Identifying and analyzing glue maps using artificial intelligence

Abstract

The invention relates to identification and analysis of gel maps using artificial intelligence. According to the scheme, on the basis of the artificial intelligence visual processing technology and the image processing technology, plasmids, products and scale lanes are recognized, the specific positions of the corresponding lanes and strips in the image are determined, tiny gaps between similar glue maps can be perceived, and analysis of the sizes of the product strips and genome strips in the glue maps is carried out. And the judgment on the analysis result is not influenced by personal subjective factors, so that the accuracy and the stability of the analysis result are improved.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims priority to Chinese patent application No. 202311300510.6 filed on October 9, 2023, the entire contents of which are incorporated by reference. Technical Field

[0003] The present disclosure relates to electrophoresis band technology in the biological field, and more specifically to using artificial intelligence tools to identify and analyze electrophoresis band imaging gel images. Background Art

[0004] Electrophoresis technology, or electrophoretic band technology, has now been widely used in fields such as biology and medicine.

[0005] Many molecules, such as amino acids, peptides, and nucleotides, have ionizable groups. Under the action of an electric field, these charged molecules will move toward the electrode with the opposite polarity of their charge, which is the phenomenon of electrophoresis. Due to the different movement rates of various ions, various substances can be divided into several rows, which appear as strips of light under optical instruments. These are electrophoresis strips. People can use existing data or software to analyze the composition and content of various substances based on the width and position of the strips.

[0006] Electrophoresis bands can be displayed in the form of gel images through optical instruments. Therefore, researchers can identify and analyze gel images, specifically the lanes and bands on the gel images. Both gel image identification and gel image analysis currently involve a large amount of manual operations.

[0007] Manual analysis of gel images first requires manual confirmation of sample information, judging the expected results of each sample on the gel image, and comparing and judging whether the information content matches. Then, the image is cut and saved based on the judgment results, and whether there are any problems is marked. This manual operation process may have inconsistent judgment standards for the same sample, and human factors have a relatively large interference. The operation process not only takes a lot of time, energy and cost, but also has a high probability of error, and there are errors in the judgments of different people.

[0008] There are studies on automatic recognition of glue images in the prior art, but the current research requires manual active adjustment of parameters and has no automatic analysis function, or the recognition and analysis functions are far from meeting actual needs.

[0009] The PCT international application with publication number WO 2022 / 188696 A1 proposes a method for automatically identifying and analyzing gel images. The method introduces image processing technology, and under the guidance of automatic recognition and analysis of computer software, automatically identifies and analyzes band information with high accuracy. At the same time, the method can automatically determine whether the gel is running correctly, and for incorrect gel running, it can also specifically analyze the type of problem. By reference, the entire content of this document is incorporated into the present disclosure and becomes part of the present disclosure. Summary of the invention

[0010] The present invention is based on artificial intelligence (AI) visual processing technology and image processing technology, and is used to complete the automatic cutting and quality control (QC) result judgment of the original glue image taken by the glue shooting instrument, and automatically synchronize the production execution system (manufacturing execution system, MES), and establish a complete set of QC inspection standards.

[0011] According to a first aspect of the present disclosure, a method for identifying and analyzing gel images using artificial intelligence is provided. The method may include: identifying lanes and bands on the lanes in the original gel image of the electrophoresis image using an artificial intelligence recognition network model; and analyzing the identified lanes and bands on the lanes.

[0012] Preferably, the method according to the first aspect of the present disclosure may further include: in the case where the gel image includes more than one electrophoresis sample, determining the plasmid, product and scale lanes of each electrophoresis sample based on the lanes and bands identified by using the artificial intelligence recognition network model, and cutting the gel image based on each electrophoresis sample.

[0013] In the method according to the first aspect of the present disclosure, the artificial intelligence recognition network model can be constructed by the following steps: obtaining a training set, the training set including multiple training samples, each of the training samples including an original picture of the electrophoresis gel image and an annotation of the picture, wherein the annotation of the picture may include annotations of lanes in the picture and bands on the lanes; based on the original picture of the electrophoresis gel image and the annotation of the picture, training the initial artificial intelligence recognition network model to obtain an artificial intelligence recognition network model for identifying lanes and bands on the lanes in the original picture of the electrophoresis gel image.

[0014] Preferably, the marking of the picture may further include marking of the glue holes in the picture.

[0015] Preferably, the initial artificial intelligence recognition network model can be a convolutional neural network (CNN) model.

[0016] Preferably, the initial artificial intelligence recognition network model can be a Yolo network model or an SSD network model.

[0017] In the method according to the first aspect of the present disclosure, the use of an artificial intelligence recognition network model to identify lanes and bands on the lanes in the original gel image of the electrophoresis image may include: using an artificial intelligence recognition network model to identify gel holes in the original gel image; performing horizontal deflection correction on the original gel image based on the deflection angle of the horizontal position of the identified gel holes; and using the artificial intelligence recognition network model to identify lanes and bands on the lanes in the corrected gel image.

[0018] In the method according to the first aspect of the present disclosure, the analysis of the identified lanes and bands on the lanes includes performing at least one of the following operations: genome detection; cleavage or degradation detection; product band size calculation; and impurity detection.

[0019] Preferably, the genome detection may include: using an artificial intelligence prediction network model to preliminarily predict whether a genome exists in the gel image; when the artificial intelligence prediction network model preliminarily predicts the presence of a genome in the gel image, determining the content of the genome band according to the positional characteristics of the genome band to be determined in the gel image and a quantitative analysis method. Preferably, the positional characteristics of the genome may be the position of the band that exceeds the top edge of the ruler on the product lane. Specifically, based on the position of the genome in the image, it is compared with the ruler to check whether it meets the positional characteristics of the genome to determine whether it is a false alarm.

[0020] Preferably, the blurring or degradation detection may include: using an artificial intelligence prediction network model to preliminarily predict whether there is a blurring or degradation area in the glue map; when the artificial intelligence prediction network model predicts that there is a blurring or degradation area in the glue map, further judging the degree of blurring or degradation based on the upper and lower positions of the preliminarily predicted blurring or degradation area and the percentage of the brightness of the blurring or degradation area to the strip brightness of the identification area.

[0021] Preferably, the product band size calculation includes: determining the scale type based on the bands in the scale lane identified by using the artificial intelligence recognition network model; determining the size of the bands in the product lane based on the relative position relationship between the bands in the product lane identified by using the artificial intelligence recognition network model and the bands in the scale lane.

[0022] Preferably, the determining of the ruler type may further include: determining the ruler type according to the number of bands in the ruler lane and the distance between the last two bands.

[0023] Preferably, after determining the type of ruler and before determining the size of the band in the product lane, the position of the band can be precisely adjusted.

[0024] Preferably, the step of determining the size of the bands in the product lane according to the relative positional relationship between the bands in the product lane and the bands in the ruler lane identified by using the artificial intelligence recognition network model further includes: determining the size of each band in the ruler lane according to the determined ruler type and with reference to the band sizes corresponding to each band in the standard ruler diagram of the corresponding type; and determining the size of the bands in the product lane according to the relative positional relationship between the bands in the product lane and the bands in the ruler lane.

[0025] Preferably, determining the size of the band in the product lane according to the relative positional relationship between the band in the product lane and the band in the ruler lane may further include: judging whether the band in the product lane is flush with a band in the ruler lane; when the band in the product lane is flush with a band in the ruler lane, the size of the band in the product lane is the size of the band in the ruler lane that is flush with it.

[0026] Preferably, determining the size of the band in the product lane based on the relative positional relationship between the band in the product lane and the band in the ruler lane may further include: determining the size of the band in the product lane by calculating the relative positional relationship between the upper edge, lower edge and center line of the band in the product lane and the band in the ruler lane.

[0027] Preferably, the method of determining the size of the band in the product lane by calculating the relative positional relationship between the upper edge, lower edge and center line of the band in the product lane and the band in the ruler lane can further include: determining the middle value of the size of the band in the product lane by calculating the relative positional relationship between the upper edge, lower edge of the band in the product lane and the band in the ruler lane; determining the centerline value of the size of the band in the product lane by calculating the relative positional relationship between the centerline of the band in the product lane and the band in the ruler lane; comparing the middle value with the centerline value, and taking the one closest to the theoretical value of the band in the product lane as the size of the band in the product lane.

[0028] In addition, when calculating the size of the stripe, adjustments may be made based on the perception error of the human eye.

[0029] Preferably, the impurity detection includes: using an artificial intelligence prediction network model to preliminarily predict whether there are impurities in the gel image; when the artificial intelligence prediction network model preliminarily predicts that there are impurities in the gel image, judging whether it affects the determination of the product band based on the preliminarily predicted impurity position, size and brightness.

[0030] Preferably, the analysis of the identified lanes and bands on the lanes may further include performing at least one of the following operations: ruler integrity detection; product enzyme incomplete cleavage detection; plasmid band detection; and mixed band or dropped band detection.

[0031] In the case of genome testing, the artificial intelligence prediction network model can be constructed through the following steps: obtaining a training set, the training set including multiple training samples, each of the training samples including an original picture of the electrophoresis gel image and an annotation of the picture, wherein the annotation of the picture includes an annotation of the genome and its location; based on the original picture of the electrophoresis gel image and the annotation of the picture, the initial artificial intelligence prediction network model is trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether a genome exists in the product lane.

[0032] In the case of blurring or degradation detection, the artificial intelligence prediction network model is constructed through the following steps: obtaining a training set, the training set includes multiple training samples, each of the training samples includes an original picture of the electrophoresis gel map and an annotation of the picture, wherein the annotation of the picture includes annotation of the blurring or degradation area and its location; based on the original picture of the electrophoresis gel map and the annotation of the picture, the initial artificial intelligence prediction network model is trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether there is a blurring or degradation area in the gel map.

[0033] In the case of impurity detection, the artificial intelligence prediction network model is constructed through the following steps: obtaining a training set, the training set includes multiple training samples, each of the training samples includes an original picture of the electrophoresis gel map and an annotation of the picture, wherein the annotation of the picture includes an annotation of the impurities and their locations; based on the original picture of the electrophoresis gel map and the annotation of the picture, the initial artificial intelligence prediction network model is trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether there are impurities in the gel map.

[0034] Preferably, the initial artificial intelligence prediction network model can be a convolutional neural network (CNN) model.

[0035] Preferably, the initial artificial intelligence prediction network model can be a Yolo network model or an SSD network model.

[0036] According to a second aspect of the present disclosure, a method for constructing an artificial intelligence recognition network model for identifying gel images is provided. The method may include: obtaining a training set, the training set including a plurality of training samples, each of the training samples including an original image of an electrophoresis gel image and annotations of the image, wherein the annotations of the image include annotations of lanes in the image and bands on the lanes; based on the original image of the electrophoresis gel image and the annotations of the image, an initial artificial intelligence recognition network model is trained to obtain an artificial intelligence recognition network model for identifying lanes and bands on the lanes in the original image of the electrophoresis gel image.

[0037] Preferably, the marking of the picture may further include marking of the glue holes in the picture.

[0038] Preferably, the initial artificial intelligence recognition network model can be a convolutional neural network (CNN) model.

[0039] Preferably, the initial artificial intelligence recognition network model can be a Yolo network model or an SSD network model.

[0040] According to a third aspect of the present disclosure, there is provided an artificial intelligence recognition network model for identifying glue images obtained by the method described in the second aspect of the present disclosure.

[0041] According to a fourth aspect of the present disclosure, a method for constructing an artificial intelligence prediction network model for analyzing gel images is provided. The method may include: obtaining a training set, the training set including a plurality of training samples, each of the training samples including an original image of an electrophoresis gel image and an annotation of the image, wherein the annotation of the image includes annotations of items with gel image quality problems and their locations; based on the original image of the electrophoresis gel image and the annotation of the image, an initial artificial intelligence prediction network model is trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether there are quality problems in the gel image.

[0042] Preferably, the annotation of the image includes annotation of the genome and its location. Preferably, based on the original image of the electrophoresis gel and the annotation of the image, the initial artificial intelligence prediction network model is trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether the genome exists in the product lane.

[0043] Preferably, the annotation of the image includes annotation of the cut or degraded area and its location. Preferably, based on the original image of the electrophoresis gel image and the annotation of the image, the initial artificial intelligence prediction network model is trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether there is a cut or degraded area in the gel image.

[0044] Preferably, the annotation of the image includes annotation of impurities and their locations. Preferably, based on the original image of the electrophoresis gel image and the annotation of the image, the initial artificial intelligence prediction network model is trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether there are impurities in the gel image.

[0045] Preferably, the initial artificial intelligence prediction network model can be a convolutional neural network (CNN) model.

[0046] Preferably, the initial artificial intelligence prediction network model can be a Yolo network model or an SSD network model.

[0047] According to a fifth aspect of the present disclosure, there is provided an artificial intelligence prediction network model for analyzing glue images obtained by the method described in the fourth aspect of the present disclosure.

[0048] According to a sixth aspect of the present disclosure, a system for identifying and analyzing gel images using artificial intelligence is provided. The system may include: an acquisition module for acquiring an original gel image of an electrophoresis image; a processing module for identifying and analyzing the gel image using artificial intelligence, the processing module further including the following submodules: an identification submodule for identifying lanes and bands on the lanes in the original gel image of the electrophoresis image using an artificial intelligence identification network model; and an analysis submodule for analyzing the lanes and bands on the lanes identified by the identification submodule.

[0049] According to the sixth aspect of the present disclosure, the system for identifying and analyzing gel images using artificial intelligence can be implemented by a computer. Preferably, the following operations can be implemented by executing a computer program: obtaining an original image of the gel image of the electrophoresis image; identifying and analyzing the gel image using artificial intelligence, and the operation further includes the following sub-operations: identifying the lanes and bands on the lanes in the original image of the gel image of the electrophoresis image using an artificial intelligence recognition network model; and analyzing the identified lanes and bands on the lanes.

[0050] According to a seventh aspect of the present disclosure, a non-transitory computer-readable storage medium is provided for storing a computer program. The computer program includes instructions. When the instructions are executed by a processor of an electronic device, the electronic device implements the method for identifying and analyzing glue images using artificial intelligence according to the first aspect of the present disclosure, or the method for constructing an artificial intelligence recognition network model for identifying glue images according to the second aspect of the present disclosure, or the method for constructing an artificial intelligence prediction network model for analyzing glue images according to the fourth aspect of the present disclosure.

[0051] According to an eighth aspect of the present disclosure, a computer system is provided. The computer system includes: a processor, a memory, and a computer program. The computer program is stored in the memory and is configured to be executed by the processor. The computer program includes instructions for implementing the method for identifying and analyzing glue images using artificial intelligence according to the first aspect of the present disclosure, or the method for constructing an artificial intelligence recognition network model for identifying glue images according to the second aspect of the present disclosure, or the method for constructing an artificial intelligence prediction network model for analyzing glue images according to the fourth aspect of the present disclosure.

[0052] According to the contents of the present disclosure, when performing glue image cutting, artificial intelligence visual processing technology is used to identify and classify the target, and the position of the target object in the image is obtained, and then algorithm processing is performed to achieve the ideal effect.

[0053] The solution proposed in this disclosure mainly solves the following two problems:

[0054] a) When a large number of samples are processed in the same gel map, manual cutting of the gel map requires manual verification using image processing software such as Photoshop, and then correction and preservation of the cut images. The whole process is time-consuming and laborious, and large-volume orders and manual processing are prone to problems such as order stringing;

[0055] b) When manually reviewing QC results, due to the influence of human subjective factors, there is no unified QC testing standard. Under large-volume orders, it is easy to cause eye fatigue and fail to detect subtle differences in similar glue images.

[0056] Specifically, in the actual production process, there are some business scenarios that cannot be effectively handled, resulting in errors, such as the inability to detect weak bands, unseparated bands during gel running, and the inability to identify impurities of different shapes. However, it is easy to use AI vision to handle such problems.

[0057] The present disclosure is based on AI machine vision processing technology and traditional image processing technology to identify plasmids, products, and scale lanes, and determine the specific positions of corresponding lanes and bands in the image. It can perceive the subtle differences between similar gel images, and the judgment of quality control results is not affected by personal subjective factors. All results are judged according to standards. Although the technology disclosed in the present disclosure takes enzyme-cut gel images as an example and is applied to automatic cutting of gel images and automatic judgment of QC results, it can also be applied to other application scenarios that can perform image recognition and processing, such as SDS-PAGE protein gel images, PCR bacterial detection gel images, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The present disclosure will be more fully understood from the following detailed description taken in conjunction with the accompanying drawings, in which like elements are numbered in a similar manner, and in which:

[0059] Figure 1 The present invention is a flowchart of a method for identifying and analyzing glue images using artificial intelligence according to an embodiment of the present invention.

[0060] Figure 2 It is a standard scale reference diagram.

[0061] Figure 3A It is a flowchart of a method for constructing an artificial intelligence recognition network model used in the present disclosure for recognizing glue images.

[0062] Figure 3B It is a flowchart of the method for constructing the artificial intelligence prediction network model used in the present disclosure for analyzing glue images.

[0063] Figure 4 It is a schematic diagram for marking and classifying glue drawings in the annotation software.

[0064] Figure 5 is a schematic diagram of the original QC gel map.

[0065] Figure 6 It is the AI ​​model recognition result of the original QC glue image.

[0066] Figure 7 The process of scale calibration is shown.

[0067] Figure 8 It is a schematic diagram for saving the glue image after cutting.

[0068] Fig. 9 This is an example of a QC testing process.

[0069] Fig.10 Used to illustrate the division of lanes in a gel sample.

[0070] Fig.11 is an example of a genomic feature.

[0071] Fig.12 This is a schematic diagram of annotating a genome within the annotation software.

[0072] Fig.13 Schematic diagram of the genome prediction process.

[0073] Fig.14 The genomic regions and corresponding peak profiles are illustrated.

[0074] Fig.15 The original gel image, band identification results, genome identification results and product lane band content calculation are shown.

[0075] Fig.16 It is a schematic diagram for detecting cutting and degradation.

[0076] Fig.17 It is a schematic diagram for marking and classifying the cut and degradation in the marking software.

[0077] Fig.18 The predicted degradation process is shown.

[0078] Fig.19 The prediction process of cutting paste is illustrated.

[0079] Fig. 20 It is a schematic diagram of labeling and classifying bands and lanes in the labeling software.

[0080] Fig.21 It shows the result after the AI ​​model identifies the bands in the lane and performs algorithm correction.

[0081] Fig. 22 The sharp corners of the strip are shown.

[0082] Fig.23 It is a schematic diagram of product band size calculation.

[0083] Fig.24 This is a schematic diagram of labeling and classifying impurities in the labeling software.

[0084] Fig.25 The impurity prediction process is illustrated.

[0085] Fig.26 is a schematic diagram of the standard strip.

[0086] Fig. 27 It is a schematic block diagram of a system for identifying and analyzing glue images using artificial intelligence according to the present disclosure. DETAILED DESCRIPTION

[0087] Unless otherwise defined, technical and scientific terms used in the present disclosure have the same meaning as commonly understood by one of ordinary skill in the art to which the present disclosure belongs.

[0088] The technical scheme of the present disclosure is further described in detail below by way of examples and in conjunction with the accompanying drawings. Unless otherwise specified, the methods and materials of the embodiments described below are conventional products that can be purchased from the market. It will be understood by those skilled in the art that the present disclosure belongs to that the methods and materials described below are merely exemplary and should not be construed as limiting the scope of the present disclosure.

[0089] a) Application scope of computer vision and its artificial intelligence (AI) framework

[0090] Computer vision refers to the process of processing and analyzing images or videos using artificial intelligence algorithms, including functions such as object recognition, object classification, object tracking, and optical character recognition. Computer vision applications run on a type of artificial neural network algorithm, using convolutional layers to filter out useful information from the input. The convolution operation requires the combined use of input data (feature maps) and convolution kernels (filters) to generate the transformed feature maps. Convolutional layer filters can be modified according to learning parameters to extract the most useful information for a specific task, which enables it to provide more accurate image analysis. This common artificial neural network mainly includes R-CNN, Fast R-CNN, Faster R-CNN, Mask R-CNN, Yolo, SSD, etc.

[0091] R-CNN (Region with CNN features): It is mainly divided into two stages: candidate region extraction and candidate region classification. First, the input is superpixel merged through selective search to generate basic sub-regions. Then small sub-regions are continuously merged into large regions, and areas where objects may exist are found from them. This process is called candidate region extraction (Region Proposal). After extracting the candidate region containing the target, it needs to be classified to determine which category the target belongs to. Classification can be performed through algorithms such as SVM (Support Vector Machine) or CNN (Convolutional Neural Network). Please refer to https: / / arxiv.org / abs / 1311.2524. The entire content of this document is incorporated into the present disclosure by reference here, making it a part of the present disclosure.

[0092] Fast R-CNN: Improved from R-CNN, it is mainly to overcome the shortcomings of R-CNN, improve training speed and recognition accuracy, and combine the two losses after extracting the features of the ROI (Region of Interest) for unified training. Please refer to https: / / arxiv.org / abs / 1504.08083, the entire contents of which are incorporated into this disclosure by reference, making it a part of this disclosure.

[0093] Faster R-CNN: Improved from R-CNN, it makes full use of the feature maps generated by R-CNN to perform region recommendation, thereby achieving a higher training speed. Please refer to https: / / arxiv.org / abs / 1506.01497, the entire contents of which are incorporated into this disclosure by reference, making it a part of this disclosure.

[0094] Mask R-CNN: Using the recognition effect of Faster R-CNN on objects, the recognition results are extended to pixel-level segmentation by adding a branch to Faster R-CNN, which outputs a binary mask indicating whether a pixel is part of an object. See https: / / arxiv.org / abs / 1703.06870, the entire contents of which are incorporated into this disclosure by reference, making it a part of this disclosure.

[0095] Yolo (You Only Look Once): The above R-CNN systems are two-stage detectors, but Yolo is a one-stage detector that implements target detection and target classification in one network, which greatly improves performance. Through the integrated architecture design, it has less computation and faster speed. It performs convolutional learning on the entire image and comprehensively considers the context information of the entire image. Therefore, it can better learn the essential expression of the data set, with better generalization performance and higher recognition accuracy. Please refer to https: / / arxiv.org / abs / 2305.17786, and the entire content of this document is incorporated into this disclosure by reference, making it a part of this disclosure.

[0096] SSD (Single Shot MultiBox Detector): See https: / / arxiv.org / abs / 1512.02325. Like Yolo, it is also a one-stop detector. By running multi-scale feature maps, it has better detection effect for smaller objects. See https: / / arxiv.org / abs / 1512.02325. The entire contents of this document are incorporated into this disclosure by reference, making it a part of this disclosure.

[0097] In addition, deep learning methods (for example, see the document https: / / www.nature.com / articles / nature14539 in Nature, the entire content of which is incorporated into the present disclosure by reference here and becomes a part of the present disclosure) may also be used.

[0098] b) Object recognition and its application in traditional industries

[0099] With the continuous development of technology, computer vision has been widely used in society. The more common ones are as follows:

[0100] Face recognition: Face recognition is a biometric identification technology that identifies people based on their facial features. It automatically detects and tracks faces by collecting images or video streams containing faces, and then identifies the detected faces. It is currently widely used in public security, transportation, payment and other practical scenarios.

[0101] Smart transportation: Smart transportation is an important application field of target detection, such as traffic flow monitoring and traffic light timing control, traffic accident detection, traffic violation detection and tracking, and autonomous driving.

[0102] Industrial inspection: During the production process of products, various problems may occur due to the influence of raw materials, manufacturing processes, environment and other factors. Industrial inspection uses the target detection algorithm in computer vision technology to detect appearance defects such as cracks, deformation, and missing parts that occur during the production process of products, so as to improve product quality stability and production efficiency.

[0103] Medical field: In the field of medical image analysis, such as target detection on CT, MRI and other image data, it can help doctors identify the diseased areas inside the human body and improve the diagnosis and treatment of diseases.

[0104] c) Importance of image sources and image (object recognition) interpretation in the CRO (Contract Research Organization) field

[0105] In the CRO field, for example, in the process of gene synthesis, it is necessary to conduct quality control (QC) on the results of gene synthesis, including whether the size of the synthesized gene fragment is correct, whether it contains preset enzyme cutting sites, whether the quality of the extracted plasmid is qualified, etc. The commonly used detection method is to perform QC enzyme cutting verification. First, the synthesized gene sequence is enzyme-cut to form different gene fragment sizes; then, agarose gel electrophoresis is used to shoot the gel to obtain a gel map, and the band position target of the gel map is identified; finally, the identification result is compared with the ruler (marker) to obtain the size of the band, and then the QC test result is obtained. In order to save costs, the same gel (25 holes or even more holes) needs to be inspected for 7 to 8 orders at the same time. The gel map needs to be cut manually according to the order information, and then the cut order gel map is named and stored according to the order, and finally the order gel map is manually verified for QC results. After all steps are completed, upload to MES (Manufacturing Execution System) to execute the subsequent production process. With the continuous increase in order throughput and the influence of subjective factors in manual judgment of QC test results, a complete and effective set of QC test standards has not been formed. The original manual processing is becoming increasingly difficult to continue.

[0106] In this context, it is necessary to provide a method based on computer vision processing technology and image processing technology to use artificial intelligence (AI) to complete the automatic recognition and analysis of glue images, so as to be able to recognize and cut glue images and automatically determine QC results.

[0107] From the overall perspective of automatic gel image recognition and automatic QC result determination, Figure 1 A flow chart of a method for identifying and analyzing glue images using artificial intelligence according to an embodiment of the present disclosure is shown.

[0108] In general, the method for identifying and analyzing glue images using artificial intelligence according to the embodiment of the present disclosure can be divided into two parts: identification and quality analysis.

[0109] like Figure 1 As shown, the method 100 for identifying and analyzing gel images using artificial intelligence begins at step S110, in which an artificial intelligence (AI) recognition network model is used to identify the lanes and bands on the lanes in the original gel image of the electrophoresis image.

[0110] As mentioned above, the artificial intelligence recognition network model used in the present disclosure may be a convolutional neural network (CNN) model, such as R-CNN, Fast R-CNN, Faster R-CNN, Mask R-CNN, etc. More specifically, the artificial intelligence recognition network model used in the present disclosure may be a Yolo network model or an SSD network model. For the construction method of the artificial intelligence recognition network model for identifying the glue image, please refer to the detailed description below.

[0111] Those skilled in the art should understand that a gel image is a picture obtained by taking a photo after gel electrophoresis, and is in the form of a picture. The gel image will present an electrophoresis image corresponding to each order, that is, a combination of a plasmid lane, a product lane, and a ruler lane. In the present disclosure, the term "electrophoresis sample" is used to refer to such a combination on the gel image.

[0112] In the case where the gel map includes more than one electrophoresis sample, that is, the gel map includes multiple orders, the plasmid, product and marker lanes of each electrophoresis sample can be determined based on the lanes and bands identified by the artificial intelligence recognition network model, and the gel map can be cut based on each electrophoresis sample. Those skilled in the art should understand that, although in actual situations, it is necessary to cut the gel map before performing subsequent quality analysis operations, more generally, even if cutting is not performed, it does not affect the execution of subsequent operation steps of the present disclosure.

[0113] However, the electrophoresis image is skewed when photographed, so there may be a preferred step for correcting the gel image. In step S110, the gel holes in the original gel image may be identified using an AI recognition network model; the horizontal deflection of the original gel image may be corrected according to the deflection angle of the horizontal position of the identified gel holes; and the lanes and bands on the lanes in the corrected gel image may be identified using the AI ​​recognition network model.

[0114] Specifically, first, the AI ​​model can be used to identify all the rulers, gel holes and band targets in the gel image. Then, the horizontal position of the first gel hole and the last gel hole is calculated. Through trigonometric functions, the deflection angle can be calculated. If it exceeds 1 degree, image correction is required. After correction, the AI ​​model can be used again for identification to obtain the corrected gel hole, ruler and band information. Then, in conjunction with the order information, the image is cut from left to right according to the ruler-to-ruler logic (that is, in units of each electrophoresis sample). In practice, there may be a situation where the electrophoresis sample corresponding to a single order has two ruler lanes, but this situation does not affect the above-mentioned ruler-to-ruler cutting logic. It is just that the correct ruler lane needs to be determined in subsequent operations (when calculating the product band size, the ruler closest to the product lane is used for calculation. When it exceeds the range of the ruler, another ruler can be selected for calculation).

[0115] Next, for the uncut or cut glue images, further correction can be performed. For example, all the scale strips can be found by using image processing methods. First, the image is preprocessed, and a Gaussian filter with a kernel size of 3*3 is used to smooth the image; then a mean filter with a kernel size of 3*3 is used to reduce noise; finally, the image is converted into a grayscale image. After the preprocessing is completed, in order to solve the problem that some of the glue image product strips are too dark, it is necessary to increase the brightness of the image and improve the contrast. After processing, the image is binarized, that is, only two colors (black and white) are left. After this processing, there will still be a lot of noise, and the image needs to be further processed. Therefore, the image can be closed (the image contour (black and white junction) is first enlarged and then reduced) to remove the external small contour; then an erosion operation (reducing the image contour) is performed to remove the internal contour; finally, the contour search is used to find all possible contours; then for each contour, its circumscribed rectangle is drawn and the coordinates of the circumscribed rectangle are calculated. Finally, count whether the circumscribed rectangle of the first scale strip is on the same vertical line as the circumscribed rectangle of the last strip, and calculate the deflection angle. If it is less than 2.5 degrees, it can be corrected by rotating the image. Otherwise, it cannot be corrected automatically and manual processing is required to check whether the image needs to be retaken or can be manually rotated and corrected.

[0116] Next, in step S120, the identified swimming lanes and bands on the swimming lanes are analyzed. According to a preferred embodiment of the present disclosure, the analysis of the identified swimming lanes and bands on the swimming lanes in step S120 can be specifically a quality analysis for quality control of the gel image.

[0117] The QC result analysis can be divided into the following parts:

[0118] a) Genome detection and band analysis: The AI ​​prediction network model can be used to preliminarily predict whether there is a genome in the gel image. When the genome is preliminarily predicted to exist in the gel image, the content of the genome band is determined according to the positional characteristics of the genome band to be determined in the gel image and the quantitative analysis method. The positional characteristics of the genome can be the position of the band that exceeds the top edge of the ruler on the product lane. Specifically, according to the position of the genome in the image, it is compared with the ruler to check whether it meets the positional characteristics of the genome to determine whether it is a false alarm. If the identification position is in the plasmid lane, near the gel hole or within the ruler range, it is considered a false alarm; if it is not a false alarm, the genome content needs to be calculated. If the genome content does not exceed 15%, it can be released; otherwise, it will be intercepted. For the specific calculation method, please refer to the detailed description of the specific embodiments below.

[0119] b) Detection and analysis of fuzziness and degradation: The AI ​​prediction network model can be used to preliminarily predict whether there is a fuzziness or degradation area in the glue map. When it is preliminarily predicted that there is a fuzziness or degradation area in the glue map, the degree of fuzziness or degradation is further judged based on the upper and lower positions of the preliminarily predicted fuzziness or degradation area and the percentage of the brightness of the fuzziness or degradation area and the strip brightness of the identification area. For specific release standards, please refer to the detailed description of the specific embodiments below.

[0120] c) Calculation of product band size: The type of ruler can be determined based on the bands in the ruler lane identified by the AI ​​recognition network model in step S110. The size of the bands in the product lane is determined based on the relative position relationship between the bands in the product lane and the bands in the ruler lane identified by the AI ​​recognition network model in step S110. Specifically, the previous steps have used the AI ​​recognition network model to identify all the product bands and rulers in the gel image, and grouped the bands in the order of the three lanes: plasmid, product (may include product 1, product 2, product..., that is, there may be multiple product lanes), and ruler. The KB type ruler has 9 bands, DL3000 has 8 bands, and DL15000 has 7 bands. Figure 2 Standard scale reference diagram ( Figure 2, the appearance of the standard scale bands of KB Ladder, DL3000 and DL15000 and the molecular weight of each band are shown in the figure. According to the characteristics of the identified band data and the spacing between the last two bands, the scale type of the current sample is distinguished as KB, DL3000 or DL15000. Finally, according to the position of the scale band in the image, the product mobility, genome, brightness of the product band, center point of the band, height of the band and error are considered, and the molecular weight of the product is calculated. Then, according to the theoretical band size of the enzyme cutting of the order, it is determined whether the calculated product molecular weight meets the QC requirements. For the specific calculation method, please refer to the detailed description of the specific embodiment below.

[0121] d) Impurity detection and analysis: The AI ​​prediction network model can be used to preliminarily predict whether there are impurities in the gel image. When it is preliminarily predicted that there are impurities in the gel image, the position, size and brightness of the preliminarily predicted impurities are used to determine whether they affect the determination of the product bands.

[0122] Other error detection (i.e. quality analysis) mainly includes ruler integrity detection, product enzyme digestion incompleteness inspection, band aesthetics inspection, whether the plasmid band is supercoiled and plasmid band dropout, etc.

[0123] As mentioned above, the process of manually cutting the glue map is time-consuming and laborious, and large-volume orders and manual processing are prone to problems such as order stringing. When manually reviewing QC results, due to the influence of human subjective factors and the lack of unified QC testing standards, large-volume orders can easily cause human eye fatigue and inability to detect subtle differences in similar glue maps, such as weak bands that cannot be detected, unseparated bands that have run, and impurities of different shapes that cannot be determined.

[0124] The method disclosed in the present invention is based on AI machine vision processing technology and image processing technology to identify plasmids, products, and ruler lanes, and determine the specific positions of corresponding lanes and bands in the image. It can perceive subtle differences between similar gel images, and the judgment of analysis results is not affected by personal subjective factors, thereby providing accuracy and stability of analysis results.

[0125] The following is a detailed description of the process of constructing an artificial intelligence recognition network model for identifying glue images and the process of constructing an artificial intelligence prediction network model for quality control of glue images according to an embodiment of the present disclosure.

[0126] Figure 3A It is a flowchart of a method 300A for constructing an artificial intelligence recognition network model for identifying glue images used in the present disclosure.

[0127] It will be understood by those skilled in the art that by, for example Figure 3A The AI ​​recognition network model obtained by this method can be used for Figure 1The method for identifying and analyzing gel images is shown in .

[0128] like Figure 3A As shown in , the construction method 300A starts at step S310A, in which a training set can be obtained. The training set includes multiple training samples, each of which includes an original image of an electrophoresis gel image and an annotation of the image. The annotation of the image includes annotations of lanes in the image and bands on the lanes.

[0129] In step S320A, based on the original image of the electrophoresis gel image and the annotations of the image, an initial artificial intelligence recognition network model is trained to obtain an artificial intelligence recognition network model for identifying lanes and bands on the lanes in the original image of the electrophoresis gel image.

[0130] The marking of the picture also includes the marking of the glue holes in the picture.

[0131] The initial artificial intelligence recognition network model described in step S320A may be a convolutional neural network (CNN) model, such as R-CNN, Fast R-CNN, Faster R-CNN, Mask R-CNN, etc. More specifically, the initial artificial intelligence recognition network model may be a Yolo network model or an SSD network model.

[0132] In order to be able to automatically identify the glue image and further optionally perform automatic cutting, the AI ​​model can be trained and constructed as follows: collect the original pictures, and then use labeling software (such as LabelImg software, "Elf Labeling Assistant", etc.) to label and classify the pictures to form the original learning data set (training data set) and verification data set required for AI training; then use an AI neural network such as Yolo to perform AI training. In this process, it is also necessary to optimize hyperparameters, such as learning rate, loss coefficient, hue, saturation, rotation, brightness and other hyperparameters, using SGD (Stochastic Gradient Descent), Adam (Adaptive Gradient Algorithm, see https: / / arxiv.org / abs / 1412.6980v8, the entire contents of which are incorporated into this disclosure by reference, making it a part of this disclosure; in addition, Adam's variant NADAM: http: / / cs229.stanford.edu / proj2015 / 054_report.pdf, the entire contents of which are incorporated into this disclosure by reference, making it a part of this disclosure) optimizers to optimize model parameters. Among them, SGD continuously adjusts the parameters of the model through the method of gradient descent to minimize the loss function of the model. Its advantages are simple implementation and high efficiency, and its disadvantages are slow convergence and easy to fall into local minimum. Adam adjusts the model parameters by maintaining the first-order momentum and second-order momentum of the model's gradient and the square of the gradient. Its advantages are high computational efficiency and fast convergence, but its disadvantage is that it requires adjustment of hyperparameters. Finally, the obtained model is converted into a format such as Openvino to achieve optimal inference performance. Openvino is a semi-open source toolkit developed by Intel specifically for optimizing and deploying artificial intelligence inference, mainly used to optimize deep inference.

[0133] Those skilled in the art should realize that Figure 3A The method can obtain an artificial intelligence recognition network model for identifying glue images.

[0134] Figure 3B It is a flowchart of a method 300B for constructing an artificial intelligence prediction network model for analyzing glue images used in the present disclosure.

[0135] It will be understood by those skilled in the art that by, for example Figure 3B The AI ​​prediction network model obtained by this method can be used for Figure 1 The analysis operation in the method of identifying and analyzing gel images is shown in .

[0136] like Figure 3BAs shown in , the construction method 300B starts at step S310B, in which a training set can be obtained. The training set includes multiple training samples, each of which includes an original image of an electrophoresis gel image and an annotation of the image. The annotation of the image includes annotations of items with gel image quality problems and their locations.

[0137] In step S320B, based on the original image of the electrophoresis gel image and the annotation of the image, the initial artificial intelligence prediction network model is trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether there are quality problems in the gel image.

[0138] It should be noted that the "items of gel image quality problems" mentioned here refer to defects that may be found in QC testing or problem items that cause the gel running results to fail. In a preferred embodiment of the present disclosure, such quality problem items include but are not limited to: the presence of genomes, the presence of fuzzy cutting / degradation, and the presence of impurities. Therefore, in a preferred embodiment of the present disclosure, the artificial intelligence prediction network model can be trained and constructed specifically for a certain quality problem. Specifically, three special artificial intelligence prediction network models can be trained and constructed for QC testing of genomes, QC testing of fuzzy cutting / degradation, and QC testing of impurities, respectively.

[0139] Specifically, for the QC detection of the genome, the annotation of the image includes the annotation of the genome and its location. Therefore, based on the original image of the electrophoresis gel image and the annotation of the image, the initial artificial intelligence prediction network model is trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether there is a genome in the product lane. You can first collect images with genomic features; then use annotation software (such as "Elf Annotation Assistant") to annotate and classify the locations with genomic features; divide the data into 90% for the training set and 10% for the validation set, use the initial AI model (such as Yolo) for learning and training, obtain an AI model for genomic features, and then use the model to predict whether there is a genome in the gel image on the test set. After multiple training and adjustments, retain the model with the best prediction effect on the test set.

[0140] For QC detection of cutting blurring / degradation, the annotation of the image includes the annotation of the cutting blurring or degradation area and its location. Therefore, based on the original image of the electrophoresis gel image and the annotation of the image, the initial artificial intelligence prediction network model is trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether there is a cutting blurring or degradation area in the gel image. You can first collect images containing cutting blurring and degradation characteristics; then use annotation software (such as "Elf Annotation Assistant") to annotate and classify the locations with cutting blurring and degradation characteristics; divide the annotated data into 90% training set and 10% validation set, and use the initial AI model (such as Yolo) for model training to obtain a constructed AI model.

[0141] Similarly, for QC detection of impurities, the annotation of the image includes the annotation of the impurities and their locations. Therefore, based on the original image of the electrophoresis gel image and the annotation of the image, the initial artificial intelligence prediction network model is trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether there are impurities in the gel image. You can first collect images with impurity features; then use annotation software (such as "Elf Annotation Assistant") to annotate and classify the locations with impurity features; divide the data into 90% for the training set and 10% for the validation set, use the initial AI model (such as Yolo) for learning and training, obtain an AI model for impurities, and then use the model to predict whether there are impurities in the gel image on the test set. After multiple training and adjustments, retain the model with the best prediction effect on the test set.

[0142] The initial artificial intelligence prediction network model described in step S320B may be a convolutional neural network (CNN) model, such as R-CNN, Fast R-CNN, Faster R-CNN, Mask R-CNN, etc. More specifically, the initial artificial intelligence prediction network model may be a Yolo network model or an SSD network model.

[0143] It should be noted here that the AI ​​prediction network models used for genome detection, slicing or degradation detection, and impurity detection are different network models trained and constructed separately. In other words, if the quality analysis includes multiple detection operations, multiple AI prediction network models that are trained and constructed separately may be used simultaneously.

[0144] Those skilled in the art should realize that Figure 3B The method can obtain an artificial intelligence prediction network model for analyzing glue images.

[0145] In addition, since the initial models of the AI ​​recognition network model used to identify gel images and the AI ​​prediction network model used for analysis are both based on CNN or Yolo, SSD, in a preferred embodiment, the AI ​​recognition network model and one or more AI prediction network models used for different analysis purposes (genomics, gel cutting / degradation, impurities, etc.) can share the same CNN or Yolo, SSD framework, but belong to different feature sets and neural network parameters.

[0146] Those skilled in the art should understand that in addition to training and building the AI ​​recognition network model and the AI ​​prediction network model through the training set, the test set can also be used for testing. After the test set is tested, the parameters of the AI ​​recognition network model and the AI ​​prediction network model can be adjusted according to the test results.

[0147] On the other hand, in the calculation of product band size, the model training has been completed through the previous training of the model for automatic identification of gel images (i.e., AI recognition network model), namely: first collect original gel images of various types; use annotation software (such as "Elf Annotation Assistant") to annotate and classify bands, lanes, etc.; divide the training set and validation set according to the ratio of 90% and 10%, and the initial AI model (such as Yolo) performs model training and obtains an applicable AI model. Therefore, the model can be used to predict all product bands and scales in the gel image, and the band size can be calculated later according to the method described above.

[0148] In addition, for other QC detection processes, such as ruler integrity detection, product incomplete enzyme digestion detection, plasmid band detection, mixed band or dropped band detection, etc., the previously trained and constructed AI recognition network model for automatic identification of gel images can be used to obtain the identification results of the divided lanes and the bands on them, combined with computer vision and image processing technology to complete the corresponding detection.

[0149] Specific embodiments will be described below.

[0150] Embodiment 1:

[0151] Example 1 involves QC automatic gel image recognition and optional cutting.

[0152] A total of 370 pictures were collected from the original glue map. The pictures were labeled and classified using the software "Elf Labeling Assistant". Figure 4 As shown in the figure, position ① is the annotation object, and the gel wells, lanes (such as plasmids, products, rulers) and bands on the lanes need to be annotated; position ② is the corresponding classification information. After the annotation is completed, the data is exported, and the exported data includes pictures, annotation information and corresponding classifications.

[0153] Export the labeled data, divide the training data into 90% of the data set and 10% of the validation set, and finally perform model training.

[0154] For the original QC gel image, Figure 5 As shown, using the trained model to predict the gel image, you can get all the required target information, such as gel holes, lanes (including plasmids, products, rulers), bands, and their position coordinates in the image. Since it is a cutting requirement, you do not need to know the specific ruler type and completeness. A complete gel image is arranged in order from left to right, namely plasmid, product (product 1, product 2), and ruler. Therefore, between two adjacent rulers is a complete gel image information, that is, an electrophoresis sample corresponding to a single order. You only need to cut out this piece of image. If the ruler is incomplete, you can also refer to the gel hole, and you can also locate the specific coordinate position of the cut image, such as Figure 6 shown.

[0155] According to the coordinate position of the ruler in the image, the glue image is cut from left to right in a ruler-to-ruler manner, and the cutting height is from the glue hole above the ruler to the end of the ruler position or the glue hole below. As mentioned above, the presence of two rulers in a single order does not affect the cutting. Then, for each original glue image cut out, use Gaussian filter and mean filter to reduce image noise and other preprocessing; then convert it into a grayscale image; then adjust the brightness; use the threshold method to convert the grayscale image into a binary image (only black and white); perform further corrosion operations on the binary image; finally, find the contour and calculate the circumscribed rectangle of the outer contour to obtain the strip position, such as Figure 7 Finally, analyze the relative position of the first circumscribed rectangle and the last circumscribed rectangle, calculate the deflection angle through trigonometric functions, and complete the calibration of the ruler. Finally, save it as Figure 8 shown.

[0156] Embodiment 2:

[0157] Example 2 relates to the automatic analysis process of QC results.

[0158] The overall detection process of Example 2 is as follows Fig. 9 As shown. Those skilled in the art should understand that, in practice, for quality analysis, multiple operations are often required, that is, not just one operation, but it is hoped to implement as comprehensive quality analysis as possible in order to achieve a more comprehensive quality control effect. For example, Fig. 9 A combined process of multiple quality analysis operations is presented. Fig. 9In the process, the order of each quality analysis operation is meaningful. For example, genome and cleavage or degradation detection needs to be performed before calculating the molecular weight of the product band. That is to say, in the preferred embodiment of the present disclosure, the detected genome region or the region where cleavage / degradation is detected does not participate in the calculation of the band size. In addition, impurity detection can be performed after calculating the molecular weight of the product band. Accordingly, the AI ​​prediction network model related to these quality analysis operations will also be used in the same order for preliminary prediction. However, those skilled in the art should understand that Fig. 9 This is merely an example and is not meant to be limiting of the possible combinations or possible sequences of the mass analysis operations of the present disclosure.

[0159] In order to complete the QC result detection, AI models are also required to complete genome prediction, pasting / degradation prediction, band size calculation and impurity prediction in gel images. Based on the prediction results and calculation results, relevant algorithm tests are performed to finally confirm whether release or interception is required.

[0160] glue Figure 1 The order is usually plasmid lane, product lane, and marker lane. Fig.10 shown.

[0161] 1. Genome Prediction

[0162] Genomic characteristics: In the product lane, there will be a band that is not within the marker range and is close to the marker position, such as Fig.11 As shown, the genome framed position, the band close to the top, is the genome band.

[0163] 1. First, collect the original gel images of the genome, a total of 402 images, and use the Wizard Annotation Assistant software to annotate and classify the gel images. Due to the characteristics of the genome, we only need to know the relationship between its position and the ruler. Fig.12 As shown, ① is the marking area, and ② is the classification information corresponding to the marking.

[0164] 2. Export the labeled genome gel data and divide the data into 90% training set and 10% validation set. Then use the Yolo model for training. The training method and training steps are the same as the model training process for gel image recognition. Finally, export the trained model and convert it into openvino format.

[0165] The genome prediction results are as follows Fig.13 As shown, the genome and ruler are framed respectively, and the prediction score is given next to the genome (the higher the score, the more accurate the prediction).

[0166] 3. Genome release rules:

[0167] According to the characteristics of the genome, firstly, according to the lane position, only the identification results under the product lane are retained, and then according to the relative position of the ruler, it is determined whether the genome band exceeds the upper edge of the ruler.

[0168] Then, the genome content is calculated, and if it is less than 15%, it can be released, otherwise it cannot be released. Fig.14 The genomic band area and the corresponding brightness peak diagram shown are shown. The content calculation needs to refer to the band position in the product lane identified by the AI ​​model. The genomic content calculation process is as follows: first, all the images in the lane where the genome is located are cropped out; then a circular kernel filter with a radius of 10 is generated; the image in the lane is subjected to the OSU method (see OTSU N. "A threshold selection method from gray-level histo-grams" [J]. IEEE Transactions on Systems, Man, and Cybernetics, 1979, 9 (1): 62-66, which is incorporated into the present disclosure by reference, making it a part of the present disclosure) to calculate the segmentation threshold T of the image foreground and background; select all pixels less than the T value, and for these selected pixels, use the circular kernel filter with a radius of 10 just generated to perform a closing operation on the image (enlarge and then reduce the position with a boundary in the image, such as the position of the band boundary) to obtain the result matrix; then use the matrix to select the image, if the value is true, use the original image, otherwise subtract the background image.

[0169] After subtracting the background, combined with the band identification, calculate the average value of the 6 rows above and below the band for each band. Take the minimum of the two upper and lower values, and use this value as the basis to subtract these values ​​from the pixel values ​​within the band range. If a negative value is obtained after subtraction, it is calculated as 0. Then add up all the remaining values ​​within the band range to get the content value of the band. All bands in the lane are calculated in this way to obtain the band content. Finally, add up all the band contents, divide the content of each band by the total content, and calculate the band content. Genomic band content can be released when it is less than 15%.

[0170] like Fig.15 According to the calculation results, the contents of each band are (from top to bottom) 23749, 53832, and 3205. The first band is the genome band, and its content accounts for 29.39%, so it cannot be released.

[0171] 2. Paste cutting / degradation model prediction

[0172] Blurring / degradation characteristics: Blurring usually occurs in the product lane, which is manifested as band dragging and unclear bands or no bands at all. Degradation is manifested by the presence of dragging and unclear bands in both the plasmid lane and the product lane. If blurring or degradation occurs above the band, if the data is accumulated by row, it will show a gradual increase until it reaches the maximum value (the highest position of the band); if it occurs below the band, it will show a gradual darkening until it merges with the background color. Blurring and degradation can be as follows Fig.16 shown.

[0173] 1. First, collect the original glue pictures of specific cutting and degradation, a total of 496 pictures, use the "Elf Labeling Assistant" software to complete the labeling and classification of the pictures, ① is the labeling area, ② is the labeling classification information area, such as Fig.17 shown.

[0174] 2. Export the labeled data through the software. Then divide the data into 90% training set and 10% validation set. Use the Yolo model to train the divided data. The training method and steps are the same as the model training process of the glue map cutting. Finally, export the trained model and convert it into openvino format.

[0175] The prediction demonstration of the cutting / degradation model is as follows Fig.18 , 19 As shown, the locations of degradation and fuzz cutting are marked respectively, and the prediction scores are given in detail.

[0176] 3. Cutting and degradation judgment process

[0177] Since the characteristics of fuzzy cutting and degradation are similar, the proportion of model false alarms is high, and secondary confirmation is required through algorithms. The specific process is as follows: First, use the model to predict the gel map; if there is a predicted fuzzy cutting or degradation area, detect the relationship between the identified area and the lane, and correct the classification error; after the correction is completed, the corresponding identified area is taken out separately; if it is in the plasmid lane, first detect whether there is a tail in the area, and according to the characteristics of the tail (a process in which the brightness gradually dims), detect whether there is really a tail in the area, so as to confirm whether it is degradation, otherwise it is fuzzy cutting. After the category determination is completed, start to confirm the area. Fuzzy cutting or degradation is unlikely to occur at the gel hole position. If it is at the gel hole position and the score is less than 0.4 points and the area height is less than 40, it can be released; if it cannot be released, further detection is performed. First, calculate a background value at the upper and lower positions of the identification area and find the average value; then convert the area into a grayscale image; then find the average value by row, and only keep the data rows that exceed 8% of the background brightness for standby; then use this data to draw a peak diagram; then search for peaks with a peak width of 8 and an amplitude of 3.97; if not found, it means that the strip is completely degraded and needs to be intercepted; if multiple peaks are found, it means that there are multiple strips in the area, and each strip needs to be detected at this time. Because the tailing situation generally occurs above or below the strip. If there is only one peak, divide the data into two parts according to the location of the peak, and check the two parts separately. For the upper part, the characteristic is gradually brightening, and the lower part is gradually darkening. Through these two pieces of data, the percentage of the brightness of the cut or degraded area to the brightness of the strip can be calculated. If it is less than 10%, it can be released. For the case where multiple peaks are found, the calculation can be performed in the same way.

[0178] 3. Calculation of product band size

[0179] 1. First, collect the original glue images, a total of 603 images, and then use the "Elf Labeling Assistant" to complete the data labeling of the glue images. After exporting, divide the images and labeling data into 90% training set and 10% validation set. Use Yolo for training, and the training method and training steps are the same as the model training process of glue image cutting. Finally, export the trained model and convert it into openvino format. Fig. 20 As shown, ① is the data annotation area, and ② is the annotation category.

[0180] 2. The trained model can predict the gel images that are not in the training set. It only needs to identify all the bands and markers in the gel image, and then divide the bands according to the order of plasmid, product, and marker. The bands can be distinguished into different lanes and further processed. Fig.21 shown.

[0181] Algorithm correction process:

[0182] Each band should be from top to bottom, and there should be no intersection. If there is any intersection, it may be a false positive. In addition, according to big data statistics, the width of the band is generally about 70px and the height is about 25px. Preliminary filtering of the bands to remove the crossed bands. Calculate the average width and height of all bands in each lane, and perform secondary filtering on the bands based on the calculated average value. After filtering, refer to the standard scale chart (reference Figure 2 ). According to the standard ruler diagram, (1) the number of bands in each ruler is different, and (2) the distance between the last two bands is quite different. These two characteristics can be used to determine the type of ruler. In other words, the type of ruler can be determined based on the number of bands in the ruler lane and the distance between the last two bands.

[0183] like Fig.21 In the example, after filtering, there are 7 strips left. According to the number of strips, it should belong to the DL15000 type of ruler. However, the distance between the last two strips does not conform to the DL15000 strip. This means that the last strip is actually identified as one strip because the distance between the two strips is too close. At this time, the last strip needs to be split into two strips, so that it has 8 strips, and the ruler is finally identified as the DL3000 type.

[0184] After the scale type is determined, the position of the strip needs to be adjusted accurately. Ensuring the accurate position of the strip can reduce the error caused by the subsequent calculation of the product strip size, such as Fig. 22 The error caused by the sharp corners in the image. Through model recognition, the coordinate position of each strip in the image can be obtained, mainly including the coordinates of the upper left corner, width and height. According to the coordinates, width and height, the area where the strip is located can be obtained. After the area is cropped out, 10% of the width is deducted from the left and right sides respectively as the main body of the strip, that is, the sharp corners on the left and right sides are excluded. After getting the main content of the strip, the pixel values ​​are accumulated row by row and the average value is calculated. Then the maximum value of the main area, that is, the row with the highest brightness, can be obtained. Refer to this brightness value and compare it with the brightness in other ranges. Normally, the brightness of a strip is relatively evenly distributed. When a boundary appears, the brightness drops sharply. Based on this feature, the strip is further optimized to obtain a more accurate strip position.

[0185] 3. Stripe size calculation

[0186] The band size calculation is mainly based on the relative position of the product band and the standard band of the ruler. According to the standard ruler, the band size corresponding to each band can be known.

[0187] Specifically, determining the size of the bands in the product lane according to the relative positional relationship between the bands in the product lane and the bands in the ruler lane in the sample identified and cut out by using the artificial intelligence model further includes: determining the size of each band in the ruler lane in the sample according to the determined ruler type and referring to the band sizes corresponding to each band in the standard ruler diagram of the corresponding type; determining the size of the bands in the product lane according to the relative positional relationship between the bands in the product lane in the sample and the bands in the ruler lane.

[0188] More specifically, determining the size of the band in the product lane according to the relative positional relationship between the band in the product lane and the band in the ruler lane in the sample further includes: judging whether the band in the product lane is flush with a band in the ruler lane; when the band in the product lane is flush with a band in the ruler lane, the size of the band in the product lane is the size of the band in the ruler lane that is flush with it.

[0189] The corresponding sizes of the upper edge L1, center line L2 and lower edge L3 of the product strip can be calculated, and then the calculated results can be compared with the theoretical size to determine whether the strip size is correct. Fig.23 That is, the size of the band in the product lane is determined by calculating the relative positional relationship between the upper edge L1, the lower edge L3 and the center line L2 of the band in the product lane and the band in the ruler lane.

[0190] In the process of calculating the strip size, according to the scale type and strip calibration position coordinates determined above, the size of each strip of the scale can be known. Select the upper edge of two adjacent scale strips as the reference point to calculate the size of the product strip in the middle. If the lower edge or center line is used as the reference, the adjacent strips will be paired, and part of the entire scale range will be missing.

[0191] Product band size calculation process:

[0192] like Fig.23 As shown in Figure ①, you can see that there are two bands in the product lane to be calculated. The position of the first product band is close to a band in the ruler. So first, you need to check whether the product band is flush with the ruler band. If they are flush, the band size is consistent with the corresponding size of the ruler band, and it has the highest priority. The judgment standard is whether 60% of the product band height is within the range of a certain ruler band. Fig.23 In Figure ①, the first strip does not meet this rule, so it is necessary to calculate the upper edge L1, center line L2, and lower edge L3 of the current product strip. The advantage of this calculation is that it does not need to consider the strip crossing the scale strip (such as Fig.23To reduce the computational difficulty brought about by (as shown in Figure ③), the calculation method can be standardized.

[0193] Next, we calculate L1. Based on the height of L1 in the image, we divide the image into segments from top edge to top edge in the ruler strip, and see which segment L1 falls in. Fig.23 It can be seen that it falls between the 3rd and 4th stripes. From the standard diagram, it can be concluded that this ruler is of KB type, and the corresponding sizes of the 3rd and 4th stripes are 6000 and 5000 respectively. Through the span of the upper edge of the 3rd and 4th stripes, such as 50 lines, the stripe size difference is 6000-5000=1000, and it can be calculated that the difference in each line is 20BP. Then, calculate how many lines the difference between L1 and the upper edge of the 3rd stripe of the ruler is, for example, 40 lines, then it can be calculated that the corresponding size of L1 is 6000-40*20=5200.

[0194] Refer to the same method to calculate L2 and L3.

[0195] Calculate L1, L3, just can calculate intermediate value, calculation formula is L1-(L1-L3) / 2.Can compare this intermediate value with actual L2 calculated value, take one closest to the theoretical value of enzyme cutting (product) band from two, as final calculation result.That is to say, by calculating the relative positional relationship between the upper edge, lower edge and the band in the scale swimming lane of the band in the product swimming lane, determine the intermediate value of the size of the band in the product swimming lane.By calculating the relative positional relationship between the center line of the band in the product swimming lane and the band in the scale swimming lane, determine the center line value of the size of the band in the product swimming lane.Finally, compare described intermediate value with described center line value, take the size of the band in the product swimming lane that is closest to the theoretical value of enzyme cutting.

[0196] At the same time, the human eye difference must also be considered, so that when calculating the size of the stripe, adjustments should be made based on the human eye's perception error. Fig.23 Figure ② in the figure shows that the strip is exactly in the middle of the adjacent strips of the ruler. Therefore, in addition to calculating L1, L2, and L3 as above, it is also necessary to consider the visual differences of the human eye. From the perspective of the human eye, the center point of the strip is viewed as the entire strip, such as L6 and L7 in Figure ②. L6 corresponds to the center point from the upper edge L4 to the upper edge L5 of the strip. L7 corresponds to the center position of the adjacent strips in the human eye's perception. Due to the existence of this difference, when calculating the size, it is necessary to take into account this difference in human eye perception, that is, the actual calculation result, it is also necessary to add 33% of the upper strip height of the current ruler strip used for calculation. Assuming that the strip height is 10, 33% is 3.3, then it is necessary to add this difference to the original calculation result, that is, 3.3*20+5200=5260.6. The size calculated in this way is closer to what the human eye actually sees.

[0197] 4. Determination of impurities

[0198] 1. First, collect 198 original glue images with impurities, such as Fig.24 As shown, use the Elf Software Assistant to complete the labeling and classification of the data. Finally, use the export function of the software to export the labeled data. After exporting, split the data according to the ratio of 90% for the training set and 10% for the validation set. Use the Yolo model to complete the model training. The training process and method are the same as the glue map cutting. Finally, export the trained model to openvino format.

[0199] 2. Use model prediction

[0200] Impurity characteristics: The position, size, and shape of impurities in the glue image are not fixed, but the brightness is generally very high. It is very difficult to detect impurities using traditional image processing methods, and there is no good detection method. However, the use of AI models can solve this problem well. The trained model can be used to predict impurities to detect whether the image has impurities, such as Fig.25 shown.

[0201] 3. Impurity release standard

[0202] Impurities can be released as long as they do not affect the determination of the product band. Specifically, if the impurities are near the gel hole and are smaller than 20*20PX, they are not considered impurities; if the impurities are not within this range, they can be released if the number is less than 3 and they are not on the band and are smaller than 20*20PX; if the impurities are on the band, regardless of size, they will be intercepted.

[0203] 5. Overall testing process

[0204] The overall detection process is divided into 8 detection tasks, namely genome detection, fuzzy or degradation detection, band size calculation and detection whether it meets the requirements, scale integrity and category detection, incomplete enzyme cutting detection, impurity detection, plasmid band detection (including the presence or absence of plasmid bands, aesthetics, etc.), and miscellaneous band or dropped band detection.

[0205] The genome detection, fuzzy cutting or degradation detection, band size calculation, and impurity detection have been described in detail above. Here we will focus on the scale integrity detection and category detection, incomplete enzyme cutting detection, plasmid band detection, and mixed or dropped band detection.

[0206] 1. Scale integrity detection

[0207] According to the band size model, the position of the ruler in the image can be known. According to the standard ruler, the type of ruler can be determined by the number of strips and the distance between the last two strips in the ruler strip. If the number is insufficient or the category cannot be determined, an interception and an alarm will be given. At the same time, according to the relative position of the ruler in the image, the position of the ruler in the image can be obtained. If the coordinates of the ruler are very close to the right side of the overall gel image, when it is less than 1px, an interception and an alarm will be given. In addition, the relative width of the bands in the product lane can be checked and compared with the width of the bands in the ruler lane. If the width of the bands in the ruler lane is significantly lower than the width of the bands in the product lane, it indicates that the ruler may be incomplete and needs to be intercepted.

[0208] 2. Detection of incomplete product digestion

[0209] Incomplete enzyme digestion means that when the plasmid is digested, some of the fragments are not completely cut due to various reasons. This will cause incomplete enzyme digestion. On the gel image, there are more than three bands in the product lane and the size of the first band is the sum of the sizes of the remaining bands. When multiple bands are found, the above logic can be used to make a judgment.

[0210] 3. Plasmid band detection

[0211] Through the stripe size model, the stripes in all lanes can be obtained. At this time, it is necessary to detect whether there are stripes in the plasmid lane. If not, intercept and alarm; if there are stripes, refer to the stripe size calculation, use the same calculation method to calculate the corresponding size of the plasmid stripe. For the stripe aesthetics check, use the Hu moment (see Monge-Alvarez, J., Hoyos-Barceló, C., Lesso, P., & Casaseca-de-la-Higuera, P. (2018). Robust detection of audio-cough events using local Hu moments. IEEE Journal of Biomedical and Health Informatics, 23 (1), 184-196, hereby by citing, the entire content of this document is incorporated into the present disclosure and made part of the present disclosure) to calculate the image moment. The Hu moment is a set of 7 variables calculated using the central moment that is invariant to image transformation. The first 6 moment invariants are invariant to translation, scaling, rotation and mapping, while the 7th moment will change due to image mapping. A set of standard strip images can be prepared in advance, and the contours in the standard images can be calculated and saved. Then, the contours of the strips to be checked can be calculated and compared with the standard contours to obtain the similarity scores with each standard strip. The scores are sorted from high to low, and the highest score is taken. If the highest score is higher than 0.85, the strip is considered unsightly. Fig.26 , shows a standard strip diagram.

[0212] 4. Detection of stray or dropped tapes

[0213] The miscellaneous bands are relative to the product lanes. Under normal circumstances, there should be only two bands after the plasmid is cut by enzymes. If multiple bands are detected, and there are still multiple bands after deducting factors such as the genome, they are considered to be miscellaneous bands and need to be intercepted and alarmed.

[0214] Band drop refers to the plasmid lane. Under normal circumstances, without enzyme digestion, the molecular weight of the last band in the plasmid lane should be the sum of the band sizes in the product lane. The calculation method is: based on the value of 60% of the sum of the theoretical sizes of all bands in the product lane, and referring to the band size position of the ruler, it can be inferred that the value should be at a specific position in the image relative to the ruler; then the calculated position is compared with the position of the last band in the plasmid lane to check whether it is a band drop. If so, intercept and alarm.

[0215] All of the above processes can be presented through a graphical user interface (GUI).

[0216] Those skilled in the art should understand that based on the method of identifying and analyzing glue images using artificial intelligence disclosed in the present invention, a system for identifying and analyzing glue images using artificial intelligence can be developed. Fig. 27 The schematic block diagram of the system for identifying and analyzing glue images using artificial intelligence according to the present disclosure is shown. Specifically, the system 2700 for identifying and analyzing glue images using artificial intelligence includes an acquisition module 2710 and a processing module 2720 .

[0217] The acquisition module 2710 is used to acquire the original gel image of the electrophoresis image.

[0218] The processing module 2720 is used to identify and analyze the glue map using artificial intelligence. Fig. 27 As shown in , the processing module 2720 may further include: an identification submodule 2721, used to identify the lanes and bands on the lanes in the original gel image of the electrophoresis image by using an artificial intelligence recognition network model; an analysis submodule 2722, used to perform quality analysis on the identified lanes and bands on the lanes to perform quality control on the gel image.

[0219] Those skilled in the art should understand that the above-mentioned system for recognizing and analyzing gel images using artificial intelligence can be implemented by a computer. For example, the following operations can be implemented by executing a computer program: obtaining an original gel image of an electrophoresis image; recognizing and analyzing the gel image using artificial intelligence, and the operation further includes the following sub-operations: recognizing lanes and bands on the lanes in the original gel image of the electrophoresis image using an artificial intelligence recognition network model; performing quality analysis on the recognized lanes and bands on the lanes to perform quality control on the gel image.

[0220] That is to say, although in the system for identifying and analyzing glue images using artificial intelligence according to the present disclosure, the various operating functions corresponding to the method for identifying and analyzing glue images using artificial intelligence are described as modules or sub-modules, those skilled in the art should understand that such modules or sub-modules may be circuit components or other physical components, or may not be composed of circuit components or other physical components, but may be functional modules constructed by computer programs.

[0221] In addition, it should be recognized by those skilled in the art that the method of the present disclosure can be implemented as a computer program. As described above in conjunction with the accompanying drawings, the method of the above embodiment is executed by one or more programs, and the instructions in the program cause the computer or processor to execute the algorithm described in conjunction with the accompanying drawings. These programs can be stored and provided to the computer or processor using various types of non-transient computer-readable media. Non-transient computer-readable media include various types of tangible storage media. Examples of non-transient computer-readable media include magnetic recording media (such as floppy disks, tapes, and hard disk drives), magneto-optical recording media (such as magneto-optical disks), CD-ROMs (compact disk read-only memories), CD-Rs, CD-R / Ws, and semiconductor memories (such as ROMs, PROMs (programmable ROMs), EPROMs (erasable PROMs), flash ROMs, and RAMs (random access memories)). Further, these programs can be provided to the computer using various types of transient computer-readable media. Examples of transient computer-readable media include electrical signals, optical signals, and electromagnetic waves. Transient computer-readable media can be used to provide programs to the computer through wired communication paths or wireless communication paths such as wires and optical fibers.

[0222] For example, according to one embodiment of the present disclosure, a non-transitory computer-readable storage medium may be provided for storing a computer program, wherein the computer program includes instructions that, when executed by a processor of an electronic device, enable the electronic device to implement the above-described method of identifying and analyzing glue images using artificial intelligence.

[0223] In addition, according to the content disclosed in the present disclosure, a computer system can also be provided, the computer system comprising: a processor; a memory; and a computer program. The computer program is stored in the memory and is configured to be executed by the processor. The computer program includes instructions for implementing the above-mentioned method of using artificial intelligence to identify and analyze glue images.

[0224] On the other hand, according to one embodiment of the present disclosure, a non-transitory computer-readable storage medium may be provided for storing a computer program, wherein the computer program includes instructions, which, when executed by a processor of an electronic device, enable the electronic device to implement the above-mentioned method for constructing an artificial intelligence recognition network model for identifying glue images or the above-mentioned method for constructing an artificial intelligence prediction network model for quality control of glue images.

[0225] In addition, according to the content disclosed in the present disclosure, a computer system can also be provided, the computer system comprising: a processor; a memory; and a computer program. The computer program is stored in the memory and is configured to be executed by the processor. The computer program includes instructions for implementing the above-mentioned method for constructing an artificial intelligence recognition network model for identifying glue images or the above-mentioned method for constructing an artificial intelligence prediction network model for quality control of glue images.

[0226] The implementation methods of the present disclosure are not limited to the above-mentioned embodiments. Without departing from the spirit and scope of the present disclosure, ordinary technicians in this field can make various changes and improvements to the present disclosure in form and details, and these are considered to fall within the protection scope of the present disclosure.

Claims

1. A method for identifying and analyzing glue images using artificial intelligence, characterized in that: The method comprises: An artificial intelligence recognition network model is used to identify the lanes and bands on the lanes in the original gel image of the electrophoresis diagram; The identified lanes and bands within the lanes were analyzed.

2. The method according to claim 1, characterized in that The artificial intelligence recognition network model is constructed through the following steps: Acquire a training set, wherein the training set includes a plurality of training samples, each of the training samples includes an original image of an electrophoresis gel image and an annotation of the image, wherein the annotation of the image includes annotations of lanes in the image and bands on the lanes; Based on the original image of the electrophoresis gel image and the annotation of the image, the initial artificial intelligence recognition network model is trained to obtain an artificial intelligence recognition network model for identifying the lanes and bands on the lanes in the original image of the electrophoresis gel image.

3. The method according to claim 1 or 2, characterized in that: The initial artificial intelligence recognition network model is a convolutional neural network (CNN) model.

4. The method according to claim 3, characterized in that: The initial artificial intelligence recognition network model is a Yolo network model or an SSD network model.

5. The method according to any one of claims 1 to 4, characterized in that The method further comprises: In the case where the gel map includes more than one electrophoresis sample, the plasmid, product and scale lanes of each electrophoresis sample are determined based on the lanes and bands identified by the artificial intelligence recognition network model, and the gel map is cut based on each electrophoresis sample.

6. The method according to any one of claims 1 to 5, characterized in that The method of using an artificial intelligence recognition network model to recognize lanes and bands on lanes in the original gel image of the electrophoresis diagram includes: The glue holes in the original glue image are identified using an artificial intelligence recognition network model; According to the deflection angle of the horizontal position of the identified glue hole, the original glue image is corrected for horizontal deflection; The artificial intelligence recognition network model is used to identify the lanes and bands on the lanes in the corrected gel image.

7. The method according to any one of claims 1 to 6, characterized in that The analysis of the identified swimming lanes and bands on the swimming lanes includes performing at least one of the following operations: Genomic testing; Cutting or degradation detection; Product band size calculation; and Impurity detection.

8. The method according to claim 7, characterized in that The genome testing includes: An artificial intelligence prediction network model was used to preliminarily predict whether there were genomic bands in the gel image; When the artificial intelligence prediction network model preliminarily predicts the presence of a genomic band in the gel image, the content of the genomic band is determined according to the positional characteristics of the genomic band to be determined in the gel image and a quantitative analysis method.

9. The method according to claim 7, characterized in that: The cutting or degradation detection includes: Use artificial intelligence prediction network model to preliminarily predict whether there is a cut or degradation area in the glue map; When the artificial intelligence prediction network model preliminarily predicts that there is a shaving or degradation area in the glue map, the degree of shaving or degradation is further judged according to the upper and lower positions of the preliminarily predicted shaving or degradation area and the percentage of the brightness of the shaving or degradation area to the strip brightness of the identification area.

10. The method according to claim 7, characterized in that The product band size calculation includes: Determining the type of the ruler according to the bands in the ruler lane identified by using the artificial intelligence recognition network model; The size of the band in the product lane is determined according to the relative positional relationship between the band in the product lane and the band in the ruler lane identified by the artificial intelligence recognition network model.

11. The method according to claim 10, characterized in that The determining of the scale type further comprises: The scale type is determined by the number of bands in the scale lane and the distance between the last two bands.

12. The method according to claim 10, characterized in that After determining the type of the ruler and before determining the size of the band in the product lane, the method further comprises: accurately adjusting the position of the band.

13. The method according to claim 10, characterized in that The determining the size of the band in the product lane according to the relative positional relationship between the band in the product lane and the band in the ruler lane identified by the artificial intelligence recognition network model further comprises: According to the determined scale type, the size of each band in the scale lane is determined by referring to the band size corresponding to each band in the standard scale diagram of the corresponding type; The size of the band in the product lane is determined based on the relative position of the band in the product lane and the band in the ruler lane.

14. The method according to claim 13, characterized in that The determining the size of the band in the product lane according to the relative position relationship between the band in the product lane and the band in the ruler lane further comprises: Determine whether the band in the product lane is flush with a band in the ruler lane; When the band in the product lane is aligned with a band in the ruler lane, the size of the band in the product lane is the size of the band in the ruler lane that is aligned with it.

15. The method according to claim 13, characterized in that The determining the size of the band in the product lane according to the relative position relationship between the band in the product lane and the band in the ruler lane further comprises: The size of the band in the product lane is determined by calculating the relative positional relationship between the upper edge, lower edge and center line of the band in the product lane and the band in the ruler lane.

16. The method according to claim 15, characterized in that The step of determining the size of the band in the product lane by calculating the relative positional relationship between the upper edge, the lower edge and the center line of the band in the product lane and the band in the ruler lane further comprises: The median value of the size of the band in the product lane is determined by calculating the relative positional relationship between the upper edge and the lower edge of the band in the product lane and the band in the ruler lane; By calculating the relative positional relationship between the center line of the band in the product lane and the band in the ruler lane, the center line value of the size of the band in the product lane is determined; The intermediate value is compared with the center line value, and the value closest to the theoretical value of the band in the product lane is taken as the size of the band in the product lane.

17. The method according to claim 7, characterized in that The impurity detection includes: Use artificial intelligence prediction network model to preliminarily predict whether there are impurities in the glue map; When the artificial intelligence prediction network model preliminarily predicts that there are impurities in the gel image, it is determined whether the preliminarily predicted impurity position, size and brightness affect the determination of the product band.

18. The method according to any one of claims 7 to 17, characterized in that: The analysis of the identified swimming lanes and bands on the swimming lanes further includes performing at least one of the following operations: Scale integrity test; Incomplete detection of product digestion; Plasmid band detection; and Detection of stray or dropped tapes.

19. The method according to any one of claims 2 to 18, characterized in that The marking of the picture further includes marking the glue holes in the picture.

20. The method according to claim 8, characterized in that The artificial intelligence prediction network model is constructed through the following steps: Obtaining a training set, wherein the training set includes a plurality of training samples, each of the training samples includes an original image of an electrophoresis gel image and an annotation of the image, wherein the annotation of the image includes annotations of genome bands and their locations; Based on the original images of the electrophoresis gel images and the annotations of the images, the initial artificial intelligence prediction network model was trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether there are genomic bands in the product lanes.

21. The method according to claim 9, characterized in that The artificial intelligence prediction network model is constructed through the following steps: Obtaining a training set, wherein the training set includes a plurality of training samples, each of the training samples includes an original image of an electrophoresis gel image and an annotation of the image, wherein the annotation of the image includes an annotation of a cut or degraded area and its location; Based on the original images of the electrophoresis gel images and the annotations of the images, the initial artificial intelligence prediction network model is trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether there are shaving or degradation areas in the gel images.

22. The method according to claim 17, characterized in that The artificial intelligence prediction network model is constructed through the following steps: Obtaining a training set, wherein the training set includes a plurality of training samples, each of the training samples includes an original image of an electrophoresis gel image and an annotation of the image, wherein the annotation of the image includes annotations of impurities and their locations; Based on the original images of the electrophoresis gel images and the annotations of the images, the initial artificial intelligence prediction network model is trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether there are impurities in the gel images.

23. The method according to any one of claims 20 to 22, characterized in that The initial artificial intelligence prediction network model is a convolutional neural network model.

24. The method according to claim 23, characterized in that The initial artificial intelligence prediction network model is a Yolo network model or an SSD network model.

25. A method for constructing an artificial intelligence recognition network model for identifying glue images, characterized in that: The method comprises: Acquire a training set, wherein the training set includes a plurality of training samples, each of the training samples includes an original image of an electrophoresis gel image and an annotation of the image, wherein the annotation of the image includes annotations of lanes in the image and bands on the lanes; Based on the original image of the electrophoresis gel image and the annotation of the image, the initial artificial intelligence recognition network model is trained to obtain an artificial intelligence recognition network model for identifying the lanes and bands on the lanes in the original image of the electrophoresis gel image.

26. The method according to claim 25, characterized in that The marking of the picture further includes marking the glue holes in the picture.

27. The method according to claim 24 or 25, characterized in that The initial artificial intelligence recognition network model is a convolutional neural network (CNN) model.

28. The method according to claim 27, characterized in that The initial artificial intelligence recognition network model is a Yolo network model or an SSD network model.

29. An artificial intelligence recognition network model for identifying glue images obtained according to the method described in any one of claims 25-28.

30. A method for constructing an artificial intelligence prediction network model for analyzing adhesive images, characterized in that: The method comprises: Obtaining a training set, wherein the training set includes a plurality of training samples, each of the training samples includes an original image of an electrophoresis gel image and annotations of the image, wherein the annotations of the image include annotations of items with gel image quality problems and their locations; Based on the original images of the electrophoresis gel images and the annotations of the images, an initial artificial intelligence prediction network model is trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether there are quality problems in the gel images.

31. The method according to claim 30, characterized in that The labeling of the image includes labeling of the genome and its location; and Based on the original images of the electrophoresis gel images and the annotations of the images, the initial artificial intelligence prediction network model is trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether a genome exists in the product lane.

32. The method according to claim 30, characterized in that The marking of the image includes marking the cut or degraded area and its location; and Based on the original images of the electrophoresis gel images and the annotations of the images, the initial artificial intelligence prediction network model is trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether there are shaving or degradation areas in the gel images.

33. The method according to claim 30, characterized in that The marking of the image includes marking of impurities and their locations; and Based on the original images of the electrophoresis gel images and the annotations of the images, the initial artificial intelligence prediction network model is trained to obtain an artificial intelligence prediction network model for preliminarily predicting whether there are impurities in the gel images.

34. The method according to any one of claims 30 to 33, characterized in that The initial artificial intelligence prediction network model is a convolutional neural network (CNN) model.

35. The method according to claim 34, characterized in that The initial artificial intelligence prediction network model is a Yolo network model or an SSD network model.

36. An artificial intelligence prediction network model for analyzing gel images obtained according to the method described in any one of claims 30-35.

37. A system for identifying and analyzing glue images using artificial intelligence, characterized in that: The system comprises: An acquisition module is used to obtain the original gel image of the electrophoresis image; The processing module is used to identify and analyze the glue map using artificial intelligence, and the processing module further includes the following submodules: The recognition submodule is used to recognize the lanes and bands on the lanes in the original gel image of the electrophoresis image by using an artificial intelligence recognition network model; The analysis submodule is used to analyze the lanes and bands on the lanes identified by the identification submodule.

38. A computer-implemented system for identifying and analyzing glue images using artificial intelligence, wherein the computer-implemented system implements the following operations by executing a computer program: Get the original gel image of the electrophoresis diagram; Using artificial intelligence to identify and analyze the glue map, this operation further includes the following sub-operations: An artificial intelligence recognition network model is used to identify the lanes and bands on the lanes in the original gel image of the electrophoresis diagram; The identified lanes and bands within the lanes were analyzed.

39. A non-transitory computer-readable storage medium for storing a computer program, wherein the computer program includes instructions, which, when executed by a processor of an electronic device, causes the electronic device to implement the method for identifying and analyzing glue images using artificial intelligence as described in any one of claims 1-24, or the method for constructing an artificial intelligence recognition network model for identifying glue images as described in any one of claims 25-28, or the method for constructing an artificial intelligence prediction network model for analyzing glue images as described in any one of claims 30-35.

40. A computer system, comprising: processor; Memory; and A computer program, wherein the computer program is stored in the memory and is configured to be executed by the processor, the computer program comprising instructions for implementing the method for identifying and analyzing glue images using artificial intelligence as described in any one of claims 1-24, or the method for constructing an artificial intelligence recognition network model for identifying glue images as described in any one of claims 25-28, or the method for constructing an artificial intelligence prediction network model for analyzing glue images as described in any one of claims 30-35.

Images