Intelligent method and system for recognizing test strip detection result

Abstract

The application provides a test strip detection result recognition intelligent method and system, and belongs to the field of biological sample detection, which comprises the following steps: collecting the original image of a reagent box, positioning the test strip area and correcting the perspective distortion; according to the prior position interval of the quality control line C, extracting the pixel gray sequence of the strip where the C line is located, and calculating the quality control reference index; dynamically constructing the gray threshold function of the detection line T, locating the strip where the T line is located, and calculating the detection original index; calculating the normalized detection index, and obtaining the batch correction coefficient for correction; inputting into the pre-trained color development response-concentration inversion model, and outputting the concentration estimation value of serum amyloid A; comparing the concentration estimation value with the dynamic clinical positive threshold value, if the concentration estimation value is less than the dynamic clinical positive threshold value, then judging as positive, otherwise judging as negative. The application solves the problems of insufficient accuracy and consistency of test strip judgment caused by perspective distortion, fixed threshold and batch difference.

Description

Technical Field

[0001] This invention relates to the field of biological sample detection, specifically to an intelligent method and system for identifying test strip detection results. Background Technology

[0002] Serum amyloid A (SAA), as an acute-phase reactant protein, has important clinical value in the early diagnosis of infectious diseases. Currently, immunochromatographic test strips are widely used for point-of-care testing of indicators such as SAA due to their ease of use and speed. During the testing process, specialized card readers or visual interpretation of the color intensity of the control line (C-line) and the test line (T-line) on the test strip are typically required to obtain qualitative or semi-quantitative results.

[0003] Existing detection and interpretation methods have the following main drawbacks: Firstly, visual interpretation relies on the operator's visual experience and is easily affected by ambient light, observation angle, and individual color vision differences, resulting in strong subjectivity and poor repeatability. Especially in the case of a borderline positive result with weak T-line color development, it is difficult for humans to make accurate and consistent judgments. Secondly, existing card reading devices mostly use fixed grayscale thresholds to distinguish between positive and negative results, failing to fully consider the inherent batch-to-batch differences in test strips, and also failing to consider the need for dynamic threshold adjustment triggered by C-line color development fluctuations within the same batch.

[0004] In summary, how to eliminate the interference of perspective distortion on grayscale extraction, adaptively adjust the interpretation threshold to suppress the effects of batch differences and C-line color fluctuations, and achieve robust inversion from image grayscale to sample concentration has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] This invention provides an intelligent method and system for identifying test strip detection results, which solves the problems of insufficient accuracy and consistency in test strip interpretation caused by perspective distortion, fixed thresholds, and batch-to-batch differences.

[0006] In view of the above problems, the present invention provides an intelligent method and system for identifying test strip detection results.

[0007] In a first aspect, the present invention provides an intelligent method for recognizing test strip detection results, comprising: Acquire the original image of the test kit, perform test strip area localization and perspective distortion correction on the original image, and obtain the corrected image; In the corrected image, based on the prior position range of the quality control line C on the test strip, the pixel grayscale sequence of the strip where the C line is located is extracted, and the ratio of the grayscale peak value of the C line to the background grayscale value of the C line is calculated as a quality control reference index. Based on the quality control reference index, a grayscale threshold function for the detection line T is dynamically constructed, and the strip where the T line is located is located in the corrected image. The ratio of the grayscale peak value of the T line to the grayscale value of the T line background is extracted as the original detection index. Based on the original detection index and the quality control reference index, a normalized detection index is calculated, and an inter-batch correction coefficient associated with the current test strip batch number is obtained. The normalized detection index is then corrected using the inter-batch correction coefficient to obtain a corrected detection index. The modified detection index is input into the pre-trained colorimetric response-concentration inversion model, and the output is the estimated concentration of serum amyloid A in the sample. The concentration estimate is compared with the dynamic clinical positive threshold. If the concentration estimate is less than the dynamic clinical positive threshold, it is interpreted as positive; otherwise, it is interpreted as negative.

[0008] Secondly, the present invention provides an intelligent system for recognizing test strip detection results, comprising: The image acquisition and preprocessing module is used to acquire the original image of the test kit, perform test strip area localization and perspective distortion correction on the original image, and obtain the corrected image. The quality control reference index calculation module is used to extract the pixel grayscale sequence of the strip where the quality control line C is located in the corrected image, based on the prior position range of the quality control line C on the test strip, and calculate the ratio of the grayscale peak value of line C to the background grayscale value of line C as the quality control reference index. The detection original index extraction module is used to dynamically construct the grayscale threshold function of the detection line T based on the quality control reference index, locate the strip where the T line is located in the corrected image, and extract the ratio of the grayscale peak value of the T line to the grayscale value of the T line background as the detection original index. The inter-batch correction module is used to calculate the normalized detection index based on the original detection index and the quality control reference index, obtain the inter-batch correction coefficient associated with the current test strip batch number, and use the inter-batch correction coefficient to correct the normalized detection index to obtain the corrected detection index. The concentration inversion module is used to input the corrected detection index into the pre-trained colorimetric response-concentration inversion model and output the estimated concentration of serum amyloid A in the sample. The interpretation output module is used to compare the concentration estimate with the dynamic clinical positive threshold. If the concentration estimate is less than the dynamic clinical positive threshold, it is interpreted as positive; otherwise, it is interpreted as negative.

[0009] One or more technical solutions provided in this invention have at least the following technical effects or advantages: The technical solution of this invention first captures the original image by triggering a camera, then filters long line segments through edge detection, and accurately filters the target test strip area based on orthogonal parallel edges and color-marked verification points. Finally, it uses a perspective transformation matrix to remap the distorted test strip area to a standard orthophoto view, and outputs the corrected image. This eliminates geometric distortion and perspective shrinkage caused by shooting angle deviation at the physical level, ensuring the repeatability of the test from the source.

[0010] Furthermore, in the corrected image, a strip window is extracted based on the prior location interval of the C-line. The pixel grayscale sequence perpendicular to the tomography direction is extracted, the grayscale peak of the C-line is searched, and the ratio of the grayscale peak to the background grayscale of the C-line is calculated. After dual verification with the quality control effective threshold and the background noise threshold, a quality control reference index is output. A quality control benchmark independent of sample concentration and reflecting only the effectiveness of the test strip itself is established, which can automatically identify and intercept invalid test strips and background anomalies.

[0011] Furthermore, using the quality control reference index as the independent variable, the T-line recognition threshold specific to the current test strip is dynamically calculated through a pre-calibrated linear gray-scale threshold function. In the corrected image, the search interval is defined based on the prior offset distance between the T-line and C-line, the T-line gray-scale sequence is extracted and the bands are identified, and the ratio of the T-line gray-scale peak value to the T-line background gray-scale value is calculated as the original detection index. This achieves adaptive matching between the T-line recognition threshold and the actual chromatographic efficiency of each test strip, eliminating sensitivity differences between different test strips with a fixed threshold, and ensuring that the original detection index truly reflects the relative level of the analyte.

[0012] Furthermore, the initial ratio of the original detection index to the quality control reference index is calculated, and a background fluctuation ratio is introduced for normalization penalty to obtain the normalized detection index. The inter-batch correction coefficient associated with the current batch number is obtained by scanning the barcode, and the normalized detection index is corrected to output the corrected detection index. First, intra-batch individual differences are eliminated through C-line internal reference normalization, and then inter-batch process drift is eliminated through batch correction coefficients. This maps the original signals of different batches and different test strips to a unified standardized dimension, providing a stable, comparable input across batches for concentration inversion.

[0013] Furthermore, the modified detection index is input into the pre-trained colorimetric response-concentration inversion model, and through forward inference via a deep learning network, the estimated concentration of serum amyloid A in the sample is output. This completes end-to-end intelligent analysis from standardized optical signals to clinical quantitative results. The model pre-learns the mapping relationship between signals and concentrations using a large amount of standard data, ensuring the accuracy and generalization ability of the concentration estimation.

[0014] Finally, the concentration estimate is compared with a dynamic clinical positive threshold, which is dynamically generated based on the statistical distribution of negative samples and the quality control offset factor of the current test strip. The result is then output as a qualitative interpretation of positive or negative. This approach achieves personalized adaptation of interpretation criteria to baseline population characteristics and the real-time status of the test strip, effectively improving diagnostic sensitivity and specificity under different patient populations and batch conditions.

[0015] In summary, the technical solution of this invention starts with the acquisition of the original image, and then successively eliminates perspective distortion through geometric correction, ensures detection effectiveness through C-line quality control, achieves adaptive extraction of T-line through dynamic thresholding, eliminates signal drift through intra-batch normalization and inter-batch correction, inverts concentration through deep learning model, and performs personalized interpretation through dynamic clinical thresholding. This constructs a complete intelligent interpretation link from optical image to clinical qualitative result, thereby effectively solving the problem of insufficient accuracy and consistency of test strip interpretation caused by perspective distortion, fixed thresholds, and inter-batch differences. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an intelligent method for identifying test strip detection results provided by the present invention.

[0017] Figure 2 This is a schematic diagram of the calculation process for correcting the detection index in an intelligent method for identifying test strip detection results provided by the present invention.

[0018] Figure 3 This is a schematic diagram of the structure of an intelligent system for recognizing test strip detection results provided by the present invention.

[0019] In the attached diagram, the labels representing each component are as follows: Image acquisition and preprocessing module 11, quality control reference index calculation module 12, detection raw index extraction module 13, inter-batch correction module 14, concentration inversion module 15, interpretation output module 16. Detailed Implementation

[0020] This invention provides an intelligent method and system for identifying test strip detection results, addressing the problem of insufficient accuracy and consistency in test strip interpretation caused by perspective distortion, fixed thresholds, and batch-to-batch differences.

[0021] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0022] Example 1, as Figure 1As shown, the present invention provides an intelligent method for recognizing test strip detection results, the method comprising: S100: Acquire the original image of the test kit, perform test strip area localization and perspective distortion correction on the original image, and obtain the corrected image.

[0023] This step acquires the original image of the reagent kit into which the test strip is inserted, locates the band reaction area of ​​the test strip using methods such as target detection or template matching, and performs geometric correction on the area based on the perspective transformation algorithm to eliminate trapezoidal or tilt distortion caused by the shooting angle deviating from the vertical direction. Finally, a corrected image of the test strip area that is orthophoto and restored to its original size is output.

[0024] Step S100 in the method provided by the present invention includes: Based on the received image acquisition trigger signal, the original image of the test kit is acquired through the camera; Edge detection is performed on the original image, and continuous edge segments with lengths exceeding a length threshold are extracted to form an edge segment set, wherein the length threshold is determined according to a fixed proportion of the diagonal length of the original image; Select two pairs of parallel edge segments that are perpendicular to each other from the set of edge segments, and connect the two ends of each pair of parallel edge segments to form a closed quadrilateral, which serves as the set of candidate test strip regions. Locate at least four pre-placed color marker blocks on the reagent kit shell in the original image, and extract the centroid coordinates of each color marker block as a verification point; From the set of candidate test strip regions, select the candidate quadrilateral that simultaneously contains all the verification points and has the smallest average distance between the verification points and the corresponding vertices of the quadrilateral, as the target test strip region, and obtain the coordinates of the four vertices of the target test strip region; Construct a first quadrilateral using the coordinates of the four vertices, and construct a second quadrilateral using the corresponding vertex coordinates under the standard physical size of the test strip. Calculate the perspective transformation matrix from the first quadrilateral to the second quadrilateral. The pixels within the target test strip region of the original image are remapped into the second quadrilateral based on the perspective transformation matrix to generate the corrected image.

[0025] In this step, the camera is first triggered to capture the current image of the inserted test kit based on the received image acquisition trigger signal, thus obtaining the original image containing the complete area of ​​the test kit.

[0026] For example, the operator places the SAA test kit with the sample added flat in the positioning slot of the test tray and presses the physical acquisition button on the side of the device; after receiving the electrical signal triggered by the button, the 8-megapixel CMOS camera is immediately driven to take a picture and obtain an original RGB image containing the outer contour of the test kit, the test strip window and the surrounding background.

[0027] Secondly, after grayscale conversion and Gaussian filtering preprocessing of the original image, the Canny edge detection algorithm is used to extract all edge points. Chain code tracing is performed based on the eight-neighbor connectivity of the edge points, and continuous edge segments with lengths exceeding the length threshold are retained to form an edge segment set.

[0028] The length threshold is determined based on a fixed proportion of the original image's diagonal length, with a recommended range of 1 / 50 to 1 / 15. Values ​​that are too small, such as below 1 / 100, may not effectively filter out noise; values ​​that are too large, such as exceeding 1 / 5, may miss crucial short edges. A typical neutral setting is 1 / 20, which effectively filters out noise and scattered short edges.

[0029] The aforementioned grayscale conversion and Gaussian filtering preprocessing are standard front-end operations in the field of image processing for converting color images into single-channel grayscale images and using Gaussian kernels for smoothing and denoising. Canny edge detection is a classic edge extraction algorithm based on gradient magnitude and double threshold hysteresis processing. Both are basic image processing techniques commonly used in this field.

[0030] For example, after the original image acquired in the previous example was grayscaled and smoothed with a 5×5 Gaussian kernel, the Canny algorithm used high and low thresholds of 80 and 200 respectively to detect edge points. Since the image diagonal is 1200 pixels long, the length threshold was set to 1 / 20 of the image diagonal, or 60 pixels. Through connected component tracing, a total of 127 continuous edge segments with a length exceeding 60 pixels were obtained, forming a candidate edge segment set.

[0031] Next, iterate through the set of edge segments and calculate the angle of the straight line direction for each segment. Search for any pair of line segments whose angle difference is within the vertical tolerance range, such as 90°±5°, and mark them as a pair of orthogonal directions. Then find another line segment that is parallel to each of the two line segments in the orthogonal direction pair, forming two pairs of parallel edge segments with mutually perpendicular directions. Extend each pair of parallel line segments and find their intersection points, connecting them to form a closed quadrilateral. All quadrilaterals that meet the conditions constitute the set of candidate test strip regions.

[0032] For example, among the 127 edge segments, 12 horizontal segments with an angle of 0°±5° and 9 vertical segments with an angle of 90°±5° were calculated. From these, combinations containing two parallel horizontal segments and two parallel vertical segments, where the projections of the beginning and end of each segment form a closed quadrilateral, were selected, ultimately generating 5 candidate quadrilaterals. Quadrilateral A corresponds to the actual outer frame of the test strip window, quadrilaterals B and C correspond to the printed border of the reagent kit casing, and quadrilaterals D and E represent false outlines caused by environmental reflections.

[0033] Next, in the original image, pre-designed marker colors, such as circular markers of a specific blue color, are segmented based on the HSV color space. Noise is removed by morphological opening operations, and the binary connected components of each color marker block are extracted. The zeroth and first moments of the connected components are calculated to obtain the centroid coordinates of each marker block, which serve as a set of verification points.

[0034] Among them, HSV color space segmentation is a conventional method for extracting color targets in image processing by utilizing the separation characteristics of the three channels of hue, saturation, and brightness; morphological opening operation is a combination operation of erosion and dilation, which is a standard denoising method for removing small-area noise and isolated points in binary images; obtaining the region area by calculating the zero-order moment of the binary connected region, obtaining the gray-level center coordinates by calculating the first-order moment, and then normalizing to obtain the centroid of the connected region is a general technical process for extracting target location coordinates in this field.

[0035] For example, the reagent kit casing has four blue circular markers with a diameter of 3 mm printed on it, located outside the four corner points of the test strip window. After segmentation and noise reduction using the blue hue component threshold, four clear and connected white regions are obtained. The zeroth and first moments of each region are calculated, and the coordinates of the four verification points are (145, 218), (532, 221), (148, 405), and (529, 408), respectively.

[0036] Furthermore, iterate through each quadrilateral in the candidate test strip region set and determine whether its internal region contains all the verification points. For candidate quadrilaterals that meet the inclusion condition, calculate the sum of the Euclidean distances from the four verification points to the four vertices of the quadrilateral, take the quadrilateral with the smallest Euclidean distance as the target test strip region, and output the coordinates of its four vertices.

[0037] For example, point inclusion tests were performed sequentially on five candidate quadrilaterals. Quadrilateral A contained all four checkpoints, and the average distances from the four checkpoints to the four corners of quadrilateral A were 5.3 pixels, 4.8 pixels, 6.1 pixels, and 4.2 pixels, respectively, for a total distance of 20.4 pixels. Quadrilateral B contained all checkpoints, but the average total distance was 87.5 pixels. Based on the comparison, quadrilateral A was selected as the target test strip area, with the coordinates of its four vertices being (137, 210), (540, 213), (140, 412), and (538, 410).

[0038] Further, a first quadrilateral is constructed using the coordinates of the four vertices, and a second quadrilateral is constructed using the corresponding vertex coordinates under the standard physical size of the test strip. The perspective transformation matrix from the first quadrilateral to the second quadrilateral is calculated, including: The coordinates of the four vertices of the target test strip area are arranged in a clockwise direction and are respectively denoted as the first vertex, the second vertex, the third vertex, and the fourth vertex. Based on the standard physical dimensions of the test strip, the coordinates of the four corner points of the standard rectangle in the plane coordinate system are defined and denoted as the first standard point, the second standard point, the third standard point, and the fourth standard point, respectively. The plane coordinate system takes the upper left corner of the standard rectangle as the origin, the horizontal direction to the right as the positive direction of the horizontal axis, and the vertical direction downward as the positive direction of the vertical axis. Pair the first vertex with the first standard point, the second vertex with the second standard point, the third vertex with the third standard point, and the fourth vertex with the fourth standard point to establish four pairs of vertex-standard point correspondences; For each pair of correspondences, an unknown perspective transformation matrix is ​​introduced to establish a projection equation, where the homogeneous coordinates of the vertex are equal to the perspective transformation matrix multiplied by the homogeneous coordinates of the corresponding standard point. Expanding the projection equation according to the horizontal and vertical coordinate components respectively yields two independent scalar equations about the pixel coordinates. The eight independent scalar equations generated by solving the four pairs of correspondences together form a system of eight equations. Solving the system of equations yields eight degrees of freedom parameters for the perspective transformation matrix, wherein the eight degrees of freedom parameters are used to characterize the perspective projection mapping relationship from the first quadrilateral to the second quadrilateral.

[0039] Specifically, the coordinates of the four vertices of the target test strip area are first sorted in a clockwise direction and recorded as the first vertex, the second vertex, the third vertex, and the fourth vertex.

[0040] For example, the original coordinates of the four vertices of the target test strip area obtained after step S100 are A(312,189), B(718,205), C(705,508), and D(298,492). After sorting in a clockwise direction, starting from the top left corner, the coordinates of the first vertex P1 are (298,492), the second vertex P2 are (312,189), the third vertex P3 are (718,205), and the fourth vertex P4 are (705,508).

[0041] Furthermore, based on the standard physical dimensions of the test strip, the coordinates of the four corner points of the standard rectangle in the plane coordinate system are defined and denoted as the first standard point, the second standard point, the third standard point, and the fourth standard point, respectively. The plane coordinate system takes the upper left corner of the standard rectangle as the origin, the horizontal direction to the right as the positive direction of the horizontal axis, and the vertical direction downward as the positive direction of the vertical axis.

[0042] For example, the standard physical dimensions of the reaction area on an SAA test strip are 80mm wide and 25mm long. Establish a coordinate system with the top-left corner as the origin, with the horizontal axis to the right and the vertical axis downwards. Assuming an image resolution of 120 pixels / mm, the standard rectangle has a width of 9600 pixels and a length of 3000 pixels. Therefore, the first standard point Q1 has coordinates (0,0) corresponding to the top-left corner, the second standard point Q2 has coordinates (0,3000) corresponding to the bottom-left corner, the third standard point Q3 has coordinates (9600,3000) corresponding to the bottom-right corner, and the fourth standard point Q4 has coordinates (9600,0) corresponding to the top-right corner.

[0043] Furthermore, pair the first vertex with the first standard point, the second vertex with the second standard point, the third vertex with the third standard point, and the fourth vertex with the fourth standard point to establish four pairs of vertex-standard point correspondences. For example, establish four mapping pairs: (P1,Q1), (P2,Q2), (P3,Q3), and (P4,Q4).

[0044] Furthermore, for each mapping pair, an unknown 3×3 perspective transformation matrix H is introduced to establish a projection equation, where the homogeneous coordinates of the vertices on the target image are equal to the perspective transformation matrix multiplied by the homogeneous coordinates of the corresponding standard point on the source image.

[0045] Establishing a homogeneous coordinate projection mapping relationship between the source and target images using a 3×3 perspective transformation matrix is ​​a standard geometric transformation modeling method for perspective distortion correction in the field of computer vision. The 3×3 matrix contains eight independent degrees of freedom parameters, and the transformation matrix can be uniquely determined by solving a simultaneous system of homogeneous coordinate projection equations for at least four pairs of corresponding points.

[0046] For example, consider the first mapping relationship (P1, Q1). Let the perspective transformation matrix H contain 9 parameters: h11, h12, h13, h21, h22, h23, h31, h32, and h33. Since h33 is normalized to 1, the perspective transformation matrix H contains eight independent degrees of freedom, from h11 to h32. The homogeneous coordinates of Q1 are (0,0,1)ᵀ. The coordinates of P1 are (298,492), and the projection equation is: [λ·298,λ·492,λ]ᵀ=H·[0,0,1]ᵀ, where λ is the scaling factor of the homogeneous coordinates.

[0047] Furthermore, expanding the projection equations separately for the abscissa and ordinate components, and eliminating the scaling factor λ, yields two independent scalar equations concerning the pixel coordinates. For example, expanding (P1, Q1), from the homogeneous coordinate relationship, we obtain: abscissa equation: 298 = (h11·0 + h12·0 + h13) / (h31·0 + h32·0 + 1), ordinate equation: 492 = (h21·0 + h22·0 + h23) / (h31·0 + h32·0 + 1), which simplifies to: h13 = 298, h23 = 492.

[0048] Furthermore, by simultaneously solving the four pairs of correspondences, all eight independent scalar equations are generated, forming a system of eight equations. For example, the four sets of mapping pairs generate a total of eight equations: from (P1,Q1), we get h13=298, h23=492; from (P2,Q2), we get (h12·3000+h13) / (h32·3000+1)=312, (h22·3000+h23) / (h32·3000+1)=189; from (P3,Q3), we get: (h11·9600+h12·3000+h13) / ( h31·9600+h32·3000+1)=718, (h21·9600+h22·3000+h23) / (h31·9600+h32·3000+1)=205; From (P4,Q4), we get: (h11·9600+h13) / (h31·9600+1)=705, (h21·9600+h23) / (h31·9600+1)=508.

[0049] Finally, by solving the above system of equations, we obtain the eight degrees of freedom parameters h11, h12, h13, h21, h22, h23, h31, and h32 of the perspective transformation matrix H, which fully characterize the perspective projection mapping relationship from the standard rectangle, i.e., the second quadrilateral, to the target test strip area, i.e., the first quadrilateral.

[0050] For example, by solving the eight equations simultaneously using the substitution elimination method, we obtain the eight parameter values ​​of the H matrix. Let the solution be h11=0.98, h12=0.02, h13=298, h21=-0.01, h22=1.03, h23=492, h31=0.0002, h32=0.0001. The aforementioned perspective transformation matrix can accurately map the coordinates of any standard rectangle back to the corresponding position in the original image, thereby completing the perspective correction of the reaction area of ​​the SAA test strip and obtaining a standard front view without geometric distortion.

[0051] After obtaining the aforementioned perspective transformation matrix, based on the solved perspective transformation matrix H, the reverse mapping operation is performed on each integer pixel coordinate in the corrected image, i.e., the standard rectangular region defined by the second quadrilateral. After bilinear interpolation sampling, grayscale values ​​are filled to generate the final corrected image.

[0052] The specific process includes: First, a target canvas is created. Based on the standard rectangle size defined by the second quadrilateral, the width is 9600 pixels and the height is 3000 pixels. A blank image of the same size is then created as a container for the corrected image.

[0053] Furthermore, an inverse coordinate mapping is performed, traversing each pixel (x', y') in the target canvas, and multiplying its homogeneous coordinates [x', y', 1]ᵀ by the inverse matrix H of the perspective transformation matrix H. -1 This yields the corresponding subpixel coordinates (x, y) in the original image.

[0054] Furthermore, bilinear interpolation sampling is performed. Based on the calculated subpixel coordinates (x, y), the gray values ​​of four integer pixels in the neighborhood of the original image are taken, and bilinear interpolation is performed according to the distance weight to obtain the gray value estimate at that position.

[0055] Finally, grayscale filling is performed, and the interpolated grayscale value is assigned to the corresponding (x', y') position in the target canvas. After completing the filling pixel by pixel, the corrected image with perspective distortion eliminated and orthographic projection restored is obtained.

[0056] For example, continuing with the aforementioned SAA detection scenario, the perspective transformation matrix H has been obtained, whose parameters allow the coordinates of the standard rectangle to be directly mapped to the coordinates of the original image. Now, the pixel grayscale value at position (x'=4800, y'=1500) in the corrected image is generated, with the point located at the exact center of the standard rectangle.

[0057] Next, homogeneous coordinate mapping calculation is performed. The sub-pixel coordinates in the original image are calculated as (x=515.3, y=348.7) by the equation [λx,λy,λ]ᵀ=H·[4800,1500,1]ᵀ.

[0058] Finally, bilinear interpolation is performed. In the original image, the four nearest integer pixels of the subpixel coordinate and their gray values ​​are taken as follows: top left (515, 348), gray value 182; top right (516, 348), gray value 190; bottom left (515, 349), gray value 179; bottom right (516, 349), gray value 187.

[0059] The specific operation of the bilinear interpolation described above is as follows: Calculate the horizontal weight, which is the proportion of the sub-pixel coordinates in the original image within the horizontal distance between the top left and top right; calculate the vertical weight, which is the proportion of the sub-pixel coordinates in the original image within the vertical distance between the top left and bottom left. Perform two interpolations in the horizontal direction: top gray value = top left gray value × (1 - horizontal weight) + top right gray value × horizontal weight; bottom gray value = bottom left gray value × (1 - horizontal weight) + bottom right gray value × horizontal weight. Perform one interpolation in the vertical direction: final gray value = top gray value × (1 - vertical weight) + bottom gray value × vertical weight. The result is truncated to a decimal to obtain the final gray value.

[0060] For example, calculate the horizontal weight = (515.3-515) / (516-515) = 0.3, and the vertical weight = (348.7-348) / (349-348) = 0.7. Perform two interpolations in the horizontal direction to obtain the upper gray value = 182×(1-0.3)+190×0.3 = 184.4 and the lower gray value = 179×(1-0.3)+187×0.3 = 181.4. Perform one interpolation in the vertical direction to obtain the final gray value = 184.4×(1-0.7)+181.4×0.7 = 182.3. After truncating the decimal, the final gray value 182 is obtained and assigned to the position (4800, 1500) of the corrected image. The above process is repeated across all pixels of the target canvas to complete the remapping, ultimately generating an orthorectified image with a size of 9600×3000 pixels.

[0061] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0062] In summary, this step, by precisely locating the reaction area of ​​the test strip on the reagent kit and correcting for fluoroscopic distortion, effectively eliminates geometric deformation caused by shooting angle deviation, ensuring that subsequent grayscale extraction is performed under the same standard orthophoto projection, thereby improving the repeatability and spatial accuracy of the test results from the source.

[0063] S200: In the corrected image, based on the prior position range of the quality control line C on the test strip, the pixel grayscale sequence of the strip where the C line is located is extracted, and the ratio of the grayscale peak value of the C line to the background grayscale value of the C line is calculated as a quality control reference index.

[0064] This step, based on the obtained geometrically corrected image, utilizes the fixed prior position range of the C-line quality control line on the test strip during the product design phase to accurately locate the pixel row region where the C-line band is located in the corrected image. The grayscale values ​​of each column of pixels within this region are extracted along the vertical axis of the band perpendicular to the tomography direction, forming a complete C-line grayscale sequence. Peak search is performed on the C-line grayscale sequence to extract the grayscale peak at the center of the C-line band. Simultaneously, the average background grayscale is calculated from the uncolored background area near the C-line. Finally, the ratio of the C-line grayscale peak to the C-line background grayscale is used as the quality control reference index.

[0065] Step S200 in the method provided by the present invention includes: In the corrected image, a rectangular strip window covering the area of ​​the control line C on the test strip is extracted along the chromatography direction based on the prior position range of the control line C on the test strip. The average gray value of each column of pixels perpendicular to the tomography direction within the rectangular strip window is calculated, and the pixels are arranged along the tomography direction to obtain the pixel gray value sequence of the strip where the C line is located. In the pixel grayscale sequence, the point with the smallest grayscale value is searched as the C-line peak position. A preset number of sampling points are extended to both sides of the C-line peak position, and the minimum grayscale value within the covered area is taken as the C-line grayscale peak value. In the pixel grayscale sequence, a background sequence is taken from each side of the C-line peak position and within a range that is more than a preset distance from the C-line peak position. The total average grayscale value of all pixels in the two background sequences is calculated as the C-line background grayscale. Divide the peak gray level of line C by the background gray level of line C to obtain the quality control reference index.

[0066] In this step, firstly, based on the prior location range of the control line C on the test strip (i.e., the known physical distance range of the C line from the upper edge of the test strip), and the pixel coordinate range after image resolution conversion, a rectangular strip window covering the C line area is extracted along the tomographic direction. The width of this rectangular strip window covers all columns of pixels in the reaction area of ​​the test strip, and its height is slightly larger than the estimated width of the C line band to ensure complete capture of the C line color band and its adjacent upper and lower backgrounds.

[0067] The control line (C line) is a fixed band pre-coated with control antibodies on the immunochromatographic test strip, located at a specific position in the reaction area of ​​the strip. Regardless of whether the sample contains the target analyte, as long as the test strip is effective and the chromatography process is normal, the marker, such as colloidal gold particles, will be captured and colored at the C line.

[0068] For example, in the obtained 9600×3000 pixel corrected image, the prior position of line C on the SAA test strip is 18mm to 22mm from the upper edge of the strip. Converted to an image resolution of 120 pixels / mm, this range corresponds to pixel row coordinates of 2160 to 2640. A rectangular window with row coordinates [2100, 2700] is cropped along the tomographic direction, i.e., vertically downwards from the image. The window size is 600 pixels high × 9600 pixels wide, completely covering line C and the approximately 40-pixel background area above and below it.

[0069] Next, the average gray value of each column of pixels perpendicular to the tomographic direction within the rectangular strip window is calculated, and these values ​​are arranged row by row along the tomographic direction to obtain a pixel gray value sequence for the strip containing a one-dimensional C-line. The length of the pixel gray value sequence is equal to the height of the rectangular window, i.e., the number of rows, and each element in the sequence represents the average gray value of all pixels in that row.

[0070] For example, for the 600×9600 rectangular window mentioned above, we traverse each row and calculate the average grayscale value of 9600 pixels, for a total of 600 rows. For example, in the first row, the average column value corresponding to the local row coordinate 0 in the window is 198.3, in the second row it is 199.1, and in the third row it is 200.5, thus obtaining a grayscale sequence G of length 600.

[0071] Next, in the pixel grayscale sequence, the point with the smallest grayscale value is searched as the C-line peak position. With the C-line peak position as the center, a preset number of sampling points are extended to both sides, and the minimum grayscale value within the covered area is taken as the C-line grayscale peak value.

[0072] The preset number is used to extend a local search window near the peak position of the C-line to robustly determine the true grayscale peak of the band. The number of extended sampling points depends on the resolution of the corrected image and the expected physical width of the C-line band. The calculation method is: preset number = (half-width of C-line band × image resolution) / 2, which covers approximately half the width of the central region of the band.

[0073] For example, the physical half-width of line C is 0.167mm, the image resolution is 120 pixels / mm, and the preset number is (0.167×120) / 2=10 pixels, which means the preset extension sampling number is 10.

[0074] For example, in a 600-point grayscale sequence G, the minimum value is found in row 312, with a grayscale value of 82. Taking row 312 as the peak position of line C, extend 10 sampling points upwards and downwards, covering the interval [302, 322]. Within the above interval, the minimum grayscale value, i.e., the deepest color depth, is located in row 312, therefore the peak grayscale value of line C is 82.

[0075] Furthermore, in the pixel grayscale sequence, a background sequence segment is taken from each of the intervals on either side of the C-line peak position and at a distance exceeding a preset distance. The total average grayscale value of all pixels in the two background sequences is calculated as the C-line background grayscale value.

[0076] The preset distance is used to define the sampling range of the C-line background grayscale, ensuring that the selected background area falls completely outside the C-line color band and its diffusion halo, thus accurately reflecting the grayscale level of the nitrocellulose membrane background. The preset distance is set to be greater than the half-width of the C-line band plus the diffusion margin. To ensure absolute safety, a margin equal to the half-width of the band is usually reserved, i.e., the diffusion margin is 20 pixels. Therefore, the preset distance = 20 + 20 = 40 pixels, or 40 sampling points.

[0077] For example, a preset distance is set to avoid 40 sampling points on each side of the C-line peak position to ensure that the background area is not affected by the diffusion of the C-line color band. At the peak position, i.e., the upper part of row 312, 40 sampling points are taken in the interval [252, 271], with an average grayscale value of 198.5; the lower part is taken in the interval [353, 372], with an average grayscale value of 197.3. The total average grayscale value of the two background sequences, totaling 80 pixels, is (198.5 + 197.3) / 2 = 197.9. Therefore, the background grayscale value of the C-line is 197.9.

[0078] Finally, the peak grayscale value of line C is divided by the background grayscale value of line C to obtain the quality control reference index, which is used to quantitatively evaluate the effectiveness of this test. For example, if the above process yields a peak grayscale value of line C of 82 and a background grayscale value of line C of 197.9, then the quality control reference index = 82 / 197.9 ≈ 0.414.

[0079] This step, after obtaining the quality control reference index, also includes: The quality control reference index is compared with the quality control effective threshold, wherein the quality control effective threshold is determined based on the historical statistical upper limit of the quality control reference index of the same batch of test strips in multiple batches of quality verification; If the quality control reference index exceeds the quality control effective threshold, the interpretation process is terminated and an invalid reagent kit interpretation result is output. If the quality control reference index does not exceed the quality control effective threshold, then the gray standard deviation of the background sequence on both sides of the C-line peak position in the pixel gray-scale sequence is calculated, and the gray standard deviation is compared with the background noise threshold, wherein the background noise threshold is a fixed proportion of the C-line background gray level. If the grayscale standard deviation exceeds the background noise threshold, the membrane strip background is determined to be uneven, the interpretation process is terminated, and the interpretation result of invalid reagent kit is output. If the grayscale standard deviation does not exceed the background noise threshold, the quality control line is confirmed to be valid.

[0080] Specifically, after obtaining the quality control reference index, it is first compared with the preset quality control effective threshold. If the quality control reference index exceeds the quality control effective threshold, it indicates that the C-line color development is too weak, meaning the difference between the grayscale peak and the background grayscale is too small, indicating an abnormality in the test strip chromatography process or reagent failure. The interpretation process is terminated, and an invalid kit result is output. If the quality control reference index does not exceed the quality control effective threshold, it indicates that the C-line color development is strong enough, and the next step, background uniformity detection, proceeds. The quality control effective threshold is determined based on the historical statistical upper limit of the quality control reference index of the same batch of test strips in multiple batches of quality validation.

[0081] For example, in the three batches of SAA test strips tested before leaving the factory, a total of 1500 test strips were tested. The statistical upper limit of its quality control reference index is the 95th percentile, which is 0.60. Therefore, the effective quality control threshold is set to 0.60. In the previous example, the quality control reference index obtained was 0.414, which is less than the effective quality control threshold. This indicates that the C line is developing well, the termination condition was not triggered, and background noise detection continues. Conversely, if the quality control reference index of a certain test is 0.78, then since 0.78 > 0.60, it indicates that the C line is developing too weakly, and the test kit is directly output as invalid, prompting a replacement test strip for retesting.

[0082] Next, calculate the standard deviation of the grayscale values ​​of the background sequence on both sides of the C-line peak position in the pixel grayscale sequence, reflecting the dispersion of pixel grayscale values ​​within the background region. For example, in the previous step, the upper background interval was determined to be rows 252 to 271, with a total of 20 sampling points and a grayscale mean of 198.5, and the lower background interval was determined to be rows 353 to 372, with a total of 20 sampling points and a grayscale mean of 197.3. Combining the two segments into a total of 40 sampling points, the grayscale standard deviation is calculated to be 3.8.

[0083] Finally, the grayscale standard deviation is compared with the background noise threshold. If the grayscale standard deviation does not exceed the background noise threshold, the quality control line is confirmed to be valid. The background noise threshold is a fixed percentage of the background grayscale of line C, for example, 5% of the background grayscale. For example, if the background grayscale value of line C is 197.9, and the fixed percentage is set to 5%, then the background noise threshold = 197.9 × 5% = 9.895. In this test, the grayscale standard deviation of 3.8 < the threshold of 9.895, therefore the quality control line is confirmed to be valid.

[0084] If the grayscale standard deviation exceeds the background noise threshold, the membrane strip is considered to have an uneven background, indicating that the nitrocellulose membrane may have defects such as scratches, contamination, or uneven drying, affecting the accuracy of grayscale extraction. The interpretation process is then terminated, and a result indicating the kit is invalid is output. For example, if the background standard deviation of a test reaches 15.2, which is greater than the threshold of 9.895, the membrane strip is considered to have an uneven background, and a result indicating the kit is invalid with an abnormal background is directly output.

[0085] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0086] In summary, this step, by accurately locating the prior position of line C in the geometrically corrected standard image and extracting its grayscale features, calculates a quality control reference index that is unaffected by sample interference and reflects only the effectiveness of the test strip itself, effectively ensuring the reliability and traceability of a single test result.

[0087] S300: Based on the quality control reference index, dynamically construct the grayscale threshold function of the detection line T, locate the strip where the T line is located in the corrected image, and extract the ratio of the grayscale peak value of the T line to the grayscale value of the T line background as the original detection index.

[0088] This step uses the output quality control reference index as the input benchmark and dynamically constructs a T-line grayscale threshold function suitable for the current test strip based on a preset mapping relationship for accurate positioning of the T-line band. At the same time, using the prior position range of the T-line in the product design stage, the strip region where the T-line is located is located in the calibrated image, the pixel grayscale sequence perpendicular to the tomography direction is extracted, the grayscale peak of the T-line is searched, and the background grayscale of the background area adjacent to the T-line is calculated. Finally, the ratio of the grayscale peak of the T-line to the background grayscale of the T-line is used as the original detection index.

[0089] Step S300 in the method provided by the present invention includes: Obtain the corresponding data points of the quality control reference index and T-line gray peak value of the same batch of test strips under multiple concentration standards, and perform linear regression on the data points to determine the sensitivity coefficient and offset constant; Using the quality control reference index as the independent variable, a linear grayscale threshold function is constructed, wherein the grayscale threshold function is parameterized by the sensitivity coefficient and the offset constant; Substitute the quality control reference index into the grayscale threshold function to calculate the grayscale threshold of the T-line. In the corrected image, the search interval for the T-line is defined based on the prior offset distance of the detection line T-line relative to the control line C-line; Extract the average gray value of each column of pixels perpendicular to the tomography direction within the T-line search interval, and arrange them along the tomography direction to obtain the T-line pixel gray value sequence; In the T-line pixel grayscale sequence, the intervals where the grayscale value is continuously lower than the T-line grayscale threshold are identified as T-line strip candidate areas, and the interval with the smallest minimum grayscale value is selected as the strip where the T-line is located. The minimum gray value within the strip containing the T-line is extracted as the peak gray value of the T-line. A background sequence is taken from each of the regions on both sides of the strip containing the T-line and adjacent to the strip containing the T-line. The average gray value of all pixels in the two background sequences is calculated as the background gray value of the T-line. The peak gray value of the T-line is divided by the background gray value of the T-line to obtain the original detection index.

[0090] In this step, the corresponding data points of the quality control reference index and the gray peak of the T line under multiple concentration standards for the same batch of test strips are first obtained. Linear regression is performed on the above data points to determine the sensitivity coefficient and the offset constant, and a gray threshold function in linear form with the quality control reference index as the independent variable is obtained.

[0091] The test line (T-line) is a fixed band pre-coated with specific capture antibodies on the immunochromatographic test strip, located within the reaction zone and downstream of the control line (C-line). When the sample contains the target analyte, such as serum amyloid A, the analyte binds to the labeled antibody to form a complex. Under chromatography, this complex migrates to the T-line, where it is recognized and fixed by the capture antibody, forming a visible or instrumentally detectable colored band. Theoretically, the color intensity of the T-line is positively correlated with the concentration of the analyte in the sample.

[0092] For example, SAA test strips from the same batch were tested at five standard concentrations of 0 mg / L, 5 mg / L, 20 mg / L, 50 mg / L, and 100 mg / L. For each test strip, the quality control reference index was first obtained following the previous steps, and then the T-line grayscale peak value was extracted. Five sets of corresponding data points (quality control reference index, T-line grayscale peak value) were obtained: (0.41, 195), (0.42, 188), (0.43, 172), (0.41, 153), and (0.44, 135). A linear regression was performed with the quality control reference index on the horizontal axis and the T-line grayscale peak value on the vertical axis, yielding the regression equation: T-line grayscale peak value = -580 × quality control reference index + 436, which is the grayscale threshold function. The sensitivity coefficient is -580, and the offset constant is 436.

[0093] Next, the quality control reference index for this test is substituted into the grayscale threshold function to calculate the T-line grayscale threshold applicable to the current test strip. For example, the quality control reference index for this test is 0.414. Substituting it into the function, the T-line grayscale threshold is -580 × 0.414 + 436 = 195.88. After rounding, the T-line grayscale threshold is set to 196.

[0094] Next, in the corrected image, the search interval for the T-line is defined based on the prior offset distance between the detection line T-line and the control line C-line. The prior offset distance between the detection line T-line and the control line C-line is the physical distance between them, provided by the manufacturer in the product specifications, and is expressed in millimeters.

[0095] The T-line search interval is defined based on the known physical positional relationship between the T-line and C-line. A finite rectangular area is delineated as the search range for the T-line band, avoiding blind searching across the entire image. Specifically, using the row index of the C-line peak position as a reference, the center row of the T-line search interval is determined by offsetting the detection line T-line relative to the control line C-line by the prior offset distance along the tomographic direction. Then, using the center row as a reference, a certain margin is added both above and below. The margin = half-width of the C-line band × expansion ratio, with the expansion ratio set to 0.5, meaning an expansion of 50% outward from the half-width of the band as a safety margin, resulting in the start and end rows of the interval. The column range of the interval is consistent with the C-line extraction window, covering the entire width of the test strip's reaction area.

[0096] For example, the physical distance between the T line and C line on the SAA test strip is 5mm, which corresponds to 600 pixels at a resolution of 120 pixels / mm. Taking the peak position of the C line, i.e., row 312, as the baseline, offset downwards by 600 pixels to determine the center row of the T line search interval as row 912. The half-width of the C line band is 60 pixels, and the margin is calculated as half-width of the C line band × expansion ratio = 60 × 0.5 = 30 pixels. That is, the T line search interval is defined with the starting row as row 882 and the ending row as row 942, with an interval height of 60 pixels and a width covering all 9600 columns of the test strip's reaction area.

[0097] Next, the average grayscale value of each column of pixels perpendicular to the tomography direction is extracted within the T-line search interval, and arranged along the tomography direction to obtain the T-line pixel grayscale sequence. For example, for the interval of 60 rows from row 882 to 942, the average grayscale value of 9600 columns of pixels is calculated row by row. This results in a T-line grayscale sequence of length 60, where a grayscale valley region appears near row 895, with the grayscale value decreasing from 202 to 148.

[0098] Furthermore, in the T-line pixel grayscale sequence, intervals where the grayscale value continuously exceeds the T-line grayscale threshold are identified as candidate areas for the T-line strip, and the interval with the smallest minimum grayscale value is selected as the strip containing the T-line. For example, the T-line grayscale threshold is 196. Traversing the 60-point sequence, it is found that the grayscale values ​​of 11 sampling points from row 892 to 902 are continuously lower than 196, with grayscale values ​​of [195, 188, 172, 158, 148, 150, 161, 175, 189, 195, 197]. This interval is selected as the only candidate area and confirmed as the strip containing the T-line.

[0099] Finally, the minimum gray value within the T-line band is extracted as the T-line gray value peak; a background sequence is taken from each of the regions on both sides of the T-line band and adjacent to the band, and the average gray value of all pixels in the two background sequences is calculated as the T-line background gray value; the T-line gray value peak is divided by the T-line background gray value to obtain the detection raw index.

[0100] For example, the minimum grayscale value within the T-line band is 148, meaning the peak grayscale value of the T-line is 148. Four sampling points are taken from the upper side of the T-line band, in the interval [888, 891], with a mean grayscale value of 201.5; four sampling points are taken from the lower side, in the interval [903, 906], with a mean grayscale value of 200.8. Therefore, the background grayscale value of the T-line is (201.5 + 200.8) / 2 = 201.15, and the original detection index is approximately 148 / 201.15 ≈ 0.736.

[0101] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0102] In summary, this step eliminates the adaptability differences of fixed thresholds between different test strips by dynamically linking the T-line recognition threshold with the quality control reference index of each test strip. This ensures that the sensitivity of T-line positioning and grayscale extraction can be adjusted synchronously with the actual chromatography efficiency of the test strip, thereby obtaining a more accurate original detection index that reflects the relative level of the analyte.

[0103] S400: Based on the original detection index and the quality control reference index, calculate the normalized detection index, obtain the inter-batch correction coefficient associated with the current test strip batch number, and use the inter-batch correction coefficient to correct the normalized detection index to obtain the corrected detection index.

[0104] like Figure 2 As shown, this step takes the output original detection index and quality control reference index as input, and through normalization operation, normalizes the T-line signal into a normalized detection index relative to the C-line intensity to eliminate signal fluctuations caused by individual differences in gold label release amount, chromatography speed, etc. among different test strips within the same batch. Subsequently, based on the production batch number of the current test strip, the pre-calibrated inter-batch correction coefficient of the batch is obtained from the batch information database, and the normalized detection index is multiplied or linearly corrected. Finally, the corrected detection index that can be directly used for concentration inversion is output.

[0105] Step S400 in the method provided by the present invention includes: Calculate the ratio of the original detection index to the quality control reference index to obtain the initial ratio; Extract the standard deviation of grayscale values ​​of the background sequence on both sides of the peak position of line C, and calculate the ratio of the standard deviation of grayscale values ​​to the grayscale value of the background of line C as the background fluctuation ratio; Based on the initial ratio and the background fluctuation ratio, a normalized detection index is determined, wherein the normalized detection index is directly proportional to the initial ratio and inversely proportional to the background fluctuation ratio; The current test strip batch number is obtained by scanning the one-dimensional barcode attached to the test kit, and the batch-to-batch correction coefficient associated with the current test strip batch number is obtained. Wherein, the batch-to-batch correction coefficient is the ratio between the normalized detection index of the same batch of test strips measured under a known concentration standard and the measured index corresponding to the same concentration standard used when pre-training the colorimetric response-concentration inversion model. The normalized detection index is multiplied by the batch-to-batch correction coefficient to obtain the corrected detection index.

[0106] In this step, the ratio of the original detection index obtained in the previous step to the quality control reference index is first calculated as the initial ratio. For example, in the previous example, the original detection index = 0.736 and the quality control reference index = 0.414, then the initial ratio = 0.736 / 0.414 ≈ 1.778.

[0107] Secondly, extract the standard deviation of gray levels of the background sequence on both sides of the C-line peak position, and calculate the ratio of the standard deviation of gray levels to the gray level of the C-line background as the background fluctuation ratio. For example, if the standard deviation of gray levels of the background sequence on both sides of the C-line, with a total of 40 sampling points, has been calculated in the previous step, and the gray level of the C-line background is 197.9, then the background fluctuation ratio = 3.8 / 197.9 ≈ 0.0192.

[0108] Next, the normalized detection index is determined based on the initial ratio and the background fluctuation ratio. The normalized detection index is directly proportional to the initial ratio and inversely proportional to the background fluctuation ratio. That is, the greater the background noise, the more the normalized signal is suppressed to inhibit the influence of background inhomogeneity on the detection results. Specifically, the normalized detection index = initial ratio / (1 + background fluctuation ratio).

[0109] For example, based on the initial ratio of 1.778 and the background fluctuation ratio of 0.0192 calculated in the previous example, the normalized detection index is calculated as 1.778 / (1+0.0192)≈1.745, which is the standardized detection signal after batch normalization, thus reducing the uncertainty caused by background fluctuation.

[0110] Next, the batch number of the current test strip is obtained by scanning the one-dimensional barcode attached to the test kit, and the batch information database is queried to obtain the inter-batch correction coefficient associated with that batch number. The inter-batch correction coefficient is the ratio between the normalized detection index of the same batch of test strips measured with a known concentration of standard and the measured index corresponding to the same concentration of standard used in the pre-training of the colorimetric response-concentration inversion model.

[0111] For example, after obtaining the batch number through barcode scanning, a database query shows that the average normalized detection index of the current batch of SAA standard at 50 mg / L is 1.75, while the measured index of the standard used in the pre-trained model at 50 mg / L is 1.65. Therefore, the inter-batch correction coefficient = 1.65 / 1.75 ≈ 0.943.

[0112] Finally, the normalized detection index is multiplied by the batch-to-batch correction factor to obtain the corrected detection index. The corrected detection index eliminates the systematic bias introduced by batch-to-batch process differences and can be directly input into the colorimetric response-concentration inversion model for concentration estimation. For example, using the data from the previous example, the corrected detection index is calculated as 1.745 × 0.943 ≈ 1.646.

[0113] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0114] In summary, this step eliminates the interference of individual differences in test strips and membrane noise on the T-line signal by jointly normalizing the quality control reference index and the background fluctuation ratio. It also performs batch-specific correction by combining batch-specific correction coefficients to map the original signals of different batches to a unified dimension, providing a stable and standardized input that can be compared across batches for the concentration inversion model.

[0115] S500: Input the corrected detection index into the pre-trained colorimetric response-concentration inversion model and output the estimated concentration of serum amyloid A in the sample.

[0116] This step takes the modified detection index as input and feeds it into a pre-trained colorimetric response-concentration inversion model. The model uses a large number of SAA standard samples with known concentrations to pre-learn and solidify the mapping relationship between the modified detection index and SAA concentration. After receiving the modified detection index, the model performs forward inference and directly outputs the estimated concentration of serum amyloid A in the current sample.

[0117] In this step, the pre-training step of the colorimetric response-concentration inversion model includes: Multiple concentration standards were collected on the test strip to obtain corrected detection index samples, and the corresponding serum amyloid A concentration value was labeled for each corrected detection index sample to form a training sample set. Based on deep learning, an initial colorimetric response-concentration inversion model is constructed, with the corrected detection index as the model input and the serum amyloid A concentration value as the supervisory signal for the model output. The network parameters in the initial colorimetric response-concentration inversion model are iteratively optimized with the training objective of minimizing the error between the concentration estimate output by the initial colorimetric response-concentration inversion model and the supervision signal. Training is stopped once the error on the independent validation set converges to a stable value, resulting in a pre-trained colorimetric response-concentration inversion model.

[0118] Specifically, firstly, corrected detection index samples of multi-concentration standards are collected on test strips, and the corresponding serum amyloid A concentration value is labeled for each corrected detection index sample to form a training sample set.

[0119] For example, using SAA test strips from the same batch, at seven gradients of SAA standards (0 mg / L, 5 mg / L, 10 mg / L, 20 mg / L, 50 mg / L, 100 mg / L, and 200 mg / L), each concentration was tested 50 times. The corrected detection index was calculated for each sample following the previous steps, resulting in 350 corrected detection index samples. Each sample was labeled with its corresponding known standard concentration value, forming a training sample set containing 350 pairs of corrected detection index-concentration data.

[0120] Furthermore, based on a deep learning network, an initial colorimetric response-concentration inversion model is constructed, with the modified detection index as the model input and the serum amyloid A concentration value as the supervisory signal for the model output.

[0121] For example, a fully connected neural network with three hidden layers (64, 32, and 16 neurons in each layer, respectively) is constructed. The ReLU activation function is used, and the output layer is a single neuron. This network is used to regress and predict SAA concentration values. The model input is a one-dimensional modified detection index, and the output is a concentration estimate. The weights are initialized using the He initialization method, and all weights are random values ​​during the initial training phase. Therefore, the model does not yet possess accurate concentration prediction capabilities.

[0122] Furthermore, with the goal of minimizing the error between the concentration estimate output by the initial colorimetric response-concentration inversion model and the supervision signal, the network parameters in the initial colorimetric response-concentration inversion model are iteratively optimized.

[0123] For example, mean squared error (MSE) can be used as the loss function. 350 training samples are input into the model in batches. After each forward propagation to obtain a concentration estimate, the MSE compared to the true concentration value is calculated. The gradient of each layer's weights is calculated using the backpropagation algorithm. The Adam optimizer is used with a learning rate of 0.001 to update the network parameters. If, in a certain iteration, the input correction detection index is 1.646, the initial output concentration estimate is 63.2 mg / L, while the actual concentration of the labeled sample is 50 mg / L, then the squared error is calculated as (63.2 - 50). 2 =174.24. After updating the parameters via gradient backpropagation, the concentration estimate obtained in the next round with the same input was adjusted to 55.8 mg / L, and the error gradually decreased. After about 200 iterations, the model's error on the training set tended to stabilize.

[0124] Finally, training is stopped after the error on the independent validation set converges to a stable value, resulting in a pre-trained colorimetric response-concentration inversion model.

[0125] For example, a separate batch of test strips, with a different batch number than the batch used for training, is reserved. Seventy samples collected at the same seven concentration gradients serve as an independent validation set and are not used in training. During training, every 10 epochs, the current model is used to infer from the validation set, and the mean squared error of the validation set is calculated. Training stops when the validation set error no longer decreases for 20 consecutive epochs. The model parameters at this point are saved, forming the pre-trained colorimetric response-concentration inversion model.

[0126] By inputting the calculated corrected detection index into the colorimetric response-concentration inversion model, the corresponding serum amyloid A concentration value, i.e., the SAA concentration value, can be obtained. For example, by inputting a corrected detection index of 1.646, the estimated SAA concentration output by the colorimetric response-concentration inversion model is approximately 50.2 mg / L, which is highly consistent with the actual concentration of 50 mg / L.

[0127] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0128] In summary, this step utilizes a deep learning model pre-trained with a large amount of standard data to accurately map the normalized and batch-to-batch corrected standardized signal into SAA concentration values, achieving end-to-end intelligent analysis from test strip optical signals to clinical quantitative results.

[0129] S600: The concentration estimate is compared with the dynamic clinical positive threshold. If the concentration estimate is less than the dynamic clinical positive threshold, it is interpreted as positive; otherwise, it is interpreted as negative.

[0130] This step compares the estimated SAA concentration with a dynamic clinical positive threshold. If the estimated concentration is less than the dynamic threshold, it is interpreted as positive; otherwise, it is interpreted as negative. The final output is the qualitative interpretation result and the estimated concentration.

[0131] In this step, the setting of the dynamic clinical positive threshold includes: The corrected detection index of the same batch of test strips under multiple negative samples is obtained to form a negative sample index set, wherein the negative sample is a sample that has been clinically confirmed to not contain serum amyloid A; Calculate the arithmetic mean of all corrected detection indices in the negative sample index set, and denot it as the negative mean; Calculate the standard deviation of all corrected detection indices in the negative sample index set, and denot it as the negative standard deviation; Obtain multiple quality control reference indices corresponding to the negative sample index set, calculate the arithmetic mean of the multiple quality control reference indices, and denot it as the negative quality control mean; Obtain the quality control reference index obtained in the current test of the same batch of test strips, and calculate the ratio of the quality control reference index to the mean value of the negative control, as the quality control deviation factor; The dynamic clinical positive threshold is obtained by subtracting the negative mean from the negative standard deviation by a preset multiple and then combining it with the quality control offset factor.

[0132] Specifically, the corrected detection indices of the same batch of test strips for multiple negative samples are obtained to form a negative sample index set. Among them, negative samples are those confirmed by clinical gold standard methods, such as immunoturbidimetry, to be free of serum amyloid A. The arithmetic mean and standard deviation of all corrected detection indices in the negative sample index set are calculated and denoted as the negative mean and negative standard deviation, respectively.

[0133] For example, using SAA test strips from the same batch, 200 clinically confirmed negative samples were tested throughout the entire process. These negative samples were from healthy individuals undergoing physical examinations, and the SAA concentration was measured to be <2 mg / L using immunoturbidimetry. A total of 200 corrected detection indices were obtained, forming a negative sample index set. The average of the corrected detection indices of the 200 negative samples was calculated, yielding a negative mean of 2.08, and the negative standard deviation was approximately 0.12.

[0134] Furthermore, multiple quality control reference indices corresponding to the negative sample index set are obtained, and the arithmetic mean of these quality control reference indices is calculated and denoted as the negative quality control mean. For example, the above 200 negative samples each generated a quality control reference index, with an average value of 0.43. Therefore, the negative quality control mean = 0.43.

[0135] Furthermore, the quality control reference index of the same batch of test strips obtained in the current test is obtained, and the ratio of the quality control reference index to the mean of the negative control is calculated as the quality control deviation factor. For example, if the quality control reference index of the current test sample is 0.414, then the quality control deviation factor is approximately 0.414 / 0.43 ≈ 0.963.

[0136] Finally, the negative mean is subtracted from the negative standard deviation at a preset fold, and then combined with the quality control offset factor to obtain the dynamic clinical positive threshold. The preset fold determines the balance between sensitivity and specificity in the interpretation; its value is derived from statistical process control and clinical testing conventions. Generally, when the tested samples approximately follow a normal distribution, the preset fold is taken as 3 times the standard deviation. Dynamic clinical positive threshold = (negative mean - 3 × negative standard deviation) × quality control offset factor.

[0137] For example, if the preset multiple is 3 times the standard deviation, the dynamic clinical positive threshold is approximately (2.08 - 3 × 0.12) × 0.963 ≈ 1.66. The estimated SAA concentration for the current test is 22.5 mg / L, corresponding to a corrected detection index of approximately 1.65. Since the corrected detection index is 1.65 < the dynamic clinical positive threshold of 1.66, it is interpreted as positive. In another possible embodiment, if the corrected detection index for a test is 2.41, since the corrected detection index is 2.41 > the dynamic clinical positive threshold of 1.66, it is interpreted as negative.

[0138] It should be noted that the above values ​​are for illustrative purposes only and do not constitute a limitation on the present invention.

[0139] In summary, this step, by comparing individualized concentration estimates with a positive threshold dynamically generated based on the statistical distribution of negative samples and the current quality control status of the test strips, achieves adaptive matching between the interpretation criteria and the population baseline and testing conditions, effectively improving diagnostic sensitivity and specificity under different patient groups and batch conditions.

[0140] In summary, this invention constructs a complete intelligent interpretation chain from optical images to clinical qualitative results. The entire process requires no manual intervention. Using the quality control signal of each test strip as a dynamic benchmark, it eliminates various interference factors introduced by geometric, optical, process, and population baseline differences step by step, ultimately achieving high-precision, high-consistency, batch-comparable quantitative detection and individualized positive and negative interpretation of serum amyloid A.

[0141] Example 2, as Figure 3 As shown, the present invention provides an intelligent system for recognizing test strip detection results, the system comprising: The image acquisition and preprocessing module 11 is used to acquire the original image of the test kit, perform test strip area localization and perspective distortion correction on the original image, and obtain the corrected image.

[0142] The process includes acquiring the original image of the test kit, performing test strip region localization and perspective distortion correction on the original image to obtain a corrected image, including: Based on the received image acquisition trigger signal, the original image of the test kit is acquired through the camera; Edge detection is performed on the original image, and continuous edge segments with lengths exceeding a length threshold are extracted to form an edge segment set, wherein the length threshold is determined according to a fixed proportion of the diagonal length of the original image; Select two pairs of parallel edge segments that are perpendicular to each other from the set of edge segments, and connect the two ends of each pair of parallel edge segments to form a closed quadrilateral, which serves as the set of candidate test strip regions. Locate at least four pre-placed color marker blocks on the reagent kit shell in the original image, and extract the centroid coordinates of each color marker block as a verification point; From the set of candidate test strip regions, select the candidate quadrilateral that simultaneously contains all the verification points and has the smallest average distance between the verification points and the corresponding vertices of the quadrilateral, as the target test strip region, and obtain the coordinates of the four vertices of the target test strip region; Construct a first quadrilateral using the coordinates of the four vertices, and construct a second quadrilateral using the corresponding vertex coordinates under the standard physical size of the test strip. Calculate the perspective transformation matrix from the first quadrilateral to the second quadrilateral. The pixels within the target test strip region of the original image are remapped into the second quadrilateral based on the perspective transformation matrix to generate the corrected image.

[0143] Specifically, a first quadrilateral is constructed using the coordinates of its four vertices, and a second quadrilateral is constructed using the corresponding vertex coordinates under the standard physical dimensions of the test strip. The perspective transformation matrix from the first quadrilateral to the second quadrilateral is calculated, including: The coordinates of the four vertices of the target test strip area are arranged in a clockwise direction and are respectively denoted as the first vertex, the second vertex, the third vertex, and the fourth vertex. Based on the standard physical dimensions of the test strip, the coordinates of the four corner points of the standard rectangle in the plane coordinate system are defined and denoted as the first standard point, the second standard point, the third standard point, and the fourth standard point, respectively. The plane coordinate system takes the upper left corner of the standard rectangle as the origin, the horizontal direction to the right as the positive direction of the horizontal axis, and the vertical direction downward as the positive direction of the vertical axis. Pair the first vertex with the first standard point, the second vertex with the second standard point, the third vertex with the third standard point, and the fourth vertex with the fourth standard point to establish four pairs of vertex-standard point correspondences; For each pair of correspondences, an unknown perspective transformation matrix is ​​introduced to establish a projection equation, where the homogeneous coordinates of the vertex are equal to the perspective transformation matrix multiplied by the homogeneous coordinates of the corresponding standard point. Expanding the projection equation according to the horizontal and vertical coordinate components respectively yields two independent scalar equations about the pixel coordinates. The eight independent scalar equations generated by solving the four pairs of correspondences together form a system of eight equations. Solving the system of equations yields eight degrees of freedom parameters for the perspective transformation matrix, wherein the eight degrees of freedom parameters are used to characterize the perspective projection mapping relationship from the first quadrilateral to the second quadrilateral.

[0144] The quality control reference index calculation module 12 is used to extract the pixel grayscale sequence of the strip where the quality control line C is located in the corrected image based on the prior position range of the quality control line C on the test strip, and calculate the ratio of the grayscale peak value of the C line to the background grayscale value of the C line as the quality control reference index.

[0145] In the corrected image, based on the prior location range of the quality control line C on the test strip, the pixel grayscale sequence of the band containing line C is extracted, and the ratio of the grayscale peak value of line C to the background grayscale value of line C is calculated as a quality control reference index, including: In the corrected image, a rectangular strip window covering the area of ​​the control line C on the test strip is extracted along the chromatography direction based on the prior position range of the control line C on the test strip. The average gray value of each column of pixels perpendicular to the tomography direction within the rectangular strip window is calculated, and the pixels are arranged along the tomography direction to obtain the pixel gray value sequence of the strip where the C line is located. In the pixel grayscale sequence, the point with the smallest grayscale value is searched as the C-line peak position. A preset number of sampling points are extended to both sides of the C-line peak position, and the minimum grayscale value within the covered area is taken as the C-line grayscale peak value. In the pixel grayscale sequence, a background sequence is taken from each side of the C-line peak position and within a range that is more than a preset distance from the C-line peak position. The total average grayscale value of all pixels in the two background sequences is calculated as the C-line background grayscale. Divide the peak gray level of line C by the background gray level of line C to obtain the quality control reference index.

[0146] After obtaining the quality control reference index, the method further includes: The quality control reference index is compared with the quality control effective threshold, wherein the quality control effective threshold is determined based on the historical statistical upper limit of the quality control reference index of the same batch of test strips in multiple batches of quality verification; If the quality control reference index exceeds the quality control effective threshold, the interpretation process is terminated and an invalid reagent kit interpretation result is output. If the quality control reference index does not exceed the quality control effective threshold, then the gray standard deviation of the background sequence on both sides of the C-line peak position in the pixel gray-scale sequence is calculated, and the gray standard deviation is compared with the background noise threshold, wherein the background noise threshold is a fixed proportion of the C-line background gray level. If the grayscale standard deviation exceeds the background noise threshold, the membrane strip background is determined to be uneven, the interpretation process is terminated, and the interpretation result of invalid reagent kit is output. If the grayscale standard deviation does not exceed the background noise threshold, the quality control line is confirmed to be valid.

[0147] The detection original index extraction module 13 is used to dynamically construct the grayscale threshold function of the detection line T based on the quality control reference index, locate the strip where the T line is located in the corrected image, and extract the ratio of the grayscale peak value of the T line to the grayscale value of the T line background as the detection original index.

[0148] Specifically, based on the quality control reference index, a grayscale threshold function for the detection line T is dynamically constructed, and the strip containing the T line is located in the corrected image. The ratio of the T line's grayscale peak value to the T line's background grayscale value is extracted as the original detection index, including: Obtain the corresponding data points of the quality control reference index and T-line gray peak value of the same batch of test strips under multiple concentration standards, and perform linear regression on the data points to determine the sensitivity coefficient and offset constant; Using the quality control reference index as the independent variable, a linear grayscale threshold function is constructed, wherein the grayscale threshold function is parameterized by the sensitivity coefficient and the offset constant; Substitute the quality control reference index into the grayscale threshold function to calculate the grayscale threshold of the T-line. In the corrected image, the search interval for the T-line is defined based on the prior offset distance of the detection line T-line relative to the control line C-line; Extract the average gray value of each column of pixels perpendicular to the tomography direction within the T-line search interval, and arrange them along the tomography direction to obtain the T-line pixel gray value sequence; In the T-line pixel grayscale sequence, the intervals where the grayscale value is continuously lower than the T-line grayscale threshold are identified as T-line strip candidate areas, and the interval with the smallest minimum grayscale value is selected as the strip where the T-line is located. The minimum gray value within the strip containing the T-line is extracted as the peak gray value of the T-line. A background sequence is taken from each of the regions on both sides of the strip containing the T-line and adjacent to the strip containing the T-line. The average gray value of all pixels in the two background sequences is calculated as the background gray value of the T-line. The peak gray value of the T-line is divided by the background gray value of the T-line to obtain the original detection index.

[0149] The inter-batch correction module 14 is used to calculate the normalized detection index based on the original detection index and the quality control reference index, obtain the inter-batch correction coefficient associated with the current test strip batch number, and use the inter-batch correction coefficient to correct the normalized detection index to obtain the corrected detection index.

[0150] Specifically, based on the original detection index and the quality control reference index, a normalized detection index is calculated, and an inter-batch correction coefficient associated with the current test strip batch number is obtained. The normalized detection index is then corrected using the inter-batch correction coefficient to obtain a corrected detection index, including: Calculate the ratio of the original detection index to the quality control reference index to obtain the initial ratio; Extract the standard deviation of grayscale values ​​of the background sequence on both sides of the peak position of line C, and calculate the ratio of the standard deviation of grayscale values ​​to the grayscale value of the background of line C as the background fluctuation ratio; Based on the initial ratio and the background fluctuation ratio, a normalized detection index is determined, wherein the normalized detection index is directly proportional to the initial ratio and inversely proportional to the background fluctuation ratio; The current test strip batch number is obtained by scanning the one-dimensional barcode attached to the test kit, and the batch-to-batch correction coefficient associated with the current test strip batch number is obtained. Wherein, the batch-to-batch correction coefficient is the ratio between the normalized detection index of the same batch of test strips measured under a known concentration standard and the measured index corresponding to the same concentration standard used when pre-training the colorimetric response-concentration inversion model. The normalized detection index is multiplied by the batch-to-batch correction coefficient to obtain the corrected detection index.

[0151] The concentration inversion module 15 is used to input the corrected detection index into the pre-trained colorimetric response-concentration inversion model and output the estimated concentration of serum amyloid A in the sample.

[0152] The pre-training steps of the colorimetric response-concentration inversion model include: Multiple concentration standards were collected on the test strip to obtain corrected detection index samples, and the corresponding serum amyloid A concentration value was labeled for each corrected detection index sample to form a training sample set. Based on deep learning, an initial colorimetric response-concentration inversion model is constructed, with the corrected detection index as the model input and the serum amyloid A concentration value as the supervisory signal for the model output. The network parameters in the initial colorimetric response-concentration inversion model are iteratively optimized with the training objective of minimizing the error between the concentration estimate output by the initial colorimetric response-concentration inversion model and the supervision signal. Training is stopped once the error on the independent validation set converges to a stable value, resulting in a pre-trained colorimetric response-concentration inversion model.

[0153] The interpretation output module 16 is used to compare the concentration estimate with the dynamic clinical positive threshold. If the concentration estimate is less than the dynamic clinical positive threshold, it is interpreted as positive; otherwise, it is interpreted as negative.

[0154] The steps for setting the dynamic clinical positive threshold include: The corrected detection index of the same batch of test strips under multiple negative samples is obtained to form a negative sample index set, wherein the negative sample is a sample that has been clinically confirmed to not contain serum amyloid A; Calculate the arithmetic mean of all corrected detection indices in the negative sample index set, and denot it as the negative mean; Calculate the standard deviation of all corrected detection indices in the negative sample index set, and denot it as the negative standard deviation; Obtain multiple quality control reference indices corresponding to the negative sample index set, calculate the arithmetic mean of the multiple quality control reference indices, and denot it as the negative quality control mean; Obtain the quality control reference index obtained in the current test of the same batch of test strips, and calculate the ratio of the quality control reference index to the mean value of the negative control, as the quality control deviation factor; The dynamic clinical positive threshold is obtained by subtracting the negative mean from the negative standard deviation by a preset multiple and then combining it with the quality control offset factor.

[0155] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0156] This specification and accompanying drawings are merely illustrative examples of the invention and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its scope. Therefore, if such modifications and modifications fall within the scope of the invention and its equivalents, the invention is intended to include these modifications and modifications.

Claims

1. An intelligent method for recognizing test strip detection results, characterized in that, The method includes: Acquire the original image of the test kit, perform test strip area localization and perspective distortion correction on the original image, and obtain the corrected image; In the corrected image, based on the prior position range of the quality control line C on the test strip, the pixel grayscale sequence of the strip where the C line is located is extracted, and the ratio of the grayscale peak value of the C line to the background grayscale value of the C line is calculated as a quality control reference index. Based on the quality control reference index, a grayscale threshold function for the detection line T is dynamically constructed, and the strip where the T line is located is located in the corrected image. The ratio of the grayscale peak value of the T line to the grayscale value of the T line background is extracted as the original detection index. Based on the original detection index and the quality control reference index, a normalized detection index is calculated, and an inter-batch correction coefficient associated with the current test strip batch number is obtained. The normalized detection index is then corrected using the inter-batch correction coefficient to obtain a corrected detection index. The modified detection index is input into the pre-trained colorimetric response-concentration inversion model, and the output is the estimated concentration of serum amyloid A in the sample. The concentration estimate is compared with the dynamic clinical positive threshold. If the concentration estimate is less than the dynamic clinical positive threshold, it is interpreted as positive; otherwise, it is interpreted as negative.

2. The intelligent method for identifying test strip detection results according to claim 1, characterized in that, Acquire the original image of the test kit, perform test strip region localization and perspective distortion correction on the original image to obtain the corrected image, including: Based on the received image acquisition trigger signal, the original image of the test kit is acquired through the camera; Edge detection is performed on the original image, and continuous edge segments with lengths exceeding a length threshold are extracted to form an edge segment set, wherein the length threshold is determined according to a fixed proportion of the diagonal length of the original image; Select two pairs of parallel edge segments that are perpendicular to each other from the set of edge segments, and connect the two ends of each pair of parallel edge segments to form a closed quadrilateral, which serves as the set of candidate test strip regions. Locate at least four pre-placed color marker blocks on the reagent kit shell in the original image, and extract the centroid coordinates of each color marker block as a verification point; From the set of candidate test strip regions, select the candidate quadrilateral that simultaneously contains all the verification points and has the smallest average distance between the verification points and the corresponding vertices of the quadrilateral, as the target test strip region, and obtain the coordinates of the four vertices of the target test strip region; Construct a first quadrilateral using the coordinates of the four vertices, and construct a second quadrilateral using the corresponding vertex coordinates under the standard physical size of the test strip. Calculate the perspective transformation matrix from the first quadrilateral to the second quadrilateral. The pixels within the target test strip region of the original image are remapped into the second quadrilateral based on the perspective transformation matrix to generate the corrected image.

3. The intelligent method for identifying test strip detection results according to claim 2, characterized in that, Construct a first quadrilateral using the coordinates of its four vertices, and construct a second quadrilateral using the corresponding vertex coordinates under the standard physical dimensions of the test strip. Calculate the perspective transformation matrix from the first quadrilateral to the second quadrilateral, including: The coordinates of the four vertices of the target test strip area are arranged in a clockwise direction and are respectively denoted as the first vertex, the second vertex, the third vertex, and the fourth vertex. Based on the standard physical dimensions of the test strip, the coordinates of the four corner points of the standard rectangle in the plane coordinate system are defined and denoted as the first standard point, the second standard point, the third standard point, and the fourth standard point, respectively. The plane coordinate system takes the upper left corner of the standard rectangle as the origin, the horizontal direction to the right as the positive direction of the horizontal axis, and the vertical direction downward as the positive direction of the vertical axis. Pair the first vertex with the first standard point, the second vertex with the second standard point, the third vertex with the third standard point, and the fourth vertex with the fourth standard point to establish four pairs of vertex-standard point correspondences; For each pair of correspondences, an unknown perspective transformation matrix is ​​introduced to establish a projection equation, where the homogeneous coordinates of the vertex are equal to the perspective transformation matrix multiplied by the homogeneous coordinates of the corresponding standard point. Expanding the projection equation according to the horizontal and vertical coordinate components respectively yields two independent scalar equations about the pixel coordinates. The eight independent scalar equations generated by solving the four pairs of correspondences together form a system of eight equations. Solving the system of equations yields eight degrees of freedom parameters for the perspective transformation matrix, wherein the eight degrees of freedom parameters are used to characterize the perspective projection mapping relationship from the first quadrilateral to the second quadrilateral.

4. The intelligent method for identifying test strip detection results according to claim 1, characterized in that, In the corrected image, based on the prior location range of the quality control line C on the test strip, the pixel grayscale sequence of the band containing line C is extracted, and the ratio of the grayscale peak value of line C to the background grayscale value of line C is calculated as a quality control reference index, including: In the corrected image, a rectangular strip window covering the area of ​​the control line C on the test strip is extracted along the chromatography direction based on the prior position range of the control line C on the test strip. The average gray value of each column of pixels perpendicular to the tomography direction within the rectangular strip window is calculated, and the pixels are arranged along the tomography direction to obtain the pixel gray value sequence of the strip where the C line is located. In the pixel grayscale sequence, the point with the smallest grayscale value is searched as the C-line peak position. A preset number of sampling points are extended to both sides of the C-line peak position, and the minimum grayscale value within the covered area is taken as the C-line grayscale peak value. In the pixel grayscale sequence, a background sequence is taken from each side of the C-line peak position and within a range that is more than a preset distance from the C-line peak position. The total average grayscale value of all pixels in the two background sequences is calculated as the C-line background grayscale. Divide the peak gray level of line C by the background gray level of line C to obtain the quality control reference index.

5. The intelligent method for identifying test strip detection results according to claim 1, characterized in that, After obtaining the quality control reference index, the method further includes: The quality control reference index is compared with the quality control effective threshold, wherein the quality control effective threshold is determined based on the historical statistical upper limit of the quality control reference index of the same batch of test strips in multiple batches of quality verification; If the quality control reference index exceeds the quality control effective threshold, the interpretation process is terminated and an invalid reagent kit interpretation result is output. If the quality control reference index does not exceed the quality control effective threshold, then the gray standard deviation of the background sequence on both sides of the C-line peak position in the pixel gray-scale sequence is calculated, and the gray standard deviation is compared with the background noise threshold, wherein the background noise threshold is a fixed proportion of the C-line background gray level. If the grayscale standard deviation exceeds the background noise threshold, the membrane strip background is determined to be uneven, the interpretation process is terminated, and the interpretation result of invalid reagent kit is output. If the grayscale standard deviation does not exceed the background noise threshold, the quality control line is confirmed to be valid.

6. The intelligent method for identifying test strip detection results according to claim 1, characterized in that, Based on the quality control reference index, a grayscale threshold function for the detection line T is dynamically constructed, and the strip containing the T line is located in the corrected image. The ratio of the grayscale peak value of the T line to the background grayscale value of the T line is extracted as the original detection index, including: Obtain the corresponding data points of the quality control reference index and T-line gray peak value of the same batch of test strips under multiple concentration standards, and perform linear regression on the data points to determine the sensitivity coefficient and offset constant; Using the quality control reference index as the independent variable, a linear grayscale threshold function is constructed, wherein the grayscale threshold function is parameterized by the sensitivity coefficient and the offset constant; Substitute the quality control reference index into the grayscale threshold function to calculate the grayscale threshold of the T-line. In the corrected image, the search interval for the T-line is defined based on the prior offset distance of the detection line T-line relative to the control line C-line; Extract the average gray value of each column of pixels perpendicular to the tomography direction within the T-line search interval, and arrange them along the tomography direction to obtain the T-line pixel gray value sequence; In the T-line pixel grayscale sequence, the intervals where the grayscale value is continuously lower than the T-line grayscale threshold are identified as T-line strip candidate areas, and the interval with the smallest minimum grayscale value is selected as the strip where the T-line is located. The minimum gray value within the strip containing the T-line is extracted as the peak gray value of the T-line. A background sequence is taken from each of the regions on both sides of the strip containing the T-line and adjacent to the strip containing the T-line. The average gray value of all pixels in the two background sequences is calculated as the background gray value of the T-line. The peak gray value of the T-line is divided by the background gray value of the T-line to obtain the original detection index.

7. The intelligent method for identifying test strip detection results according to claim 1, characterized in that, Based on the original detection index and the quality control reference index, a normalized detection index is calculated, and an inter-batch correction coefficient associated with the current test strip batch number is obtained. The normalized detection index is then corrected using the inter-batch correction coefficient to obtain a corrected detection index, including: Calculate the ratio of the original detection index to the quality control reference index to obtain the initial ratio; Extract the standard deviation of grayscale values ​​of the background sequence on both sides of the peak position of line C, and calculate the ratio of the standard deviation of grayscale values ​​to the grayscale value of the background of line C as the background fluctuation ratio; Based on the initial ratio and the background fluctuation ratio, a normalized detection index is determined, wherein the normalized detection index is directly proportional to the initial ratio and inversely proportional to the background fluctuation ratio; The current test strip batch number is obtained by scanning the one-dimensional barcode attached to the test kit, and the batch-to-batch correction coefficient associated with the current test strip batch number is obtained. Wherein, the batch-to-batch correction coefficient is the ratio between the normalized detection index of the same batch of test strips measured under a known concentration standard and the measured index corresponding to the same concentration standard used when pre-training the colorimetric response-concentration inversion model. The normalized detection index is multiplied by the batch-to-batch correction coefficient to obtain the corrected detection index.

8. The intelligent method for identifying test strip detection results according to claim 1, characterized in that, The pre-training steps of the colorimetric response-concentration inversion model include: Multiple concentration standards were collected on the test strip to obtain corrected detection index samples, and the corresponding serum amyloid A concentration value was labeled for each corrected detection index sample to form a training sample set. Based on deep learning, an initial colorimetric response-concentration inversion model is constructed, with the corrected detection index as the model input and the serum amyloid A concentration value as the supervisory signal for the model output. The network parameters in the initial colorimetric response-concentration inversion model are iteratively optimized with the training objective of minimizing the error between the concentration estimate output by the initial colorimetric response-concentration inversion model and the supervision signal. Training is stopped once the error on the independent validation set converges to a stable value, resulting in a pre-trained colorimetric response-concentration inversion model.

9. The intelligent method for identifying test strip detection results according to claim 1, characterized in that, The steps for setting the dynamic clinical positive threshold include: The corrected detection index of the same batch of test strips under multiple negative samples is obtained to form a negative sample index set, wherein the negative sample is a sample that has been clinically confirmed to not contain serum amyloid A; Calculate the arithmetic mean of all corrected detection indices in the negative sample index set, and denot it as the negative mean; Calculate the standard deviation of all corrected detection indices in the negative sample index set, and denot it as the negative standard deviation; Obtain multiple quality control reference indices corresponding to the negative sample index set, calculate the arithmetic mean of the multiple quality control reference indices, and denot it as the negative quality control mean; Obtain the quality control reference index obtained in the current test of the same batch of test strips, and calculate the ratio of the quality control reference index to the mean value of the negative control, as the quality control deviation factor; The dynamic clinical positive threshold is obtained by subtracting the negative mean from the negative standard deviation by a preset multiple and then combining it with the quality control offset factor.

10. An intelligent system for recognizing test strip detection results, characterized in that, The system is used to perform the intelligent method for identifying test strip detection results according to any one of claims 1 to 9, the system comprising: The image acquisition and preprocessing module is used to acquire the original image of the test kit, perform test strip area localization and perspective distortion correction on the original image, and obtain the corrected image. The quality control reference index calculation module is used to extract the pixel grayscale sequence of the strip where the quality control line C is located in the corrected image, based on the prior position range of the quality control line C on the test strip, and calculate the ratio of the grayscale peak value of line C to the background grayscale value of line C as the quality control reference index. The detection original index extraction module is used to dynamically construct the grayscale threshold function of the detection line T based on the quality control reference index, locate the strip where the T line is located in the corrected image, and extract the ratio of the grayscale peak value of the T line to the grayscale value of the T line background as the detection original index. The inter-batch correction module is used to calculate the normalized detection index based on the original detection index and the quality control reference index, obtain the inter-batch correction coefficient associated with the current test strip batch number, and use the inter-batch correction coefficient to correct the normalized detection index to obtain the corrected detection index. The concentration inversion module is used to input the corrected detection index into the pre-trained colorimetric response-concentration inversion model and output the estimated concentration of serum amyloid A in the sample. The interpretation output module is used to compare the concentration estimate with the dynamic clinical positive threshold. If the concentration estimate is less than the dynamic clinical positive threshold, it is interpreted as positive; otherwise, it is interpreted as negative.

Images