B includes the features of stained lymphocytes; TA30: The white blood cells are eosinophilic granulocytes, and the white blood cell feature dataset A or B includes the features of stained eosinophilic granulocytes; TA40: The white blood cells are basophilic granulocytes, and the white blood cell feature dataset A or B includes the features of stained basophilic granulocytes; TA50: The white blood cells are neutrophilic stab granulocytes, and the white blood cell feature dataset A or B includes the features of stained neutrophilic stab granulocytes; TA60: The white blood cells are neutrophilic segmented granulocytes, and the white blood cell feature dataset A or B includes the features of stained neutrophilic segmented granulocytes; TA70: The white blood cells are monocytes, and the white blood cell feature dataset A or B includes the features of stained monocytes.

9. The AI training method for blood white blood cell recognition according to claim 1 includes any one of the following technical features: TE10: The white blood cells are neutrophil white blood cells, and the white blood cell feature dataset A or B includes the features of unstained neutrophil white blood cells; TE20: The white blood cells are lymphocytes, and the white blood cell feature dataset A or B includes the features of unstained lymphocytes; TE30: The white blood cells are eosinophilic granulocytes, and the white blood cell feature dataset A or B includes the features of unstained eosinophilic granulocytes; TE40: The white blood cells are basophilic granulocytes, and the white blood cell feature dataset A or B includes the features of unstained basophilic granulocytes; TE50: The white blood cells are neutrophilic stab granulocytes, and the white blood cell feature dataset A or B includes the features of unstained neutrophilic stab granulocytes; TE60: The white blood cells are neutrophilic segmented granulocytes, and the white blood cell feature dataset A or B includes the features of unstained neutrophilic segmented granulocytes; TE70: The white blood cells are monocytes, and the white blood cell feature dataset A or B includes the features of unstained monocytes.

10. The AI training method for blood white blood cell recognition according to any one of claims 1 to 9, characterized in that the white blood cell feature dataset A is a white blood cell feature dataset, and the white blood cell feature dataset B is a sub-item white blood cell feature dataset.

11. A computing and processing device, characterized in that the computing and processing device includes the AI training method for blood white blood cell recognition described in any one of claims 1 to 10.

12. A computing and processing device, characterized in that the memory of the computing and processing device includes the white blood cell AI feature dataset A or the white blood cell AI feature dataset B described in any one of claims 1 to 10.

13. A data storage device, characterized in that the data storage includes the white blood cell AI feature dataset A or the white blood cell AI feature dataset B described in any one of claims 1 to 10.

Images