Method and system for rapidly detecting negative endpoint titer
By automatically interpreting the results of Treponema pallidum antigen serological tests using a deep learning model, the subjectivity and inefficiency of manual visual interpretation are solved, enabling rapid and accurate detection of the negative endpoint titer for syphilis. This method is suitable for syphilis diagnosis and quality control.
Patent Information
- Application Number
- CN202511445025.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-10
- Publication Date
- 2026-02-06
AI Technical Summary
The interpretation of existing Treponema pallidum antigen serological test results relies on manual visual inspection, which has problems such as strong subjectivity, poor consistency, environmental sensitivity, and low efficiency, resulting in poor repeatability and accuracy of test results.
By employing a deep learning-based reaction intensity classification model, the dilution factor is automatically determined through image acquisition and analysis, skipping unnecessary dilution steps, and enabling rapid detection of syphilis negative endpoint titers.
It improves the accuracy and efficiency of testing, reduces reagent and time costs, is suitable for large-scale sample testing, and reduces the risk of misjudgment.
Smart Images
Figure CN121476170A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of syphilis detection technology, and in particular to a method and system for rapidly detecting negative endpoint titers. Background Technology
[0002] Syphilis is a chronic, systemic sexually transmitted disease caused by infection with Treponema pallidum. Timely and accurate diagnosis is crucial for disease control and effective treatment. Serological testing is currently the primary means of diagnosing syphilis and evaluating treatment efficacy in clinical practice.
[0003] Serological testing methods for syphilis generally fall into two main categories: non-treponemal antigen serological tests and treponemal antigen serological tests. Among these, non-treponemal antigen serological tests, such as the toluidine red unheated serum test (TRUST) or the rapid plasma reagin (RPR) ring card test, are widely used clinically due to their simplicity, low cost, and suitability for large-scale screening. These tests primarily detect the presence of antibodies against lipid antigens, known as reagin, in the patient's serum.
[0004] Taking the RPR test as an example, its basic principle is to mix cardiolipin antigen suspended in activated charcoal particles with the serum sample to be tested on a specific reaction card. If sufficient reagin is present in the serum, the antigen and antibody will undergo a flocculation reaction, forming visible black agglomerated particle clumps under the guidance of activated charcoal particles; this is a positive reaction. If there is no reagin in the serum or its content is extremely low, the carbon particles in the mixture remain in a uniform free suspension; this is a negative reaction. By serially diluting the serum, the highest dilution at which a positive reaction occurs can be determined, which is the antibody negative endpoint titer of the sample, used to assess disease activity and treatment efficacy.
[0005] In current technical practices, the interpretation of RPR test results relies entirely on the visual observation of the testing technician. Operators need to visually observe the degree of carbon particle aggregation within the reaction wells under specific lighting conditions to determine whether the result is "negative," "weakly positive," or "positive." However, this method, which relies entirely on manual visual interpretation, has several inherent drawbacks: 1. High subjectivity and poor consistency: The interpretation of results heavily depends on the personal experience and subjective judgment of the testing technician. For "weakly positive" samples, which fall between negative and positive, different technicians may reach different conclusions. Even the same technician's interpretation of the same sample at different times (such as under fatigue) may produce deviations, leading to poor repeatability and consistency of test results. 2. Sensitive to environmental conditions: The effectiveness of visual interpretation is easily affected by external environmental factors, such as the intensity of the light source, color temperature, and observation angle. Unstable environmental conditions directly affect the observation results, leading to misjudgments or missed judgments. 3. Low efficiency and high labor intensity: For the testing of large batches of samples, manual interpretation of each sample is time-consuming and labor-intensive, limiting the testing throughput and becoming a bottleneck in the laboratory workflow. Prolonged visual observation can also easily cause visual fatigue for testing personnel, further increasing the risk of misinterpretation.
[0006] Therefore, how to overcome the problems of subjectivity, inefficiency and low standardization of traditional RPR test manual interpretation methods, and provide a more objective, efficient and more accurate and reliable result interpretation technology, is a technical problem that urgently needs to be solved in this field. Summary of the Invention
[0007] To address the problems existing in the prior art, the present invention provides a method and system for rapidly detecting negative endpoint titers.
[0008] In a first aspect, the present invention provides a method for detecting a negative endpoint titer, comprising: (1) Obtain the reaction result image of the non-treponemal antigen serum test; (2) Input the reaction result image into the reaction intensity classification model and output the reaction intensity classification; determine the dilution factor for the next step of detection based on the reaction intensity classification and skip unnecessary dilution factors; (3) Prepare a new sample to be tested according to the dilution factor determined in (2), and repeat steps (1) and (2) until the output result of the reaction intensity classification model is negative, and obtain the negative endpoint titer; The reaction intensity classification model is constructed based on the reaction result images of non-treponemal antigen serum test results of syphilis samples at different titers and the corresponding reaction intensity classification.
[0009] This invention proposes a rapid method for detecting the negative endpoint titer of syphilis based on the reaction result image of the test sample and a reaction intensity classification model. The dilution factor is determined through the reaction intensity classification model, and the dilution step is adjusted, which can eliminate some testing steps and save reagents, consumables, and time costs. Furthermore, it also achieves high accuracy (low false positive and false negative rates).
[0010] Furthermore, the rapid syphilis titer detection method provided by this invention is based on non-disease treatment and diagnostic purposes. Besides detecting syphilis titers in patient serum, it can also be used for quality control of standards for different syphilis titers (e.g., commercially available syphilis titer standards), verifying their effectiveness. It can also be used in the research, development, and quality control of syphilis diagnostic kits.
[0011] It can also be used for the detection of lipid antigens in in vitro culture processes, such as when donor-derived cells (e.g., umbilical cord blood, bone marrow, peripheral blood mononuclear cells) are used in the production of cell therapy products (e.g., CAR-T, stem cell therapy) and lipid antigens may be present.
[0012] Furthermore, the reaction intensity classification model is trained using a training set, and the method for constructing the training set includes: The reaction result images of the non-treponemal antigen serological test results for each training sample at each dilution gradient in the complete serial dilution series were collected and grouped; the complete serial dilution series includes dilutions from undiluted samples to the negative endpoint; The reaction result images corresponding to the first dilution gradient of the negative endpoint are classified into the weak positive group; The reaction result images corresponding to the first two dilution gradients of the negative endpoint were classified into the medium-positive group; The reaction result images corresponding to ≥3 dilution gradients of the negative endpoint are classified into the strong positive group; The negative endpoints are classified into negative groups; Preferably, the dilution gradient in the complete serial dilution series is ≥2.
[0013] More preferably, the dilution gradient in the complete serial dilution series is an integer of ≥2.
[0014] More preferably, the dilution gradient in the complete serial dilution series is 2 to 10 times; for example, gradients of 2, 3, 4, 5, 6, 7, etc.
[0015] Furthermore, the negative endpoint is determined based on the standard WS / T491-2024.
[0016] The negative endpoint described in this invention can be determined by experts in the field, for example, based on the standard WS / T491-2024: the carbon particles in the reaction mixture always remain in a uniform free suspension state.
[0017] Furthermore, the method for constructing the training set also includes: Remove reaction result images that exhibit prozone phenomena.
[0018] Furthermore, the reaction intensity classification model is constructed using the following method: (1) Input the image data of the training set into the initial deep learning classification model, and extract image features and output the predicted classification through forward propagation of the convolutional neural network; (2) Calculate the error value based on the predicted classification and the real label, and use error backpropagation to calculate the gradient of the weight parameters in the model; (3) The model weight parameters are updated using the stochastic gradient descent algorithm, and iterative optimization is performed until the loss function meets the preset convergence condition.
[0019] Furthermore, the loss function is one or more of the cross-entropy loss function or the focus loss function; Preferably, the loss function is the cross-entropy loss function.
[0020] Furthermore, the method for determining the dilution factor used in the next step of detection in (2) includes: If the reaction intensity is classified as weakly positive, then the dilution factor is determined to be a serial dilution step of 0.5 to 1.5 times; If the reaction intensity is classified as moderately positive, then the dilution factor is determined to be a serial dilution step of 1.5 to 2.5 times; If the reaction intensity is classified as strongly positive, then the dilution factor is determined to be a serial dilution step of 2.5 to 3.5 times.
[0021] Furthermore, the non-treponemal antigen serum test is either the toluidine red unheated serum test or the rapid plasma reagin ring card test.
[0022] In a second aspect, the present invention provides a system for detecting syphilis titers, comprising: The image acquisition module is used to acquire reaction result images of non-treponemal antigen serum tests; The sample classification module is used to input the reaction result image into the reaction intensity classification model and output the reaction intensity classification; based on the reaction intensity classification, it determines the dilution factor for the next step of detection and skips unnecessary dilution factors. The dilution module is used to prepare new test samples according to the dilution factor output by the sample classification module, and re-enter the image acquisition module to repeat the process until the sample classification module outputs a negative result. Preferably, the method for determining the dilution factor for the next detection step includes: If the reaction intensity is classified as weakly positive, then the dilution factor is determined to be a serial dilution step of 0.5 to 1.5 times; If the reaction intensity is classified as moderately positive, then the dilution factor is determined to be a serial dilution step of 1.5 to 2.5 times; If the reaction intensity is classified as strongly positive, then the dilution factor is determined to be a serial dilution step of 2.5 to 3.5 times.
[0023] Furthermore, it also includes: The model building module is used to construct the reaction intensity classification model based on the reaction result images and corresponding reaction intensity classifications of syphilis samples with non-treponemal antigen serum test results at different titers. Preferably, the reaction intensity classification model is trained using a training set, and the method for constructing the training set includes: The reaction result images of the non-treponemal antigen serological test results for each training sample at each dilution gradient in the complete serial dilution series were collected and grouped; the complete serial dilution series includes dilutions from undiluted samples to the negative endpoint; The reaction result images corresponding to the first dilution gradient of the negative endpoint are classified into the weak positive group; The reaction result images corresponding to the first two dilution gradients of the negative endpoint were classified into the medium-positive group; The reaction result images corresponding to ≥3 dilution gradients of the negative endpoint are classified into the strong positive group; The negative endpoints are classified into negative groups; More preferably, the method for constructing the training set further includes: Remove reaction result images that exhibit prozone phenomena; More preferably, the reaction intensity classification model is constructed using the following method: (1) Input the image data of the training set into the initial deep learning classification model, and extract image features and output the predicted classification through forward propagation of the convolutional neural network; (2) Calculate the error value based on the predicted classification and the real label, and use error backpropagation to calculate the gradient of the weight parameters in the model; (3) The model weight parameters are updated using the stochastic gradient descent algorithm, and iterative optimization is performed until the loss function meets the preset convergence condition; More preferably, the loss function is one or more of the cross-entropy loss function or the focus loss function; More preferably, the loss function is the cross-entropy loss function.
[0024] The present invention has the following beneficial effects: This invention provides a method for detecting syphilis negative endpoint titers based on reaction result images of non-treponemal antigen serological tests and a deep learning model. The method determines the dilution factor for the next detection step through a reaction intensity classification model, adaptively skipping unnecessary dilution factors. This improves detection efficiency while maintaining high accuracy and saving test reagents, which has significant application value in the field of syphilis negative endpoint titer detection. Attached Figure Description
[0025] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0026] Figure 1 These are images of RPR reaction cards at different titers provided in Example 1 of this invention.
[0027] Figure 2 This is a schematic diagram of the method for rapid detection of syphilis negative endpoint titer provided in Embodiment 1 of the present invention. Detailed Implementation
[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of this invention, not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0029] Unless otherwise specified, the experimental methods involved in the following embodiments are conventional methods in the art. For example, you can refer to the experimental manual in the art or follow the conditions recommended in the manufacturer's instructions.
[0030] Unless otherwise specified, all experimental materials and reagents used in the following examples are commercially available.
[0031] Example 1 In this embodiment, a method for rapid detection of syphilis titers includes: I. Construction of the reaction intensity classification module.
[0032] 1. Constructing the training set (1) Collect all gradient (preferably 2 times) reaction image data of each sample from undiluted (1:1) to the clinically determined negative endpoint titer. The negative endpoint titer is professionally determined by a physician with more than 3 years of clinical experience, for example, according to the standard WS / T491-2024: the carbon particles in the reaction mixture always remain in a uniform free suspension state.
[0033] (2) Data classification rules: i) Negative group: Includes reaction images corresponding to clinically determined negative titers (e.g., ... Figure 1 Middle 2 5 (titration image).
[0034] ii) Weak positive group: Response images containing the dilution gradient preceding the negative endpoint (e.g.) Figure 1 Middle 2 4 (titration image).
[0035] iii) Positive group: Response images including the two dilution gradients prior to the negative endpoint (e.g.) Figure 1 Middle 2 3 (titration image).
[0036] iv) Strongly positive group: Response images containing the first three or more dilution gradients before the negative endpoint (e.g.) Figure 1 Middle 2 2 times, 2 1 (titration image).
[0037] (3) Data quality control: To address the potential "prozone phenomenon" (i.e., the weakened response due to antibody overdose, as shown in the examples) in high antibody concentration samples... Figure 1 2 shown 0 (Images with multiple titers) are excluded from the training set during the training phase to prevent the model from learning incorrect feature associations.
[0038] During the application phase, when such samples are input into the model, due to their weak response, they may output moderate / weak positive results. The system will automatically perform gradient dilution and test the diluted samples. Since this invention uses the negative endpoint as the end point of the dilution-detection process, samples exhibiting the "pre-band phenomenon" will still correctly return to the negative endpoint after multiple consecutive dilutions (the first dilution normally shows strong / moderate positive results, and subsequent dilution-detection processes will proceed normally), without affecting the accuracy of the judgment.
[0039] 2. Model training.
[0040] (1) Input the image data of the aforementioned training set into the initial deep learning classification model, and extract image features and output the predicted classification (strong positive, medium positive, weak positive and negative) through forward propagation of the convolutional neural network.
[0041] (2) Calculate the error value based on the predicted classification and the real label, and use error backpropagation to calculate the gradient of the weight parameters in the model.
[0042] The true label is the true classification of the image data in the training set.
[0043] (3) The model weight parameters are updated by using the stochastic gradient descent algorithm, and the loss function is iterated until it meets the preset convergence condition to obtain the reaction intensity classification model; the loss function is the cross-entropy loss function.
[0044] II. Detection of syphilis titer.
[0045] 1. At the undiluted titer, take a picture of the reaction card after the RPR test of the sample to be tested, obtain the image, and input the image into the aforementioned reaction intensity classification module to obtain the reaction intensity classification at the current titer; 2. Based on the current reaction intensity classification results, automatically calculate the next dilution factor to be tested, and dilute the plasma sample accordingly according to the calculation results to prepare plasma reagents of the specified dilution for the next round of testing.
[0046] The calculation method is as follows: If the reaction intensity is classified as weakly positive, then the serial dilution step of 1x is determined (the above refers to a 2x dilution, so a serial dilution step of 1x is a 1:2 dilution).
[0047] If the reaction intensity is classified as moderately positive, then the dilution factor is determined to be a 2-fold serial dilution step (corresponding to 1:2). 2 dilution).
[0048] If the reaction intensity is classified as strongly positive, then the dilution factor is determined to be a 3-fold serial dilution step (corresponding to 1:2). 3 dilution).
[0049] 3. Perform RPR testing on the diluted sample and acquire reaction images. Input the images back into the reaction intensity classification module to obtain the reaction intensity rating of the current titer. Repeat step 2 until the reaction intensity classification module outputs a negative reaction intensity rating and generates a test report.
[0050] Figure 2 This is a schematic diagram of a method for detecting syphilis titers.
[0051] Example 2 In this embodiment, a method for rapid detection of syphilis titers includes: I. Construction of the reaction intensity classification module.
[0052] 1. Construct the training set.
[0053] (1) Sample source: Blood samples from 376 outpatients or inpatients who actively underwent syphilis nonspecific antibody testing.
[0054] (2) After standardizing the operation by using the instruction manual, a high-definition CCD microscope, model GP-660V, with an eyepiece of 0.35×, was used. The appropriate height and focal length were adjusted so that the circle of the test card could be displayed within the microscope's shooting range. The position was fixed, the supplementary light angle was fixed, and all images were taken to achieve standardization.
[0055] RPR response image data for each sample is obtained for all gradients (preferably 2x in this embodiment) from undiluted (1:1) to the clinically determined negative endpoint titer.
[0056] The negative endpoint titer is determined by a physician with more than 3 years of clinical experience, for example, according to the standard WS / T491-2024: the carbon particles in the reaction mixture always remain in a uniform free suspension state.
[0057] The RPR reaction procedure was performed in accordance with the instructions for the Rapid Plasma Reagin Diagnostic Reagent for Syphilis and the health industry standard WS / T491-2024 Guidelines for the Detection of Non-specific Antibodies for Syphilis.
[0058] (3) Data classification rules.
[0059] i) Negative group: Reaction images corresponding to the aforementioned negative endpoint (dilution factor 2) n ).
[0060] ii) Weak positive group: includes the dilution gradient preceding the negative endpoint (2) n-1 (The reaction image)
[0061] iii) Positive group: includes the two dilution gradients before the negative endpoint (2 n-2 (The reaction image)
[0062] iv) Strongly positive group: including the top three or more dilution gradients of the negative endpoint (2 n-3 The reaction images (and above).
[0063] (3) Data quality control.
[0064] Remove RPR response images exhibiting "pre-band phenomenon" from each training sample.
[0065] 2. Model training.
[0066] (1) Input the image data of the aforementioned training set into the initial deep learning classification model, and extract image features and output the predicted classification (strong positive, medium positive, weak positive and negative) through forward propagation of the convolutional neural network.
[0067] (2) Calculate the error value based on the predicted classification and the real label, and use error backpropagation to calculate the gradient of the weight parameters in the model.
[0068] The true label is the true classification of the image data in the training set.
[0069] (3) The model weight parameters are updated using the stochastic gradient descent algorithm and iteratively optimized until the loss function meets the preset convergence condition; the loss function is the cross-entropy loss function.
[0070] The model uses ResNet50 and the SGD optimizer, employing a multi-step learning rate decay strategy. The initial learning rate is 0.01, and the learning rate decays by 0.1 every 30 epochs, for a total of 60 epochs, with a batch size of 128.
[0071] Experimental Example 1 Based on the method provided in Example 2, the present invention is validated in the following process: 1. Sample collection.
[0072] The study used blood samples from 250 outpatients or inpatients who voluntarily underwent syphilis nonspecific antibody testing.
[0073] 2. Comparative example.
[0074] Comparative Example 1: The detection method of Example 1 was modified by removing the process of "dynamically adjusting the dilution step according to the current reaction intensity" and replacing it with "using a fixed dilution step to gradually obtain the next drop".
[0075] 3. Test results.
[0076] (1) Consistency The consistency between the detection results of this invention and the results of visual inspection by professional physicians showed that the consistency in predicting reaction intensity was 88.02%.
[0077] In medicine, a negative endpoint titer within 2 times (one dilution gradient) is considered clinically acceptable, meaning the detected negative endpoint titer differs from the true endpoint titer by one dilution gradient. The clinical consistency of the negative endpoint titer in this invention is 98.13%.
[0078] (2) Discrimination efficiency Using the comparative example, 1250 RPR reaction tests were required to obtain the negative endpoint titer for all samples. However, the rapid detection method of this invention, by dynamically adjusting the dilution step, skips some RPR tests at certain titers, ultimately requiring only 790 experiments to obtain the negative endpoint titer for all samples. The number of RPR tests required by this invention is only 63.2% of that in the comparative example, significantly saving testing time and reagent usage.
[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for detecting a negative endpoint titer, characterized in that, include: (1) Obtain the reaction result image of the non-treponemal antigen serum test; (2) Input the reaction result image into the reaction intensity classification model and output the reaction intensity classification; determine the dilution factor for the next step of detection based on the reaction intensity classification and skip unnecessary dilution factors; (3) Prepare a new sample to be tested according to the dilution factor determined in (2), and repeat steps (1) and (2) until the output result of the reaction intensity classification model is negative, and obtain the negative endpoint titer; The reaction intensity classification model is constructed based on the reaction result images of non-treponemal antigen serum test results of syphilis samples at different titers and the corresponding reaction intensity classification.
2. The method according to claim 1, characterized in that, The reaction intensity classification model is trained using a training set, and the method for constructing the training set includes: The reaction result images of the non-treponemal antigen serological test results for each training sample at each dilution gradient in the complete serial dilution series were collected and grouped; the complete serial dilution series includes dilutions from undiluted samples to the negative endpoint; The reaction result images corresponding to the first dilution gradient of the negative endpoint are classified into the weak positive group; The reaction result images corresponding to the first two dilution gradients of the negative endpoint were classified into the medium-positive group; The reaction result images corresponding to ≥3 dilution gradients of the negative endpoint are classified into the strong positive group; The negative endpoints are classified into negative groups; Preferably, the dilution gradient in the complete serial dilution series is ≥2.
3. The method according to claim 2, characterized in that, The negative endpoint was determined based on the standard WS / T491-2024.
4. The method according to claim 2 or 3, characterized in that, The method for constructing the training set also includes: Remove reaction result images that exhibit prozone phenomena.
5. The method according to any one of claims 2-4, characterized in that, The reaction intensity classification model was constructed using the following method: (1) Input the image data of the training set into the initial deep learning classification model, and extract image features and output the predicted classification through forward propagation of the convolutional neural network; (2) Calculate the error value based on the predicted classification and the real label, and use error backpropagation to calculate the gradient of the weight parameters in the model; (3) The model weight parameters are updated using the stochastic gradient descent algorithm, and iterative optimization is performed until the loss function meets the preset convergence condition.
6. The method according to claim 5, characterized in that, The loss function is one or more of the cross-entropy loss function or the focus loss function; Preferably, the loss function is the cross-entropy loss function.
7. The method according to any one of claims 2-6, characterized in that, (2) The method for determining the dilution factor used in the next step of detection includes: If the reaction intensity is classified as weakly positive, then the dilution factor is determined to be a serial dilution step of 0.5 to 1.5 times; If the reaction intensity is classified as moderately positive, then the dilution factor is determined to be a serial dilution step of 1.5 to 2.5 times; If the reaction intensity is classified as strongly positive, then the dilution factor is determined to be a serial dilution step of 2.5 to 3.5 times.
8. The method according to any one of claims 1-7, characterized in that, The non-treponemal antigen serum test is either the toluidine red unheated serum test or the rapid plasma reagin ring card test.
9. A system for detecting syphilis titers, characterized in that, include: The image acquisition module is used to acquire reaction result images of non-treponemal antigen serum tests; The sample classification module is used to input the reaction result image into the reaction intensity classification model and output the reaction intensity classification; based on the reaction intensity classification, it determines the dilution factor for the next step of detection and skips unnecessary dilution factors. The dilution module is used to prepare new test samples according to the dilution factor output by the sample classification module, and re-enter the image acquisition module to repeat the process until the sample classification module outputs a negative result. Preferably, the method for determining the dilution factor for the next detection step includes: If the reaction intensity is classified as weakly positive, then the dilution factor is determined to be a serial dilution step of 0.5 to 1.5 times; If the reaction intensity is classified as moderately positive, then the dilution factor is determined to be a serial dilution step of 1.5 to 2.5 times; If the reaction intensity is classified as strongly positive, then the dilution factor is determined to be a serial dilution step of 2.5 to 3.5 times.
10. The system according to claim 9, characterized in that, Also includes: The model building module is used to construct the reaction intensity classification model based on the reaction result images and corresponding reaction intensity classifications of syphilis samples with non-treponemal antigen serum test results at different titers. Preferably, the reaction intensity classification model is trained using a training set, and the method for constructing the training set includes: The reaction result images of the non-treponemal antigen serological test results for each training sample at each dilution gradient in the complete serial dilution series were collected and grouped; the complete serial dilution series includes dilutions from undiluted samples to the negative endpoint; The reaction result images corresponding to the first dilution gradient of the negative endpoint are classified into the weak positive group; The reaction result images corresponding to the first two dilution gradients of the negative endpoint were classified into the medium-positive group; The reaction result images corresponding to ≥3 dilution gradients of the negative endpoint are classified into the strong positive group; The negative endpoints are classified into negative groups; More preferably, the method for constructing the training set further includes: Remove reaction result images that exhibit prozone phenomena; More preferably, the reaction intensity classification model is constructed using the following method: (1) Input the image data of the training set into the initial deep learning classification model, and extract image features and output the predicted classification through forward propagation of the convolutional neural network; (2) Calculate the error value based on the predicted classification and the real label, and use error backpropagation to calculate the gradient of the weight parameters in the model; (3) The model weight parameters are updated using the stochastic gradient descent algorithm, and iterative optimization is performed until the loss function meets the preset convergence condition; More preferably, the loss function is one or more of the cross-entropy loss function or the focus loss function; More preferably, the loss function is the cross-entropy loss function.