Lateral flow immunoassay test paper quantitative detection system and method based on deep learning

Abstract

According to the lateral flow immunoassay test paper quantitative detection system and method based on deep learning, an innovative differential attention mechanism is integrated, key space-time dynamic characteristics of an LFA test strip within 3 minutes can be efficiently extracted and analyzed, high-precision quantitative detection can be completed within extremely short time, the diagnosis time is greatly shortened, and the detection efficiency is improved. Meanwhile, the accuracy of quantitative analysis and the objectivity of a detection result are remarkably improved; a unique time embedding module is built in a differential attention mechanism, and the module enhances the sensitivity of the model to the sequence change of an image sequence and supports built-in interpretability analysis. The requirements of instant detection on equipment miniaturization, low cost, convenient operation and highly reliable result are met.

Description

Technical Field

[0001] This invention belongs to the field of intelligent diagnostic technology of point-of-care testing (POCT), and relates to a quantitative detection system and method for lateral flow immunoassay test strips based on deep learning. Background Technology

[0002] For point-of-care testing (POCT), lateral flow immunoassay (LFA) technology has revolutionized disease diagnosis, food safety assessment, and environmental monitoring due to its simplicity, speed, and cost-effectiveness. The global market value of lateral flow immunoassay technology is approximately $12 billion and is projected to reach $32 billion by 2037. Currently, LFA plays a crucial role in public health. However, traditional LFA still faces key limitations: it typically provides qualitative or semi-quantitative results, and there are significant differences in result interpretation among different populations, which can lead to inaccurate results; furthermore, most commercial tests require 15-30 minutes to obtain results. These limitations weaken the effectiveness of LFA test strips in emergency diagnostic situations and limit their application scenarios. Summary of the Invention

[0003] To address the problems of the prior art, this invention provides a deep learning-based lateral flow immunoassay strip quantitative detection system and method.

[0004] This invention is achieved through the following technical solution: In a first aspect, the present invention provides a deep learning-based lateral flow immunoassay test strip quantitative detection system, comprising: The data acquisition module is used to acquire video data within a preset time period after the sample is added to the lateral flow immunoassay test strip, and to obtain raw time step data; The data preprocessing module is used to segment video data into a continuous time series image set at preset time intervals and identify and read out the detection line region; A deep learning model is used to process the time-series image set to obtain quantitative prediction results; The deep learning model includes: A convolutional neural module is used to extract feature maps from the time-series image set and flatten the feature maps into a sequence of feature vectors. The representation embedding module is used to combine the original time step data to perform sequence embedding and pair embedding on the feature vector sequence, and output the representation sequence and paired representation.

[0005] The core analysis module is used to calculate the attention score based on the representation sequence and paired representations, and to update the representation sequence and paired representations based on the attention score, and output the optimal result after looping. The multilayer perceptron output module is used to aggregate and perform regression analysis on the optimal results, and output the final quantitative prediction results.

[0006] Preferably, the preset time period is within A minutes after the sample is added to the lateral flow immunoassay strip, where A is 3 to 5 minutes.

[0007] Preferably, the convolutional neural module is a ResNet module.

[0008] Furthermore, the ResNet module adopts the ResNet18 architecture.

[0009] Preferably, the characterization embedding module includes a sequence embedding module and a pairing embedding module; The sequence embedding module is used to linearly project the feature vector sequence, generate a temporal position code that can capture temporal information based on the original time step data at the sinusoidal time position, and add the temporal position code to the projected feature vector sequence to form a representation sequence. The pairing embedding module is used to dynamically create attention bias tensors from the raw time-step data to obtain paired representations.

[0010] Preferably, the core analysis module is composed of multiple sub-analysis modules stacked sequentially, and the sub-analysis modules include a differential attention module and an outer product module; The differential attention module receives the representation sequence and paired representation from the previous sub-analysis module and normalizes them. Then, it calculates the attention score through the differential multi-head attention mechanism. The obtained attention score is added back to the received representation sequence through residual connections. The result is normalized and then processed by the SwiGLU activation function. The output of the SwiGLU activation function is added back to the result through residual connections to obtain the updated representation sequence, which is used as the input of the next sub-analysis module. The outer product module receives the updated representation sequence from the current differential attention module, calculates the outer product of the updated representation sequence between two different linear projections, and generates a new paired representation, which serves as the input to the next sub-analysis module.

[0011] Preferably, the data preprocessing module is further used to convert the video data into representations in RGB and HSV color spaces, and then divide the video data into a continuous time series image set at a preset time interval, and identify and read out the detection line region.

[0012] Secondly, the present invention provides a deep learning-based method for quantitative detection of lateral flow immunoassay strips, based on the system described above, comprising: The data acquisition module was used to collect video data within a preset time period after the lateral flow immunoassay test strip was used to add the sample, and the raw time step data was obtained. The video data is divided into continuous time-series image sets at preset time intervals, and the detection line regions are identified and read out. The time-series image set is processed using a deep learning model to obtain quantitative prediction results.

[0013] Thirdly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements the deep learning-based lateral flow immunoassay test strip quantitative detection method as described above.

[0014] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the deep learning-based lateral flow immunoassay strip quantitative detection method as described above.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a fast and accurate deep learning LFA (RAD-LFA) quantitative detection system with the following advantages: (1) Enhanced interpretability: RAD-LFA incorporates a unique temporal embedding module within its differential attention mechanism. This module enhances the model's sensitivity to changes in the order of image sequences and supports built-in interpretability analysis. This design enables the model to clearly reveal the key time nodes and core image patterns driving each diagnostic prediction, thereby providing clinicians with a transparent understanding of the model's decision-making process and effectively enhancing the credibility of diagnostic results and the acceptance of clinical applications. (2) Overcoming the shortcomings of traditional LFA detection, such as long detection time, weak quantitative ability, and subjective result interpretation: RAD-LFA integrates an innovative differential attention mechanism (DAM) and a lightweight ResNet module, which can efficiently extract and analyze the key spatiotemporal dynamic features of LFA test strips within 3 minutes. This design enables it to complete high-precision quantitative detection in a very short time, significantly shortening the diagnostic time and significantly improving the accuracy of quantitative analysis and the objectivity of detection results. (3) Meets the requirements of point-of-care testing (POCT) for miniaturized equipment, low cost, convenient operation, and highly reliable results: RAD-LFA only requires ordinary smartphones for image acquisition and relies on lightweight computational models for analysis, without the need for additional expensive or dedicated testing equipment. In independent tests of LFA samples of various disease types (including COVID-19, hCG, and HBV), it has demonstrated high reliability and quantitative results that are highly consistent with the gold standard laboratory method (ELISA), making it an ideal choice for POCT scenarios such as promotion in resource-scarce areas and home self-testing. (4) Improves the effectiveness of large-scale rapid screening in response to public health emergencies: By combining the spatiotemporal image analysis capabilities of deep learning with conventional LFA test strips, RAD-LFA has successfully reduced the total testing time to 3 minutes, while ensuring a high level of qualitative diagnostic accuracy and quantitative analysis precision. This rapid and accurate characteristic enables it to strongly support rapid screening of large populations, providing key technical support for public health management such as effective control of epidemics, chronic disease management, and pre-pregnancy diagnosis, and is a powerful tool for future disease monitoring and early warning. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a block diagram of the RAD-LFA deep learning model of the present invention.

[0018] Figure 2 The results show the test results of the RAD-LFA deep learning model on COVID-19 samples; (a) the quantitative regression results obtained by the RAD-LFA deep learning model on a 3-minute image sequence prediction set using an independent COVID-19 LFA test set; (b) the receiver operating characteristic (ROC) curve; and (c) the results of the RAD-LFA deep learning model, the publicly available time series model TIMESAVER, and the standard curve method on eight metrics (sensitivity, specificity, accuracy, AUC, precision, recall, F1 score, and R...). 2 (d) Performance comparison on (Quanti-fication); (d) is the absolute value of the difference between the quantitative prediction results of the RAD-LFA deep learning model on COVID-19 samples of different concentrations and the gold standard concentration.

[0019] Figure 3 The following are the test results of the RAD-LFA deep learning model of this invention on HBV samples; (a) the classification ability of the RAD-LFA deep learning model on HBV negative and positive samples; (b) the comparison of the RAD-LFA deep learning model, the published time series model TIMESAVER, and the results of clinicians on eight indicators (sensitivity, specificity, accuracy, AUC, precision, recall, F1 score, and R-value). 2 (c) Performance comparison on different groups (OP: ordinary people, SR: researchers, CD: clinicians, RF: RAD-LFA deep learning model); (d) Absolute value of the difference between the quantitative prediction results of the RAD-LFA deep learning model and the gold standard ELISA concentration on HBV samples of different concentrations; (e) Scatter plot of correlation analysis between the prediction results of the RAD-LFA deep learning model and the detection results of the gold standard ELISA; (f) Comparison of the quantitative prediction results of the RAD-LFA deep learning model and the artificial group (the matrix from top to bottom is RF group, CD group, SR group, OP group).

[0020] Figure 4 The prediction performance of the RAD-LFA deep learning model on hCG samples configured with concentration gradients in the laboratory (i.e., the hCG LFA dataset).

[0021] Figure 5 This is a schematic diagram of the core analysis module of the present invention.

[0022] Figure 6 This is a schematic diagram of a single sub-analysis module in the core analysis module of this invention. Detailed Implementation

[0023] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.

[0024] It should be noted that the process equipment or apparatus not specifically mentioned in the following embodiments are all conventional equipment or apparatus in the art.

[0025] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses. Furthermore, unless otherwise stated, the numbering of each method step is merely a convenient tool for identifying each method step, and not intended to limit the order of the method steps or define the scope of the invention. Changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the invention.

[0026] This invention provides an ultrafast lateral flow immunoassay strip quantitative detection technology based on deep learning. 1) Innovative design integrating a deep learning architecture: A deep learning model is constructed that specifically integrates a lightweight convolutional neural network (such as ResNet18) module for efficiently extracting spatial features of LFA images, and combines it with a novel DyFormer module for accurately capturing temporal dynamic features during the reaction process. Finally, a customized multilayer perceptron (MLP) output layer is used to achieve accurate quantitative or qualitative judgment of analyte concentration. 2) Utilization of spatiotemporal dynamic information in the early stage of the LFA reaction: By capturing and analyzing the weak but discriminative spatiotemporal dynamic changes in color intensity and distribution of the LFA test strip in the initial stage of the reaction (e.g., within the first 3 minutes), the model analysis is based on these changes. This full utilization of early signals allows RAD-LFA to significantly shorten the waiting time required for traditional LFA detection (usually 15-30 minutes) while maintaining or even surpassing the traditional detection accuracy and quantitative precision. 3) Enhanced Time-Sequence Awareness and Interpretability: A unique representation embedding module is designed within the DyFormer module. This module utilizes sine and cosine functions and a learnable multilayer perceptron (MLP) to effectively encode each time step in the time series. This not only enhances the model's sensitivity and understanding of the time series sequence of the LFA response process and optimizes the expression of temporal features, but also provides precise temporal location information for the differential attention mechanism, assisting the model in achieving built-in interpretability for identifying key response time nodes. 4) RAD-LFA combines the spatiotemporal image analysis capabilities of deep learning with conventional LFA test strips, ensuring high-level quantitative analysis accuracy. It provides key technical support for public health management such as effective control of epidemics, chronic disease management, and pre-pregnancy diagnosis, and is a powerful tool for future disease surveillance and early warning. This invention has low computational resource requirements and is easy to deploy on mobile devices. It is expected to provide rapid and reliable diagnosis in diverse POCT scenarios, aiming to narrow the gap between decentralized testing and laboratory-level quantitative testing, and ushering in a new era of precision POCT.

[0027] The specific implementation process of this invention is as follows.

[0028] I. Sample Dataset Preparation (1) Acquisition and processing of COVID-19 LFA dataset 1) To ensure the robustness of the RAD-LFA deep learning model training and validation, this invention adopts the standard COVID-19 LFA dataset reported by SeungminLee et al. in 2024, which is a widely accepted time-series LFA image dataset; this dataset contains a total of 594 samples, including 100 negative samples and 494 positive samples. Positive samples were prepared by serially diluting the protein in the kit run buffer (at 1 / 2-fold gradients, from 50 ng / mL to 0.39 ng / mL, resulting in eight different concentrations).

[0029] 2) The authors of the standard COVID-19 LFA dataset used a commercially available COVID-19 antigen test kit; 100 µL of samples were added to the kit and image sequences were captured at fixed intervals of 10 seconds at multiple specific time points (0.5, 1, 2, 3 and 4 minutes).

[0030] 3) In this invention, the concentration labels of all samples were first normalized for subsequent analysis. Then, the standard COVID-19 LFA dataset was randomly divided into a new training set (n=357), a validation set (n=153), and a separate test set (n=84) in a 6:3:1 ratio. Training, validation, and optimization of the RAD-LFA deep learning model were all performed on these datasets, and the best-performing model configuration on the validation set was applied to the test set for final performance evaluation.

[0031] (2) Preparation of clinical HBV LFA dataset 1) Subjects were screened according to detailed inclusion and exclusion criteria (e.g., age ≥18 years, informed consent; exclusion of other serious infectious diseases, immunosuppressive therapy, etc.); HBV serum samples were obtained from 150 qualified patients; LFA test strips were purchased from InTec PRODUCT, INC. (Xiamen).

[0032] 2) Serum samples were quantified using the gold standard ELISA method and divided into five concentration ranges (e.g., very low 0.5-50 IU / mL to very high >1000 IU / mL). Samples exceeding the very low concentration threshold were serially diluted 1:10 until the target very low concentration range was reached. Through this standardized dilution process, a total of 233 HBV LFA test video samples were generated, which were then allocated to the HBV master dataset and the clinical blind test dataset according to the established method.

[0033] (3) Preparation of hCG LFA dataset 1) Procurement and preparation of hCG standards and reagent strips: Standard hCG samples and hCG LFA reagent strips were purchased from Blue Cross Biomedical (Beijing) Co., Ltd. Twelve hCG standard solutions of different concentrations (concentration range 10, 25, 50, ..., 500 IU / mL) were prepared in the laboratory using a serial dilution method, and a negative control without hCG was also prepared.

[0034] 2) LFA Testing and Image Data Acquisition: For each concentration (including the negative control), three repeated LFA tests were performed, resulting in a total of 39 standard samples (13 concentration groups × 3 times). The entire 5-minute LFA test process for each sample was recorded on video using a smartphone, and then an image frame was captured from the video every 10 seconds.

[0035] 3) Construction and Use of the T-line Image Sequence Dataset: A grayscale thresholding method was applied to the acquired images to identify and crop the T-line region, thereby constructing an hCG LFA time-series image dataset. Due to the limited sample size, this dataset is only used for preliminary exploratory research on the RAD-LFA algorithm.

[0036] II. Development of a Deep Learning-Based lateral flow immunoassay test strip quantitative detection system This invention relates to a deep learning-based lateral flow immunoassay test strip quantitative detection system, comprising: The data acquisition module is used to acquire video data within a preset time period after the sample is added to the lateral flow immunoassay test strip, and to obtain raw time step data; The data preprocessing module is used to segment video data into a continuous time series image set at preset time intervals and identify and read out the detection line region; A deep learning model is used to process the video data to obtain quantitative detection results.

[0037] The deep learning model includes a convolutional neural module and a DyFormer temporal module. The DyFormer temporal module includes a representation embedding module, a core analysis module, and a multilayer perceptron output module.

[0038] (1) Overall architecture of RAD-LFA deep learning model RAD-LFA employs a sophisticated neural network with a Transformer-like architecture, specifically designed for processing time-series image data. It introduces innovative features such as dynamically evolving paired representations and differential attention mechanisms. Its data flow begins with a conceptual data acquisition module responsible for collecting or loading time-series image sets and their corresponding raw time-step data. For example... Figure 1The model receives this data and processes it sequentially through the following modules: First, the convolutional neural module (feature extraction) receives the time-series image set and extracts rich feature tensors for each image using pre-trained ResNet18 convolutional layers. These feature tensors are then flattened to generate a sequence of feature vectors. Next is the representation embedding module (representation initialization), which receives the feature vector sequence x1 output by the convolutional neural module and the raw time-step data t1 from the data acquisition module. This module performs two key operations in parallel: sequence embedding creates a new sequence representation by linearly projecting the feature vector sequence x1 and combining it with a temporal position code generated by processing sinusoidal signals using a small MLP, resulting in a representation sequence; paired embedding dynamically generates an attention bias tensor from the raw time-step data t through two embedding layers. Finally, the representation embedding module outputs a representation sequence x2 containing both feature and position information, as well as an initial paired representation t2, which serves as the initial attention bias for the subsequent core analysis module. Following this is the model's core analysis module (core iterative processing), which consists of multiple sub-analysis modules (Dyblocks) stacked sequentially, such as... Figure 5 As shown. Each DyBlock simultaneously optimizes the representation sequence x2 and paired representation t2 output by the previous module. Figure 6 As shown, within a single DyBlock, the processing flow is divided into three parts: In the attention path, the normalized representation sequence x2 and the normalized paired representation t2 as a dynamic bias are fed into the differential attention module. The attention score of this differential attention module is corrected by a learnable layer depth-related scalar λ. The attention score is added back to the representation sequence x2 through residual connections to generate an intermediate representation y. In the feedforward network path, the normalized intermediate representation y is processed by the SwiGLU activation function, and its output is added back to the intermediate representation y through residual connections to form the updated representation sequence x2 of the differential attention module. When updating the paired representation, the updated representation sequence x2 is fed into the outer product module, which calculates the outer product between two different linear projections of the representation sequence x2 to generate a new paired representation. This paired representation is then added to the input paired representation t2 to update the attention bias for the next DyBlock. This optimization loop is repeatedly executed according to the depth set by the model, and each DyBlock outputs the optimized representation sequence x2 and the optimized paired representation t2. Finally, after the core analysis module completes its processing, the multilayer perceptron output module receives the final optimized representation sequence x2 output by the last DyBlock. It first averages the representation sequence x2 along the sequence dimension for aggregation, then performs classification or regression through a final MLP (containing layer normalization and linear layers), and applies a Sigmoid or Softmax activation function to finally output the model's prediction results.

[0039] Specifically, the characterization embedding module includes a sequence embedding module and a pairing embedding module; The sequence embedding module is used to linearly project the feature vector sequence, generate a temporal position code that can capture temporal information based on the original time step data at the sinusoidal time position, and add the temporal position code to the projected feature vector sequence to form a representation sequence. The pairing embedding module is used to dynamically create attention bias tensors from the raw time-step data to obtain paired representations.

[0040] The core analysis module is composed of multiple sub-analysis modules stacked sequentially, including a differential attention module and an outer product module. The differential attention module receives the representation sequence and paired representation from the previous sub-analysis module and normalizes them. Then, it calculates the attention score through the differential multi-head attention mechanism. The obtained attention score is added back to the received representation sequence through residual connections. The result is normalized and then processed by the SwiGLU activation function. The output of the SwiGLU activation function is added back to the result through residual connections to obtain the updated representation sequence, which is used as the input of the next sub-analysis module. The outer product module receives the updated representation sequence from the current differential attention module, calculates the outer product of the updated representation sequence between two different linear projections, and generates a new paired representation, which serves as the input to the next sub-analysis module.

[0041] (2) Core DyFormer timing module mechanism The key innovations of DyFormer's timing module lie in its representation embedding module and differential attention module. The representation embedding module first encodes each time step as a high-dimensional sinusoidal representation using sine and cosine functions (e.g., Equation (2)), effectively capturing multiple time frequencies determined by the exponential decay function. Subsequently, a learnable MLP transforms this time frequency embedding into a final time position embedding of the required dimension (e.g., Equation (3)), where 𝑡 𝑏,𝑠 The input represents time. The differential attention module introduces a differential weighting mechanism on the basis of the traditional multi-head self-attention mechanism (firstly, the standard query Q, key K, and value V are projected, and the scaled dot product attention is calculated, see formula (4)). This differential weighting mechanism divides the attention score into two groups and uses the learned lambda parameters (λq1, λk1, λq2, λk2) for differential modulation (as shown in formula (5)) to optimize the attention weights. Finally, the output is calculated by weighted summation of the value vectors (as shown in formula (6)) and then subjected to layer normalization (RMSNorm) and output projection. During the model training phase, the mean squared error (MSE) is used as the loss function (as shown in formula (7)). This loss function optimizes the regression task by imposing a heavier penalty on larger errors. The regression performance of the model is finally measured using R 2(The coefficient of determination, as shown in formula (8)) is used for evaluation, R 2 The higher the value, the better the model's prediction matches the actual value, and the better its quantitative performance.

[0042] (1) (2) (3) (4) (5)

[0043] (6) (7) (8) Where y b,s : Represents the output feature vector of a specific image in a batch and sequence; b: Represents the batch index, indicating which sequence in the current batch the image belongs to; s: Represents the sequence index, indicating the position of the image in its sequence; W: A weight matrix (or linear transformation matrix); flatten(•): Flattening operation; ResNet(•): Deep convolutional neural network; Conv(•): Convolutional layer; x b,s b: Input image data for a specific image in a batch or sequence; b: Bias vector, a learnable parameter; ω i : Represents the frequency term at index i in a specific dimension, these frequencies are used to generate the sinusoidal pattern of the encoded position; exp(•): This is the exponential function; log(•): This is the natural logarithm function; log(•): This is the natural logarithm function; log max_period×i: Defines the maximum period of the sinusoidal function; TE(t b,s ) represents a specific time step t b,s The generated temporal embedding; MLP(•): Multi-Layer Perceptron, also known as a fully connected neural network; frequencies: the range of frequency indices; Attention Weights: a computed matrix representing the degree of attention each element in the input sequence gives to each element in the output sequence; Q (Query): the query matrix, representing the elements being searched or followed; K (Key): the key matrix, representing the elements available for query matching; λ full: Represents a complete, learnable, layer-depth-dependent scalar used to modify the attention scores within the differential attention module, allowing the model to dynamically adjust the attention pattern at each layer; exp(•): exponential function; head_dim: dimension of the attention head; (λ q1 ) i and (λ) k1 ) i Two distinct learnable vectors λ q1 and λ k1 The i-th element; (λ q2 ) i and (λ) k2 ) i Another pair of learnable vectors λ q1 and λ k1 The i-th element; λ init Learnable scalar initialization bias, which is λ full It provides a baseline value that can be learned during training; atten_weights diff : Final, adjusted attention weights; output: Final model prediction output; V (Value): Value matrix, containing the actual information or features in the input sequence to be aggregated; MSE: Mean Squared Error, a commonly used metric to measure the difference between predicted and true values. The smaller the MSE value, the closer the model's prediction is to the true value; y i y: The true value of the i-th sample; i ̂: The predicted value of the i-th sample; R 2 The coefficient of determination, also known as the R-squared value, is a statistic between 0 and 1 (it can be negative in some special cases) used to measure how well a regression model fits the data. 2 The higher the value, the better the model is for the dependent variable y. i The more variation that can be explained, the better the model fits.

[0044] III. Training and Optimization of RAD-LFA Deep Learning Model (1) Training of RAD-LFA deep learning model During the lab training phase, an open-source COVID-19 time-series image dataset was used as samples, with time intervals of 10 seconds and an input time of 3 minutes. The model training process began with data preparation: the COVID-19 time-series image dataset was loaded and divided into training, validation, and test sets according to configuration parameters (such as sequence length, batch size, image size, and dataset partitioning ratio), and corresponding data loaders were created. The model was initialized according to predefined architecture parameters and deployed to a specified computing device (CPU or GPU). Training used the Adam optimizer, with configurable learning rate (lr) and weight decay, adjusted via a learning rate scheduler that reduced the learning rate after a specific step size. The loss function used was mean squared error (MSELoss). Training loops iterated within a specified number of epochs (default 100). In each training epoch, the model was set to training mode. Training data was fed into the model in batches: input image sequences, time-step data, and corresponding target labels were transmitted to the device. The model performed forward propagation to generate predictions and then calculated the MSE loss between the predicted and true values. Next, backpropagation is performed to calculate the gradient, and the model parameters are updated via the optimizer. Simultaneously, the learning rate scheduler also updates the learning rate. After each training epoch, the total loss and R0 on the training set are calculated. 2 Scores and classification accuracy are calculated. The model is then switched to evaluation mode, where similar forward propagation and metric calculations are performed on the validation set without calculating gradients to evaluate the model's generalization ability. Performance metrics for each epoch are logged to a log file and output. R metrics on the validation set are saved during training. 2 The model with the highest score is designated as the "best model," and its predictions are saved along with the actual values ​​in an Excel file.

[0045] Figure 2 The results show the test results of the RAD-LFA deep learning model on COVID-19 samples. Figure 2 In the middle (a), the quantitative regression results of the RAD-LFA deep learning model predicted using a 3-minute image sequence on an independent COVID-19 LFA test set are shown. 2 =0.97, indicating that the predicted value of the RAD-LFA deep learning model is in high agreement with the actual value and has excellent quantitative performance. Figure 2 In the middle (b), the receiver operating characteristic curve (ROC curve) is shown. The ROC curve demonstrates the classification performance of the RAD-LFA deep learning model. The area under the curve is 0.95 (AUC = 0.95), indicating that the RAD-LFA deep learning model has excellent classification performance. Figure 2(c) shows the RAD-LFA deep learning model, the publicly available time series model TIMESAVER, and the standard curve method in eight metrics (sensitivity, specificity, accuracy, AUC, precision, recall, F1 score, and R-squared). 2 The performance comparison on the time series model shows that the RAD-LFA deep learning model of this invention has an absolute advantage over the publicly available time series model TIMESAVER and the standard curve method. Figure 2 In the middle (d), the absolute value of the difference between the quantitative prediction results of the RAD-LFA deep learning model on COVID-19 samples of different concentrations and the gold standard concentration can be seen. It can be seen that the absolute value of the difference is basically controlled at around 0.1.

[0046] Figure 4 The blue line represents the prediction performance of the RAD-LFA deep learning model on hCG samples (i.e., the hCG LFA dataset) configured with concentration gradients in the laboratory. The red line represents the performance on the training set and the red line represents the performance on the validation set. Figure 4 The results show that, regardless of the training or validation set, the prediction results and the error of the true value of RAD-LFA remain within a reasonable range. The training set R... 2 A value greater than 0.9 indicates that the performance degradation on the validation set can be attributed to the limited amount of data and the potential for optimization.

[0047] The results above show that the core performance metrics of the RAD-LFA deep learning model of this invention (with the input being an image sequence of the first 3 minutes of LFA reaction time) on the independent COVID-19 LFA dataset include: after training stabilization, R... 2 A value greater than 0.9 indicates a classification accuracy exceeding 95%, and the area under the ROC curve (AUC) reaches 0.95. Compared with existing methods such as the standard curve method and TimeSaver, it demonstrates superior sensitivity, AUC, and quantification capability (R²). 2 It outperforms or performs comparable to the model in eight key performance indicators, including a performance index as high as 0.97. The model's predicted concentrations show a high degree of consistency with the actual concentration values, and it maintains stable performance across low, medium, and high concentration ranges. Its excellent quantitative ability has also been verified in standard hCG LFA sample tests. Overall, the RAD-LFA deep learning model can serve as a reliable and efficient tool for qualitative and quantitative analysis.

[0048] (2) Optimization of RAD-LFA deep learning model In the optimization process of the RAD-LFA model, its three core components were systematically improved to significantly enhance prediction performance. Firstly, in the data preprocessing module, the length of the image sequence was optimized. Experiments showed that a sequence of 6 frames per minute and a total length of 3 minutes could simultaneously achieve excellent performance (R... 2The model achieves a resolution of 0.97 and a classification accuracy exceeding 95%, while effectively shortening the reaction time. Next, the influence of image input channels was explored in depth, comparing RGB, HSV, and their combinations. The results show that the joint input of RGB+HSV provides the best quantitative prediction and classification performance, thanks to the model's ability to simultaneously utilize the mixed color information of RGB and the perceptual attributes of HSV, achieving more comprehensive feature extraction. Secondly, the optimization of the convolutional neural module was confirmed through ablation experiments, demonstrating its crucial role in the overall prediction accuracy of the model. Further comparisons were made of four architectures: ResNet18, ResNet34, ResNet50, and ResNet101. ResNet18 was ultimately selected because it exhibited the best prediction performance in the specific task of this invention, while maintaining the lowest computational complexity (FLOPs), effectively avoiding overfitting on small-scale LFA datasets, and benefiting training efficiency and future deployment. Finally, during the refinement of the core analysis module, the nonlinear relationship between its module depth and the model's root mean square error (RMSE) was analyzed, determining the optimal depth—that is, the balance between capturing complex data patterns and avoiding overfitting. Furthermore, ablation experiments on key components of the characterization embedding module strongly validated the effectiveness of the integrated design, demonstrating the significant advantages and stability of the complete RAD-LFA configuration in quantitative detection capabilities. Its mean absolute difference in prediction was 47.2% lower than the worst ablation group, and R... 2 Significantly superior. IV. Clinical Applications of the RAD-LFA Deep Learning Model This invention uses the RAD-LFA deep learning model to test 233 HBV samples from clinical patients (i.e., the clinical HBV LFA dataset), and the results are as follows. Figure 3 As shown. Figure 3 In the middle (a), the RAD-LFA deep learning model is shown to classify HBV negative and positive samples, with a true positive rate of 95% and a true negative rate of 92%. Figure 3 (b) shows the RAD-LFA deep learning model, the publicly available time series model TIMESAVER, and the clinical application of the model across eight metrics (sensitivity, specificity, accuracy, AUC, precision, recall, F1 score, and R-value). 2 The performance comparison shows that, across all metrics, the results obtained by the RAD-LFA deep learning model of this invention are superior to the publicly available time series model TIMESAVER and those obtained by clinicians. Figure 3 In the middle (c), the classification accuracy of different groups is shown (OP: ordinary people, SR: scientific researchers, CD: clinicians, RF: RAD-LFA deep learning model). The results show that the classification accuracy of the RAD-LFA deep learning model of this invention is 94%, which is significantly higher than the human interpretation results of different groups. Figure 3In the middle (d), the absolute value of the difference between the quantitative prediction results of the RAD-LFA deep learning model on HBV samples of different concentrations and the gold standard ELISA concentration can be seen. It can be seen that the absolute value of the difference is basically controlled at around 0.2. Figure 3 The middle (e) plot shows the scatter plot of the prediction results of the RAD-LFA deep learning model and the gold standard ELISA detection results. The scatter plot and Pearson correlation analysis (r = 0.948, P<0.0001) show that the prediction results of the RAD-LFA deep learning model and the gold standard ELISA detection results are highly consistent. Figure 3 The middle (f) shows the comparison between the quantitative prediction results of the RAD-LFA deep learning model and the manual group (the matrix from top to bottom represents the RF group, CD group, SR group, and OP group), which shows that the prediction results of the RAD-LFA deep learning model are closer to the actual results.

[0049] As can be seen, the RAD-LFA deep learning model of this invention achieved an accuracy of 94% (significantly better than the three groups of manual interpretations, of which the clinician group achieved 78%), specificity of 92%, and sensitivity of 95% in tests on 233 HBV samples from clinical patients. The quantitative detection results were highly correlated with the gold standard ELISA method (Pearson r = 0.948). 2 The value reached 0.9985 (a 10% improvement over the clinician group). A single sample analysis took only 0.2 seconds. Considering all key indicators, the RAD-LFA deep learning model comprehensively outperformed manual interpretation and methods such as TimeSaver, demonstrating significant clinical application value.

[0050] This invention also provides a computer device including a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to achieve a corresponding method flow or function. The processor described in this embodiment can be used for the operation of a deep learning-based lateral flow immunoassay strip quantitative detection method.

[0051] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be Random Access Memory (RAM) or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the quantitative detection method for the deep learning-based lateral flow immunoassay test strip in the above embodiments.

[0052] This embodiment also provides a computer program product, which includes a computer program that, when executed by a processor, implements the corresponding steps of the deep learning-based lateral flow immunoassay strip quantitative detection method described in the above embodiment.

[0053] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0054] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, systems (apparatus), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0055] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0056] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0057] 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 it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A deep learning-based lateral flow immunoassay test strip quantitative detection system, characterized in that, include: The data acquisition module is used to acquire video data within a preset time period after the sample is added to the lateral flow immunoassay test strip, and to obtain raw time step data; The data preprocessing module is used to segment video data into a continuous time series image set at preset time intervals and identify and read out the detection line region; A deep learning model is used to process the time-series image set to obtain quantitative prediction results; The deep learning model includes: A convolutional neural module is used to extract feature maps from the time-series image set and flatten the feature maps into a sequence of feature vectors. The representation embedding module is used to combine the original time step data to perform sequence embedding and pair embedding on the feature vector sequence, and output the representation sequence and the paired representation; The core analysis module is used to calculate the attention score based on the representation sequence and paired representations, and to update the representation sequence and paired representations based on the attention score, and output the optimal result after looping. The multilayer perceptron output module is used to aggregate and perform regression analysis on the optimal results, and output the final quantitative prediction results.

2. The deep learning-based lateral flow immunoassay strip quantitative detection system according to claim 1, characterized in that, The preset time period is within A minutes after the sample is added to the lateral flow immunoassay test strip, where A is 3 to 5 minutes.

3. The deep learning-based lateral flow immunoassay strip quantitative detection system according to claim 1, characterized in that, The convolutional neural module is a ResNet module.

4. The deep learning-based lateral flow immunoassay test strip quantitative detection system according to claim 3, characterized in that, The ResNet module adopts the ResNet18 architecture.

5. The deep learning-based lateral flow immunoassay strip quantitative detection system according to claim 1, characterized in that, The characterization embedding module includes a sequence embedding module and a pairing embedding module; The sequence embedding module is used to linearly project the feature vector sequence, generate a temporal position code that can capture temporal information based on the original time step data at the sinusoidal time position, and add the temporal position code to the projected feature vector sequence to form a representation sequence. The pairing embedding module is used to dynamically create attention bias tensors from the raw time-step data to obtain paired representations.

6. The deep learning-based lateral flow immunoassay test strip quantitative detection system according to claim 1, characterized in that, The core analysis module is composed of multiple sub-analysis modules stacked sequentially, including a differential attention module and an outer product module. The differential attention module receives the representation sequence and paired representation from the previous sub-analysis module and normalizes them. Then, it calculates the attention score through the differential multi-head attention mechanism. The obtained attention score is added back to the received representation sequence through residual connections. The result is normalized and then processed by the SwiGLU activation function. The output of the SwiGLU activation function is added back to the result through residual connections to obtain the updated representation sequence, which is used as the input of the next sub-analysis module. The outer product module receives the updated representation sequence from the current differential attention module, calculates the outer product of the updated representation sequence between two different linear projections, and generates a new paired representation, which serves as the input to the next sub-analysis module.

7. The deep learning-based lateral flow immunoassay strip quantitative detection system according to claim 1, characterized in that, The data preprocessing module is also used to convert the video data into representations in RGB and HSV color spaces, and then divide the video data into a continuous time series image set at a preset time interval, and identify and read out the detection line region.

8. A quantitative detection method for lateral flow immunoassay strips based on deep learning, characterized in that, The system based on any one of claims 1 to 7 includes: The data acquisition module was used to collect video data within a preset time period after the lateral flow immunoassay test strip was used to add the sample, and the raw time step data was obtained. The video data is divided into continuous time-series image sets at preset time intervals, and the detection line regions are identified and read out. The time-series image set is processed using a deep learning model to obtain quantitative prediction results.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the deep learning-based lateral flow immunoassay test strip quantitative detection method as described in claim 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the deep learning-based lateral flow immunoassay test strip quantitative detection method as described in claim 8.

Images