Automatic identification method for serum protein electrophoresis chart

Abstract

The serum protein electrophoresis image automatic recognition method relates to the technical field of electrophoresis image analysis, the serum protein electrophoresis image is divided into multiple electrophoresis sub-bands with the same horizontal width by horizontal equal-width segmentation, and the electrophoresis sub-bands are taken as the first type of image; a multi-view image is generated based on a self-attention mechanism and taken as the second type of image; a machine learning model is established, features of the image are extracted, the features are taken as input, annotation results of medical experts are taken as supervision information, the machine learning model is trained, weights of different features are configured, secondary judgment results of the medical experts are taken as supervision information, and the machine learning model is trained again to obtain an accurate model for recognizing monoclonal immunoglobulin; and the serum protein electrophoresis image is recognized by using the accurate model for recognizing monoclonal immunoglobulin, thereby solving the problem that the prior art is time-consuming and laborious in analyzing the serum protein electrophoresis image, and the serum protein electrophoresis image is suitable for recognizing monoclonal immunoglobulin.

Description

Technical Field

[0001] This invention relates to the field of electrophoresis analysis technology, and in particular to an automatic identification method for serum protein electrophoresis patterns. Background Technology

[0002] Multiple myeloma is a relatively common type of malignant proliferation of bone marrow plasma cells. It is usually accompanied by the production of large amounts of monoclonal immunoglobulin in the patient's body. Monoclonal immunoglobulin is a single type of immunoglobulin produced by the proliferation of a single plasma cell clone, abbreviated as M protein.

[0003] Serum protein electrophoresis and immunofixation electrophoresis are two chemical analysis techniques based on electrophoresis and precipitation reactions. Both methods are widely used clinically. Serum protein electrophoresis is less expensive and has a wider range of applications than immunofixation electrophoresis, and is commonly used for disease screening, diagnosis, and efficacy assessment. Serum protein electrophoresis (SPE) is widely used clinically to detect the presence of M protein.

[0004] However, in the existing technology, the analysis of serum protein electrophoresis images can only be determined by relevant experts to determine whether the serum protein electrophoresis image contains M protein. This is time-consuming and labor-intensive. Due to limited manpower, it is not possible to effectively and quickly identify electrophoresis images. Summary of the Invention

[0005] The technical problem solved by this invention is to provide an automatic identification method for serum protein electrophoresis images, thereby solving the problem of time-consuming and laborious analysis of serum protein electrophoresis images in the prior art.

[0006] The technical solution adopted by this invention to solve the above-mentioned technical problems: The automatic identification method for serum protein electrophoresis images includes the following steps:

[0007] S01. Obtain a certain number of serum protein electrophoresis images;

[0008] S02. Divide the serum protein electrophoresis image into equal-width segments to obtain multiple electrophoretic sub-bands with equal horizontal distances. Use the electrophoretic sub-bands as the first type of image.

[0009] S03. Generate multi-view images based on the electrophoretic sub-bands using a self-attention mechanism, as the second type of images;

[0010] S04. Establish a machine learning model, extract features from the first type of image and the second type of image, use the features as input, and use the annotation results of medical experts as supervision information, wherein the annotation results are whether monoclonal immunoglobulins are present, train the machine learning model, and obtain an initial model capable of recognizing monoclonal immunoglobulins.

[0011] S05. Compare the recognition results of the initial model capable of recognizing monoclonal immunoglobulins with the annotation results of medical experts, identify serum protein electrophoresis images with inconsistent results, and then have medical experts make a second judgment, adjusting the weight of each feature so that the output result of the machine learning model is consistent with the second judgment result, thereby obtaining an accurate model for recognizing monoclonal immunoglobulins.

[0012] S06. The serum protein electrophoresis pattern is identified using the precise model for identifying monoclonal immunoglobulins.

[0013] Furthermore, S01 also includes removing abnormal serum protein electrophoresis images from the certain number of serum protein electrophoresis images. The abnormal serum protein electrophoresis images include serum protein electrophoresis images with non-standard detection results and serum protein electrophoresis images detected under abnormal conditions of detection equipment.

[0014] Furthermore, the machine learning model is the XGBoost model.

[0015] Furthermore, the generation of multi-view images based on the self-attention mechanism includes the following steps:

[0016] S301. Let the electrophoretic sub-band be i, and the peak value of the i-th electrophoretic sub-band be taken as the first type value, denoted as v. i

[0017] S302. Find the difference between the i-th electrophoretic sub-band and its n-th and n-th sub-bands respectively, where n ranges from 1 to max[i].

[0018] S303. The sum of the absolute values ​​of the differences is taken as the binary value of the sub-band, i.e. Where i is the number of the electrophoretic sub-band, h i For the second type value of the i-th sub-strip, v i Let v be the first type value of the i-th sub-strip. i+j For the first class value of the (i+j)th sub-strip, v i-j The first-class value of the ij-th sub-strip;

[0019] S304. A multi-view image is formed by using the number of the electrophoretic sub-bands as the horizontal axis and the second-class value as the vertical axis.

[0020] Furthermore, the features of the first type of image are the number i of the electrophoretic sub-band and the first type value, while the features of the second type of image are the number i of the electrophoretic sub-band and the second type value under different n.

[0021] Furthermore, the automatic recognition method for serum protein electrophoresis images also includes: S07, when the machine learning model recognizes the serum protein electrophoresis image, it extracts important features and displays the electrophoretic sub-bands corresponding to the important features to the user.

[0022] Furthermore, methods for extracting important features include: the machine learning model sorts the extracted features in descending order of their importance based on their role, with the top K features being the most important features, where K is a positive integer.

[0023] The beneficial effects of this invention are as follows: The automatic recognition method for serum protein electrophoresis images of this invention involves dividing a certain number of serum protein electrophoresis images horizontally into equal-width segments to obtain many electrophoretic sub-bands with equal horizontal distances. These sub-bands are used as the first type of image. Multi-view images are generated based on a self-attention mechanism and used as the second type of image. A machine learning model is established to extract features from both the first and second type of images. Using these features as input and the annotation results from medical experts as supervisory information, the machine learning model is trained to obtain an initial model capable of recognizing monoclonal immunoglobulins. The method then identifies discrepancies between the recognition results obtained using the initial model and the annotation results from medical experts. Consistent images are evaluated by medical experts, who then make secondary judgments on inconsistent images. Using first-class and second-class image features as overall input and assigning different weights, the model is trained with the expert's secondary judgment as supervisory information. This yields a model that accurately identifies monoclonal immunoglobulins. The weights of each feature are adjusted to ensure the machine learning model's output aligns with the secondary judgment results, resulting in an accurate model for monoclonal immunoglobulin identification. This accurate model is then used to identify serum protein electrophoresis images, solving the problem of time-consuming and labor-intensive analysis of serum protein electrophoresis images in existing technologies. Attached Figure Description

[0024] Appendix Figure 1 This is a schematic flowchart of the automatic recognition method for serum protein electrophoresis patterns of the present invention. Detailed Implementation

[0025] The present invention provides an automatic identification method for serum protein electrophoresis patterns, as shown in the appendix. Figure 1 As shown, it includes the following steps:

[0026] S01. Obtain a certain number of serum protein electrophoresis images;

[0027] Specifically, S01 also includes removing abnormal serum protein electrophoresis images from the certain number of serum protein electrophoresis images. The abnormal serum protein electrophoresis images include serum protein electrophoresis images with non-standard detection results and serum protein electrophoresis images detected under abnormal conditions of detection equipment, so as to ensure the accuracy of the initial data.

[0028] S02. Divide the serum protein electrophoresis image into equal-width segments to obtain multiple electrophoretic sub-bands with equal horizontal distances. Use the electrophoretic sub-bands as the first type of image.

[0029] Specifically, the width can be set according to the actual situation. The serum protein electrophoresis image is horizontally divided to obtain multiple electrophoretic sub-bands with the same horizontal distance. These electrophoretic sub-bands are used as the first type of image.

[0030] S03. Generate multi-view images based on the electrophoretic sub-bands using a self-attention mechanism, as the second type of images;

[0031] Specifically, the electrophoretic subbands are numbered sequentially from front to back, denoted by i, where i is a positive integer. The peak value of each subband is taken as the first-class value of the subband, denoted as v. i ; Calculate the differences between the i-th electrophoretic sub-band and its preceding n sub-bands and its following n sub-bands, where n ranges from 1 to max[i]; Sum the absolute values ​​of the calculated differences as the binary value of that sub-band, i.e. Where i is the number of the electrophoretic sub-band, h i For the second type value of the i-th sub-strip, v i Let v be the first type value of the i-th sub-strip. i+j For the first class value of the (i+j)th sub-strip, v i-j Let n be the first-class value of the ij-th sub-band. The graph composed of the sub-band number as the x-axis and the second-class value as the y-axis is a multi-view image. Since the n values ​​are different, there are n second-class values ​​for the same sub-band. Each n corresponds to a view image. Therefore, the second-class multi-view image focuses more on the changes in waveforms between regions of the serum protein electrophoresis map and the changes in waveforms near the sub-bands.

[0032] S04. Establish a machine learning model, extract features from the first type of image and the second type of image, use the features as input, and use the annotation results of medical experts as supervision information, wherein the annotation results are whether monoclonal immunoglobulins are present, train the machine learning model, and obtain an initial model capable of recognizing monoclonal immunoglobulins.

[0033] Specifically, an XGBoost model is built to extract features from the first and second class images. The features of the first class images are the electrophoretic sub-band number i and the first class value, such as iv. i The second type of image is characterized by the numbering of the electrophoretic sub-bands under different n values ​​and the second type of value, such as nih. i Each feature is used as input data, and the annotation results of medical experts are used as supervision information. The annotation results indicate whether monoclonal immunoglobulins are present. The XGBoost model is trained to obtain an initial model capable of recognizing monoclonal immunoglobulins.

[0034] S05. Compare the recognition results of the initial model capable of recognizing monoclonal immunoglobulins with the annotation results of medical experts, identify serum protein electrophoresis images with inconsistent results, and then have medical experts make a second judgment, adjusting the weight of each feature so that the output result of the machine learning model is consistent with the second judgment result, thereby obtaining an accurate model for recognizing monoclonal immunoglobulins.

[0035] Specifically, the model is trained by using the first type of image features and the second type of image features as the overall input, configuring different weights, and using the secondary judgment results of medical experts as supervision information to obtain an accurate model for recognizing monoclonal immunoglobulins.

[0036] S06. The serum protein electrophoresis pattern is identified using the precise model for identifying monoclonal immunoglobulins.

[0037] The automatic recognition method for serum protein electrophoresis images of the present invention may further include step S07: when the machine learning model recognizes the serum protein electrophoresis image, it extracts important features and displays the electrophoretic sub-bands corresponding to the important features to the user.

[0038] Specifically, when the XGBoost model makes predictions, the features are sorted in descending order of their importance. The top K features are considered important features, where K is a positive integer. The corresponding serum protein electrophoresis images of the K features are then displayed to the user.

Claims

1. An automatic identification method for serum protein electrophoresis patterns, characterized in that, Includes the following steps: S01. Obtain a certain number of serum protein electrophoresis images; S02. Divide the serum protein electrophoresis image into equal-width segments to obtain multiple electrophoretic sub-bands with equal horizontal distances. Use the electrophoretic sub-bands as the first type of image. S03. Generate multi-view images based on the electrophoretic sub-bands using a self-attention mechanism, as the second type of images, including the following steps: S301, denot the electrophoretic sub-bands as... , No. The peak value of each electrophoretic subband is taken as the first-class value, denoted as . ; S302. Find the difference between the i-th electrophoretic sub-band and its n-th and n-th sub-bands respectively, where n ranges from 1 to max[i]. S303. The sum of the absolute values ​​of the differences is taken as the binary value of the sub-band, i.e. ,in, For the numbering of electrophoretic sub-bands, For the first The second type value of each sub-band For the first The first type value of each sub-band For the first The first type value of each sub-band For the first The first type value of each sub-band; S304. A graph composed of the electrophoretic sub-band number as the abscissa and the second type value as the ordinate is a multi-view image. S04. Establish a machine learning model, extract features from the first type of image and the second type of image, use the features as input, and use the annotation results of medical experts as supervision information, wherein the annotation results are whether monoclonal immunoglobulins are present, train the machine learning model, and obtain an initial model capable of recognizing monoclonal immunoglobulins. S05. Compare the recognition results of the initial model capable of recognizing monoclonal immunoglobulins with the annotation results of medical experts, identify serum protein electrophoresis images with inconsistent results, and then have medical experts make a second judgment, adjusting the weight of each feature so that the output result of the machine learning model is consistent with the second judgment result, thereby obtaining an accurate model for recognizing monoclonal immunoglobulins. S06. The serum protein electrophoresis pattern is identified using the precise model for identifying monoclonal immunoglobulins.

2. The automatic recognition method for serum protein electrophoresis patterns according to claim 1, characterized in that, S01 further includes removing abnormal serum protein electrophoresis images from the certain number of serum protein electrophoresis images. The abnormal serum protein electrophoresis images include serum protein electrophoresis images with non-standard detection results and serum protein electrophoresis images detected under abnormal conditions of detection equipment.

3. The automatic recognition method for serum protein electrophoresis patterns according to claim 1, characterized in that, The machine learning model is the XGBoost model.

4. The automatic recognition method for serum protein electrophoresis patterns according to claim 1, characterized in that, The features of the first type of image are the number i of the electrophoretic sub-band and the first type value, while the features of the second type of image are the number i of the electrophoretic sub-band and the second type value under different n.

5. The method for automatic recognition of serum protein electrophoresis patterns according to any one of claims 1-3, characterized in that, The automatic recognition method for serum protein electrophoresis images further includes: S07, when the machine learning model recognizes the serum protein electrophoresis image, it extracts important features and displays the electrophoretic sub-bands corresponding to the important features to the user.

6. The automatic recognition method for serum protein electrophoresis patterns according to claim 5, characterized in that, Methods for extracting important features include: the machine learning model sorts the extracted features in descending order of their importance based on their role, and the top K features are considered important features, where K is a positive integer.

Images