Artificial intelligence-based automatic gray scale quantification method for lateral flow immunochromatographic test strip

Abstract

This invention belongs to the field of lateral flow immunochromatographic detection technology, specifically an automated grayscale quantitative method for lateral flow immunochromatographic test strips based on artificial intelligence. This invention employs a trained YOLO11 instance segmentation model to achieve precise positioning and pixel-level mask segmentation of the test strip cartridge and marking area. The entire quantitative process is completed through a region of interest (ROI) grayscale value analysis algorithm: convex quadrilateral fitting is performed on the segmentation mask to obtain the precise quadrilateral boundaries of the cartridge and marking area; perspective transformation is used to unify the spatial orientation of the cartridge, completing the positioning of the core ROI; adaptive extraction and standardized calculation of grayscale signals for the detection line and control line areas are achieved; finally, the results are substituted into a pre-stored grayscale-concentration calibration model to obtain the analyte concentration. This invention eliminates the need for a custom imaging support and enables accurate and automated quantitative detection of test strips under complex shooting angles, lighting conditions, and backgrounds, significantly improving the convenience, stability, and reliability of lateral flow immunochromatographic detection.

Description

Technical Field

[0001] This invention belongs to the field of lateral flow immunochromatographic detection technology, specifically relating to an automatic grayscale quantitative method for lateral flow immunochromatographic test strips based on artificial intelligence. Background Technology

[0002] Lateral flow immunochromatography (LFIA) technology, due to its advantages of simple operation, rapid detection, low cost, and no need for large instruments, has been widely used in clinical diagnosis, food safety, environmental monitoring, and other fields, and is one of the core technologies of point-of-care testing (POCT). Traditional LFIA test strip results interpretation mainly relies on manual visual observation, which can only achieve qualitative judgment of positive or negative results, making precise quantitative analysis difficult. Furthermore, the interpretation results are easily affected by factors such as operator subjective experience and visual fatigue, resulting in significant human error. With the development of image analysis technology, grayscale analysis methods based on software such as ImageJ have emerged, which calculate grayscale values ​​by manually selecting the detection line (T line) and control line (C line) regions to achieve quantification. However, this method still requires manual operation, resulting in low detection efficiency. Fluctuations in imaging conditions such as shooting distance, shooting angle, ambient lighting, and test strip placement can severely affect the accuracy of grayscale value calculation, leading to poor reliability of quantitative results.

[0003] In recent years, the combination of artificial intelligence technology and lateral flow immunochromatography has provided a new direction for solving the above problems. Existing AI-assisted detection methods often rely on customized imaging scaffolds or darkroom equipment to fix imaging conditions. While this improves detection stability, it increases detection costs and equipment size, negating the core advantages of lateral flow immunochromatography: portability and low cost. Other methods can only achieve positive / negative or concentration classification, failing to meet the needs of precise quantitative detection. Still others have weak anti-interference capabilities, operating only under standard imaging conditions and unable to adapt to the complex imaging environments of real-world applications, limiting their on-site deployment and application. Therefore, developing a lateral flow immunochromatographic test strip quantitative detection method that does not require customized imaging equipment and can be automated under actual imaging conditions has significant research and application value. Summary of the Invention

[0004] The purpose of this invention is to provide an automated grayscale quantitative method for lateral flow immunochromatographic test strips based on artificial intelligence, so as to overcome the above-mentioned shortcomings of existing lateral flow immunochromatographic test strip interpretation technology.

[0005] This invention provides an automated grayscale quantitative method for lateral flow immunochromatographic test strips based on artificial intelligence. The method includes using a trained YOLO11 instance segmentation model to achieve precise positioning and pixel-level mask segmentation of the test strip cartridge and identification area; and completing the entire quantitative process through a custom region of interest (ROI) grayscale value analysis algorithm: performing convex quadrilateral fitting on the segmentation mask to obtain the precise quadrilateral boundaries of the cartridge and identification area; unifying the spatial orientation of the cartridge through perspective transformation to complete the positioning of the core ROI; achieving adaptive extraction and standardized calculation of grayscale signals for the detection line and control line areas; and finally, substituting the results into a pre-stored grayscale-concentration calibration model to obtain the analyte concentration result. This invention does not rely on a customized imaging support and can automatically complete the entire process of cartridge positioning, ROI extraction, grayscale calculation, and concentration conversion under different imaging conditions, effectively solving the technical problems of large manual operation errors, weak anti-interference ability of imaging conditions, insufficient quantitative accuracy, and reliance on dedicated equipment in existing technologies.

[0006] The specific steps are as follows:

[0007] S1 acquires the raw image of the test result of the lateral flow immunochromatographic test strip as the detection input;

[0008] S2 uses a pre-trained instance segmentation model to recognize the original image and simultaneously outputs pixel-level segmentation masks for the test strip box area and the identification area.

[0009] S3 performs image geometric correction based on the segmentation results of the card box area and the identification area to unify the spatial orientation of the test strip;

[0010] S4 locates the core region of interest (ROI) containing the detection line (T line) and the control line (C line) in the corrected card image, and adaptively optimizes the extraction of the effective computational regions of the T line and the C line;

[0011] S5 calculates the gray values ​​of the T-line, C-line, and background area respectively, and obtains the standardized gray-scale ratio signal after background correction;

[0012] S6 substitutes the standardized gray-scale ratio signal into the pre-stored gray-scale-concentration calibration model of the analyte in the system to calculate the concentration result of the target analyte in the sample.

[0013] Furthermore:

[0014] In step S2, the instance segmentation model adopts the YOLO11 network architecture, and the process of pre-training to obtain the instance segmentation model is as follows:

[0015] S21 Construct the original dataset: Collect standard images of lateral flow immunochromatographic test strips of different concentrations, including valid detection images, invalid detection images, and blank cartridge images;

[0016] S22 Image Purification Processing: Background removal and contour extraction are performed on the original image to generate a standardized card image with an alpha channel;

[0017] S23 Data Augmentation and Dataset Construction: The purified card image is randomly synthesized with various background images, and multiple rounds of random data augmentation are performed by combining geometric transformation and photometric transformation to generate an augmented dataset;

[0018] S24 Dataset Partitioning: The augmented dataset is divided into training and validation sets proportionally for model training and performance validation.

[0019] S25 Model Training: An instance segmentation network architecture is adopted, and the model is trained using the training set data. The convergence of the model is monitored through the validation set to obtain the trained instance segmentation model.

[0020] S26 Model Evaluation: Construct an independent test set to evaluate model performance.

[0021] The geometric transformations in the data augmentation include one or more combinations of random rotation, horizontal flipping, scaling, translation, cropping, and perspective transformation; the photometric transformations include one or more combinations of random brightness adjustment, contrast adjustment, Gaussian blur, and Gaussian noise addition.

[0022] The instance segmentation model sets up two types of detection targets: card box regions and marker regions. The card box regions are used to provide geometric boundaries to support subsequent perspective correction processing, while the marker regions serve as directional references to achieve consistent sorting of card box coordinates, ensuring the standardization of subsequent ROI positioning and grayscale analysis.

[0023] The model performance evaluation used an independent test set consisting of test strip images under different imaging conditions (different lighting, backgrounds, and shooting angles, etc.). Precision was used to evaluate the model's false positive rate in identifying the card box region and the marker region, while recall was used to evaluate the model's false negative rate in identifying these two types of targets. mAP50 comprehensively evaluated the model's overall performance in terms of positioning accuracy and mask segmentation consistency. The confusion matrix was used to further verify the model's classification accuracy for the card box region and the marker region.

[0024] In step S3, the specific process of image geometric correction is as follows: First, morphological denoising is performed on the segmentation mask of the card box area and the label area, and the precise quadrilateral boundary of the card box and the label area is obtained by fitting a convex quadrilateral; then, geometric validity is checked to confirm that the label area is completely contained within the card box area, and invalid detection results are filtered out; finally, perspective transformation is performed based on the coordinates of the four corner points of the card box to correct the tilt, rotation and perspective distortion of the test strip, and unify the spatial pose of all card box images.

[0025] In step S4, the specific process of ROI positioning and strip optimization is as follows: In the corrected card box image, the core ROI covering the T line and C line is located according to the preset proportional coordinates; the initial positions of the T line and C line are initially determined within the ROI; a rectangle optimization algorithm based on adaptive threshold is used to finely adjust the rectangular area of ​​the T line and C line so that the area boundary fits the actual strip range and redundant background interference is eliminated.

[0026] In step S5, the normalized grayscale ratio for background correction is calculated as follows:

[0027] The average gray value T of the optimized T-line region, the average gray value C of the C-line region, and the average gray value B of the background region after subtracting the T-line and C-line within the ROI are calculated respectively. The T / C gray value ratio after background correction is calculated according to the formula (BT) / (BC) and used as the standardized signal for quantitative detection.

[0028] Furthermore, an invalid result filtering step is set after grayscale calculation: invalid test strip signals without valid C lines are screened by judging the grayscale difference, and abnormal detection results are automatically excluded.

[0029] In step S6, the gray-scale-concentration calibration model is a standard curve model pre-established for the corresponding detection item, and different detection items can be configured with independent calibration models.

[0030] Compared with the prior art, the specific beneficial effects of the present invention are as follows:

[0031] This invention eliminates the need for customized imaging supports, darkrooms, handheld signal readers, and other specialized hardware. Through algorithmic instance segmentation, geometric correction, and signal standardization, it accurately locates the T-line and C-line under varying shooting angles, lighting conditions, and background environments, retaining the core advantages of lateral flow immunochromatography: portability and low cost. It is suitable for rapid on-site testing scenarios. This method requires no manual intervention from image input to concentration output, avoiding operational errors inherent in traditional manual analysis. It effectively counteracts the interference of imaging condition fluctuations on grayscale signals, significantly improving the stability and accuracy of quantitative results. Furthermore, this solution is not limited to specific detection targets; by changing the corresponding calibration model, it can be adapted to test strips for different detection items. In addition, the training methods, including image synthesis and multi-dimensional data augmentation, reduce the cost of real sample collection, demonstrating good versatility and practical application value. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the overall process of the automated quantitative detection method for lateral flow immunochromatographic test strips described in this invention.

[0033] Figure 2This is a training result diagram of the YOLO11 instance segmentation model described in this invention. In the diagram, A represents the curves showing the change in precision and recall of the bounding box output over the training epochs; B represents the curves showing the change in precision and recall of the segmentation mask output over the training epochs; and C represents the curve showing the change in the mAP50 metric over the training epochs.

[0034] Figure 3 This is an example diagram verifying the YOLO11-RGVA localization effect of the present invention under different imaging conditions. Detailed Implementation

[0035] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to better understand the technical solution of the present invention and should not be construed as limiting the present invention. The present invention uses a lateral flow immunochromatographic test strip for alpha-fetoprotein (AFP) as an example to verify the implementation process and performance of the method of the present invention, as detailed below:

[0036] This embodiment specifically includes:

[0037] (I) Dataset Construction

[0038] 145 original AFP test strip images were collected, covering different AFP concentrations, including valid test images, invalid test images, and blank cartridge images. Background removal and contour extraction were performed on the original images to generate purified cartridge images with alpha channels. These purified cartridge images were randomly synthesized with different real-world background images at random positions, angles, and scaling ratios, resulting in 10,000 composite images. Then, geometric transformations such as rotation, scaling, translation, cropping, and perspective transformations, as well as photometric transformations such as random brightness and contrast adjustments, Gaussian blur, and Gaussian noise, were randomly combined. Each composite image was enhanced 10 times with randomly selected transformation combinations and parameters, ultimately generating an augmented dataset of 100,000 images. The augmented dataset was randomly divided into training and validation sets at a 9:1 ratio.

[0039] (II) Training and Evaluation of Instance Segmentation Model

[0040] The YOLO11s-seg instance segmentation model was adopted, defining two target classes: card box regions and AFP marker regions. The input image size was 640×640 pixels, and the training epochs were 250, performed on a GPU device. During training, precision, recall, and mAP50 metrics continuously increased and converged to a high level, such as... Figure 2 As shown, this indicates that the model is well trained and can accurately identify the card box region and the label region.

[0041] (III) Implementation of Region of Interest Gray Value Analysis (RGVA) Algorithm

[0042] The RGVA algorithm, as the core quantitative processing module of this system, performs quantitative processing immediately after the YOLO11 model identifies the card box and the marked region. This algorithm automatically extracts the grayscale features of the region of interest throughout the entire process, realizing the conversion from visual color information to quantifiable data for density analysis. The specific steps are as follows:

[0043] (1) Mask preprocessing: Morphological opening operation and erosion denoising are performed on the segmentation mask of the card box and the identification area. The four corner coordinates and precise quadrilateral boundaries of the two areas are extracted by convex quadrilateral fitting.

[0044] (2) Validity verification: The verification mark area is completely contained within the card box area and the relative position conforms to the test strip structure, filtering out invalid test results.

[0045] (3) Perspective correction: Based on the four corner coordinates of the card box, perspective transformation is performed to eliminate tilt, rotation and perspective distortion, and all test strip images are unified into the standard front view posture.

[0046] (4) ROI positioning: In the corrected front view, locate the core ROI containing the T line and the C line according to the preset scale coordinates.

[0047] (5) T / C line region optimization: Initially locate the initial rectangular regions of T and C lines within the ROI, and adjust the region boundaries through an adaptive threshold rectangle optimization algorithm to make them closely fit the actual line range and eliminate redundant background interference.

[0048] (6) Grayscale calculation and correction: Convert the RGB image to a grayscale image using the standard luminance formula. The conversion formula is as follows:

[0049] Gray = 0.2989 × R + 0.5870 × G + 0.1140 × B. Calculate the average gray value T of the T-line region, the average gray value C of the C-line region, and the average gray value B of the background region within the ROI. Calculate the normalized T / C gray ratio after background correction using the following formula: Normalized gray ratio = (BT) / (BC).

[0050] (7) Invalid result filtering: The validity of the C line is determined by the gray difference. If the gray difference between the background and the C line is lower than the preset threshold, it is determined to be an invalid result and an abnormal prompt is output.

[0051] (iv) YOLO11-RGVA test results

[0052] An independent test set was constructed using 118 test strip images not used in model training, covering different AFP concentrations and various imaging conditions. The YOLO11 model performance test results are as follows: Bounding box output: Precision 1.000, Recall 1.000, mAP50 0.995; Segmentation mask output: Precision 0.987, Recall 0.987, mAP50 0.984; YOLO11-RGVA localization results are shown below. Figure 3 As shown, this demonstrates the accurate and robust ROI identification of the method of the present invention under actual imaging conditions.

[0053] (v) Validation of quantitative detection process

[0054] An AFP lateral flow immunochromatographic assay strip was used for testing. The test result image was captured using a smartphone and input into the YOLO11-RGVA detection system of this invention. The system first outputs a segmentation mask for the card box and the marker area using an instance segmentation model. Then, based on the segmentation mask result, the RGVA algorithm performs convex quadrilateral fitting to obtain the coordinates of the card box and the marker area. A perspective transformation is then performed to obtain a standard view, which is used to locate the ROI region. The T-line and C-line are extracted and the background-corrected T / C grayscale ratio is calculated to be 1.73. This ratio is then substituted into the system's pre-stored AFP grayscale-concentration calibration model to calculate the AFP concentration in the sample as 332 ng / mL. The entire detection process requires no manual intervention, achieving accurate and automated quantitative detection.

[0055] The embodiments described above are merely illustrative of implementation methods of the present invention and are used to illustrate the technical solutions of the present invention, not to limit it. Those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. An automated grayscale quantitative method for lateral flow immunochromatographic test strips based on artificial intelligence, characterized in that, The process includes using a trained YOLO11 instance segmentation model to achieve precise localization and pixel-level mask segmentation of the test strip cartridge and labeling area; and completing the entire quantitative processing through a custom region of interest (ROI) grayscale value analysis algorithm: performing convex quadrilateral fitting on the segmentation mask to obtain the precise quadrilateral boundaries of the cartridge and labeling area, unifying the spatial orientation of the cartridge through perspective transformation, completing the localization of the core ROI, achieving adaptive extraction and grayscale signal standardization calculation of the detection line and control line areas, and finally substituting the results into the system's pre-stored grayscale-concentration calibration model to obtain the analyte concentration result; the specific steps are as follows: S1 acquires the raw image of the test results of the lateral flow immunochromatographic test strip; S2 uses a pre-trained instance segmentation model to recognize the original image and simultaneously outputs pixel-level segmentation masks for the test strip box area and the identification area. S3 performs image geometric correction based on the segmentation results of the card box area and the identification area to unify the spatial orientation of the test strip; S4 locates the core region of interest (ROI) containing the detection line (T line) and the control line (C line) in the corrected card image, and adaptively optimizes the extraction of the effective computational regions of the T line and the C line; S5 calculates the gray values ​​of the T-line, C-line, and background area respectively, and obtains the standardized gray-scale ratio signal after background correction; S6 substitutes the standardized gray-scale ratio signal into the pre-stored gray-scale-concentration calibration model of the analyte in the system to calculate the concentration result of the target analyte in the sample.

2. The detection method according to claim 1, characterized in that, The method for constructing the instance segmentation model in step S2 includes the following steps: S21 Construct the original dataset: Collect standard images of lateral flow immunochromatographic test strips of different concentrations, including valid detection images, invalid detection images, and blank cartridge images; S22 Image Purification Processing: Background removal and contour extraction are performed on the original image to generate a standardized card image with an alpha channel; S23 Data Augmentation and Dataset Construction: The purified card image is randomly synthesized with various background images, and multiple rounds of random data augmentation are performed by combining geometric transformation and photometric transformation to generate an augmented dataset; S24 Dataset Partitioning: The augmented dataset is divided into training and validation sets proportionally for model training and performance validation. S25 Model Training: An instance segmentation network architecture is adopted, and the model is trained using the training set data. The convergence of the model is monitored through the validation set to obtain the trained instance segmentation model. S26 Model Evaluation: Construct an independent test set to evaluate model performance.

3. The detection method according to claim 2, characterized in that, In step S23, the geometric transformation includes one or more combinations of random rotation, horizontal flip, scaling, translation, cropping, and perspective transformation; the photometric transformation includes one or more combinations of random brightness adjustment, contrast adjustment, Gaussian blur, and Gaussian noise addition.

4. The detection method according to claim 2, characterized in that, In step S25, the instance segmentation model defines two types of detection targets: card box regions and marker regions; wherein the card box regions are used to provide geometric boundaries for subsequent perspective correction, and the marker regions serve as directional references for consistent card box coordinate sorting, supporting standardized ROI positioning and grayscale analysis.

5. The detection method according to claim 1, characterized in that, The specific method for image geometric correction in step S3 is as follows: S31 performs morphological denoising on the segmentation mask of the card box area and the logo area, and obtains the precise quadrilateral boundary of the card box and the logo area by fitting a convex quadrilateral. S32 performs geometric validity verification to confirm that the marking area is completely contained within the card box area and filters out invalid detection results. S33 performs perspective transformation based on the corner coordinates of the card box, corrects the tilt, rotation and perspective distortion of the test strips, and unifies the spatial pose of all card box images.

6. The detection method according to claim 1, characterized in that, The specific methods for ROI positioning and T / C line optimization described in step S4 are as follows: S41 locates the core ROI covering the T-line and C-line in the corrected card box image according to the preset proportional coordinates; S42 initially locates the initial positions of the T-line and C-line within the ROI; S43 employs a rectangle optimization algorithm based on adaptive thresholds to finely adjust the rectangular regions of the T and C lines, ensuring that the region boundaries conform to the actual line range and eliminating redundant background interference.

7. The detection method according to claim 1, characterized in that, The method for calculating the standardized grayscale ratio for background correction in step S5 is as follows: Calculate the average gray value T of the T-line region, the average gray value C of the C-line region, and the average gray value B of the background region within the ROI excluding the T-line and C-line, respectively. The T / C grayscale ratio after background correction is calculated according to the formula (BT) / (BC) and used as the standardized signal for quantitative detection.

8. The detection method according to claim 1, characterized in that, Step S5 also includes an invalid result filtering step: invalid test strip signals without valid C lines are screened by judging the grayscale difference, and invalid test results are excluded.

9. The detection method according to claim 2, characterized in that, In step S26, an independent test set is constructed for performance evaluation. The target localization and mask segmentation performance of the trained instance segmentation model is evaluated using precision, recall, and mean precision (mAP50) at an intersection-union threshold of 0.

5. The classification accuracy is verified by using a confusion matrix. The test set contains test result images of lateral flow immunochromatographic test strips under different imaging conditions.

10. The detection method according to claim 1, characterized in that, The gray-concentration calibration model mentioned in step S6 is a standard curve model pre-established for the corresponding detection items, and different detection items are configured with independent calibration models.

Images