Model construction method for improving syphilis serological detection reliability and accuracy under assistance of AI

Abstract

The invention focuses on the field of biomedical detection, and aims to enhance the detection reliability and accuracy of the sensor by using artificial intelligence. Aiming at the problems that a traditional syphilis antibody agglutination test is easily interfered by subjective factors and complex sample signals are difficult to identify, a two-step convolutional neural network model is designed. The model comprises an object search algorithm and a classification algorithm, and the object search algorithm reflects a large amount of paper jam pictures through deep learning and accurately cuts agglutination and precipitation areas; the latter learns feature differences to accurately judge the agglutination intensity based on a known agglutination intensity grade sample image. A large number of experiments show that the method effectively eliminates interference, the accuracy and consistency of detection results are greatly improved, the error rate is far lower than that of a traditional method, the detection level is comparable to ideal interpretation of human beings, and the performance is better in complex sample detection. Meanwhile, the model adaptability is high, various detection items and reaction paper jam types can be rapidly adapted through different sample data training, and the result is accurately output.

Description

Technical Field

[0001] The present invention relates to the field of biomedical testing, and in particular to a method for improving the reliability and accuracy of syphilis serological testing with the aid of artificial intelligence (AI) and related model construction. Background Art

[0002] Laboratory diagnostic methods for syphilis primarily fall into two categories: etiological testing and serological testing. Serological testing is the most commonly used and important clinical method. Syphilis serological testing is a general term for a class of tests that detect antibodies to Treponema pallidum in a patient's blood. These antibodies can be divided into two categories: reagins and Treponema pallidum-specific antibodies.

[0003] The non-treponemal test is a flocculent agglutination test using cardiolipin, phosphatidylcholine, and cholesterol as antigens. Reagin forms an antigen-antibody reaction with cardiolipin. Shaking and collision cause particles to adhere to each other, forming visible agglutination and precipitation, a positive reaction. The TPPA test is a flocculent agglutination test for syphilis-specific antibodies. Refined syphilis bacterial components are coated onto artificial carrier gelatin particles. These sensitized particles react with Treponema pallidum antibodies in the sample, causing agglutination. This reaction allows the detection of Treponema pallidum antibodies in serum and plasma and can be used to determine antibody titers. Images of positive results show varying agglutinated particles, such as large aggregates and small clusters, and are unevenly distributed. In contrast, images of negative results show virtually no visible agglutinated particles, presenting a more uniform background. However, in practice, due to factors such as sample complexity, experimental procedures, and visual interpretation, image features can vary, posing a challenge for accurate identification.

[0004] Artificial intelligence (AI)-assisted approaches for biomedical applications can improve the reliability and accuracy of sensors, bringing them on par with human performance. The authors designed a two-step convolutional neural network model (object discovery and classification) that efficiently and accurately delivers outputs. The model includes an object-finding algorithm for cropping agglutination precipitates from the entire reaction paper, and another algorithm for classifying agglutination based on intensity. Summary of the Invention

[0005] The core purpose of the present invention is to provide an artificial intelligence-assisted method for the reaction results of syphilis agglutination test, aiming to improve the reliability and accuracy of result judgment.

[0006] Two-step convolutional neural network model construction: This paper designs a unique two-step convolutional neural network model, which consists of two key parts: object discovery and classification (YOLOv3 for object detection and ResNet-18 for classification).

[0007] Object search algorithm: During the entire reaction card detection process, an object search algorithm is used to crop the agglutination precipitation area from the entire reaction card. In actual biomedical testing, there may be a variety of interference factors on the reaction card, and this algorithm can accurately locate and crop the agglutination precipitation area. Its working principle is to conduct deep learning training on a large number of reaction card images containing agglutination precipitation areas. The model learns the unique image features of the agglutination precipitation area, such as color, texture, shape and other feature combinations. When a new reaction card image is input, the model can quickly and accurately identify the boundaries of the agglutination precipitation area based on the learned features and crop it out from the entire image, providing a clear and accurate target area for subsequent analysis.

[0008] Classification algorithm: Another algorithm is used to classify based on agglutination intensity. This classification algorithm begins after obtaining the cropped image of the agglutination precipitation area. It is also based on the principles of deep learning. We pre-collect a large number of standard sample agglutination reaction photos of known concentrations as the sample raw data set. After identifying clinical samples, we train the model with additional clinical data. The model can extract the differences in image features corresponding to different agglutination intensities. The model learns from the images. For example, an image with high agglutination intensity may show characteristics such as densely clustered agglutinated particles and darker colors, while an image with low agglutination intensity may show characteristics such as dispersed agglutinated particles and lighter colors. Based on these learned features, the model analyzes the input cropped image to accurately determine the level of agglutination intensity and output accurate test results.

[0009] Figures and Description

[0010] Figure 1 Schematic diagram of two-step convolutional neural network model analysis;

[0011] Figure 2 Schematic diagram of the search algorithm flow;

[0012] Figure 3 Database diagram; DETAILED DESCRIPTION

[0013] Data Collection and Preprocessing: We collected a large number of images of different types of syphilis specimens, including images of various agglutination intensity levels and those containing various interference factors. We preprocessed these images, adjusting parameters such as brightness and contrast to achieve uniform standards. We also annotated the images, accurately marking the location of the agglutination precipitation area and the corresponding agglutination intensity level in each image.

[0014] Model training: The preprocessed image dataset is divided into a training set, a validation set, and a test set. The training set is used to train the two-step convolutional neural network model. When training the object search algorithm, a large number of training images are used to allow the model to learn the characteristics of the agglomeration precipitation area. The model parameters are continuously adjusted to enable the model to accurately crop out the agglomeration precipitation area. When training the classification algorithm, cropped images with agglomeration intensity levels are used to allow the model to learn the characteristics corresponding to different intensity levels. The model parameters are continuously optimized through the backpropagation algorithm to improve the accuracy of the model classification. During the training process, the validation set is used to evaluate the performance of the model. Based on the evaluation results, the training parameters, such as the learning rate and number of iterations, are adjusted to prevent the model from overfitting. When the model performance on the validation set reaches the optimal level, the test set is used to conduct a final performance test on the model to ensure that the model can accurately output detection results.

[0015] Actual Detection Application: During the actual detection process, after the sensor acquires an image of a reaction paper jam, the image is first fed into the trained object detection algorithm. The model quickly crops the image of the agglutination precipitation area. This cropped image is then fed into the classification algorithm. The model classifies the agglutination intensity based on the learned features and ultimately outputs an accurate test result, completing the entire agglutination reaction detection process.

Claims

1. A method for constructing a model to improve the reliability and accuracy of syphilis serological detection with AI assistance, characterized in that: include: (1) a data acquisition module, which is used to obtain reaction jam image data from sensors in biomedical testing; (2) A two-step convolutional neural network model module, which includes an object search submodule and a classification submodule; (3) Object search submodule, which can cut out the coagulation precipitation area from the entire reaction card image by deep learning on a large number of reaction card images; (4) The classification submodule learns the feature differences of sample images with known agglutination intensity levels, and then analyzes the agglutination precipitation area images cropped by the object search submodule to determine their agglutination intensity levels; (5) a result output module, used to output the detection results obtained by the two-step convolutional neural network model module; (6) Data preprocessing module: This module performs preprocessing operations such as brightness adjustment, contrast enhancement, and noise removal on the collected images to improve image quality and provide a better data foundation for subsequent analysis.

2. The artificial intelligence assistance system according to claim 1, characterized in that: The two-step convolutional neural network model module has an adaptive training function, which is specifically manifested as follows: when a new detection item or a new type of reaction paper jam appears, the module can automatically incorporate the corresponding image data and accurate detection results into the training set, and by retraining the model, update and optimize the model parameters, thereby improving the model's adaptability to different detection scenarios.

3. An artificial intelligence-assisted method for improving sensor performance in biomedical applications, characterized in that: The following steps are involved: (1) Image data acquisition step: collecting reaction jam image data obtained by the sensor in biomedical testing; (2) Data preprocessing step: Perform preprocessing operations such as brightness adjustment, contrast enhancement, and noise removal on the collected image data; (3) Model analysis step: The pre-processed image data is input into a two-step convolutional neural network model that includes an object search algorithm and a classification algorithm. The object search algorithm is responsible for cutting out the agglutination precipitation area from the entire reaction cardboard image, and the classification algorithm analyzes the cut-out agglutination precipitation area image based on the characteristic differences in agglutination strength to determine its agglutination strength level; (4) Result output step: output the detection results obtained by the two-step convolutional neural network model analysis; (5) Model update step: After obtaining accurate detection data for new detection items or new types of reaction paper jams, these data are added to the training set, the two-step convolutional neural network model is retrained, and the model parameters are updated to improve the model's adaptability and detection accuracy.

Images