Hemoglobin content analysis method and system based on microscopic magnification digital image

Abstract

The application discloses a hemoglobin content analysis method and system based on microscopic magnification digital images, and relates to the technical field of hemoglobin detection.The application collects multi-band images of labeled samples through a multi-band microscope imaging module, completes denoising, illumination correction and image alignment through an image acquisition and preprocessing module, extracts morphological features and optical features of red blood cells through a feature extraction module, associates feature matrices with real contents through a model training and prediction module, constructs mapping relationships between each feature and hemoglobin content, and finally presents data through a result output and verification module.The application adopts a statistical method of morphological-optical combined features, fuses red blood cell morphology and optical properties, realizes accurate detection of hemoglobin content, adapts to complex samples such as slight cell overlap and abnormal morphology, captures complex correlations between features and content, has higher precision and stronger anti-interference ability compared with single methods, and can provide more reliable hemoglobin detection data for clinics.

Description

Technical Field

[0001] This invention relates to the field of hemoglobin detection technology, and in particular to a method and system for analyzing hemoglobin content based on magnified digital microscopic images. Background Technology

[0002] With the development of digital image processing and microscopic imaging technologies, hemoglobin detection methods based on magnified digital microscopic images have gradually attracted attention. However, existing technologies mostly use Beer-Lambert's law to calculate hemoglobin content. Chinese patent application CN114778418B discloses a hemoglobin analysis method based on magnified digital microscopic images. By combining Beer-Lambert's law with magnified digital microscopic images, the clinical perspective on hemoglobin content has been broadened from the whole sample level to the level of individual cells, enabling statistical analysis of individual hemoglobin levels and providing more valuable information for clinical practice. Furthermore, the entire measurement system is extremely simple, with maintenance-free optical and liquid paths, and simplified operation and control processes, significantly improving the overall efficiency of hemoglobin detection.

[0003] While the aforementioned patents have improved the overall efficiency of hemoglobin detection, the following problems still exist: Existing technologies rely solely on Beer-Lambert's law calculations based on optical characteristics for detection, neglecting the intrinsic relationship between red blood cell morphology and hemoglobin content. For example, red blood cells in patients with iron deficiency anemia exhibit a small cell and low pigmentation morphology, and their morphological characteristics and optical characteristics undergo synergistic changes. Single-feature detection is easily affected by cell overlap, morphological abnormalities, and other interferences, resulting in insufficient detection accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for analyzing hemoglobin content based on magnified digital microscopic images. It adopts a statistical method combining morphological and optical features, integrating red blood cell morphology and optical properties to achieve accurate detection of hemoglobin content. By fusing multiple features, it improves detection accuracy and anti-interference ability, adapts to the detection of complex samples, simplifies the operation process, and improves detection efficiency, thereby solving the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The method for analyzing hemoglobin content based on magnified digital microscopic images includes the following steps: Image acquisition: A magnified digital image of the labeled sample is acquired using a multi-band imaging device, and the acquired magnified digital image is preprocessed. The labeled sample is a blood sample with a predetermined true hemoglobin content, and the magnified digital image includes no less than three different characteristic bands. Multi-feature extraction: Identify red blood cell regions from the preprocessed magnified digital microscopic image, screen the identified red blood cell regions for target red blood cells, output the target red blood cell region image, and simultaneously extract the morphological and optical features corresponding to each target red blood cell to form a feature matrix; Model construction: Construct a mapping model between the feature matrix and the actual hemoglobin content of the labeled sample. Based on the mapping model, calculate the feature matrix of the sample to be tested to obtain the predicted value of the hemoglobin content of the sample to be tested. Prediction and Validation: Output the predicted value of hemoglobin content of the sample to be tested. At the same time, compare the predicted value with the measured true value to verify and calculate the error rate.

[0006] Furthermore, the three different characteristic wavelength bands in the image acquisition are 380-440nm, 530-550nm, and 580-600nm, respectively. The multi-band imaging device includes a switchable LED light source, a high-magnification optical microscope, and an image sensor. The light intensity stability error of the LED light source in each characteristic wavelength band does not exceed ±5%, and the image acquisition frame rate is not less than 15fps.

[0007] Furthermore, the acquired magnified digital microscopic images undergo preprocessing, specifically including: Image denoising: The acquired multi-band magnified digital images are smoothed to remove Gaussian noise and eliminate salt-and-pepper noise. Illumination correction: Adjusting the grayscale distribution of the denoised magnified digital image to correct grayscale value deviations caused by local brightness deviations; Image alignment: Spatial position calibration of digital images of each band is performed using fixed marker points in the magnified microscopic images as a reference.

[0008] Further, multi-feature extraction includes: Red blood cell recognition: Based on a pre-trained red blood cell recognition model, the target red blood cells in the pre-processed magnified digital microscopic image are identified, and the target red blood cell bounding box is generated. At the same time, the interference of non-target red blood cells and impurities is eliminated, and the target red blood cell region image containing only the red blood cell region is output. Morphological feature extraction: Calculate the morphological features of each target red blood cell based on the target red blood cell bounding box information; Optical feature extraction: Based on the magnified digital images of each band after alignment, the optical features of the target red blood cells are calculated.

[0009] Further, morphological feature extraction specifically includes: Area: The number of pixels in the target red blood cell region image within the target red blood cell bounding box is obtained, and the actual area of ​​the target red blood cell is calculated based on the magnification of the microscope objective. Equivalent diameter: Based on the area formula and calculation, the equivalent diameter of the target red blood cell is obtained. The area formula is as follows:

[0010] Where S is the actual area of ​​the target red blood cell, and d is the equivalent diameter; Roundness: The roundness is calculated by obtaining the perimeter of the target red blood cell through an edge detection algorithm and combining it with the actual area of ​​the target red blood cell, as shown in the following formula:

[0011] Where C is the roundness value and L is the circumference of the target red blood cell. The closer the roundness value is to 1, the closer the red blood cell shape is to the standard round shape. Central light-stained area proportion: Determine the grayscale thresholds of the central light-stained area and the edge dark area of ​​the target red blood cell, count the number of pixels in the central light-stained area and the entire red blood cell area respectively, and calculate the proportion of the central light-stained area.

[0012] Further, optical feature extraction specifically includes: Average gray value: Traverse all pixels within each red blood cell region, count the pixel gray values ​​in each band, and obtain the average gray value of each target red blood cell in the corresponding band; Gray-scale standard deviation: The degree of deviation between the gray-scale values ​​of all pixels within the target red blood cell region and the average gray-scale value is calculated based on the average gray-scale value of each band, and is used as the gray-scale standard deviation; Absorbance: In the magnified digital images of each wavelength band, the average gray value of the cell-free blank area is extracted as the gray value corresponding to the incident light intensity. The average gray value of the target red blood cells in the same wavelength band is retrieved as the gray value corresponding to the transmitted light intensity. The absorbance of the target red blood cells in each wavelength band is then calculated.

[0013] Further predictions and verifications include: Predicted value calculation: The extracted target red blood cell morphological features and optical features of each band are integrated to form a feature matrix, which is then input into a pre-constructed mapping relationship model. The mapping relationship model outputs the predicted value of hemoglobin content of the sample to be tested based on the correlation logic between the feature matrix and the hemoglobin content. Accurate value acquisition: The ferric cyanide method is used to simultaneously detect the hemoglobin content of the test samples to obtain the accurate value of the hemoglobin content of the test samples as a verification benchmark; Error calculation and judgment: Calculate the relative error rate between the predicted value of hemoglobin content and the actual value of hemoglobin content. If the relative error rate is higher than the preset error threshold, trigger the anomaly investigation process. After eliminating operational errors, if the error still exceeds the standard, supplement the labeled data corresponding to the sample to be tested into the model training set and re-optimize the mapping relationship model. Abnormal sample adaptation: For complex samples with slight cell overlap and morphological abnormalities, the contours of individual cells are separated by edge detection algorithm and then morphological features are extracted. The morphological and optical features of morphologically abnormal red blood cells are recorded separately and assigned corresponding feature weights when predicting in the mapping relationship model.

[0014] This invention provides another technical solution: a hemoglobin content analysis system based on magnified digital microscopic images, comprising: The multi-band microscope imaging module is used to acquire magnified digital images of labeled samples and test samples according to preset image acquisition parameters; The image acquisition and preprocessing module is used to acquire magnified digital images and sequentially perform noise reduction, illumination correction and multi-band image alignment operations. The feature extraction module is used to identify red blood cells in the preprocessed magnified digital microscopic image, filter out complete target red blood cells and output the target region image, and extract the morphological and optical features of each target red blood cell, integrating them to form a feature matrix. The model training and prediction module is used to perform statistical correlation analysis between the feature matrix and the actual hemoglobin content based on the statistical method of morphological-optical joint features, construct a mapping relationship model, input the feature matrix corresponding to the sample to be tested into the mapping relationship model, and calculate and output the predicted value of hemoglobin content. The results output and verification module is used to determine the true hemoglobin content of the sample to be tested and to calculate the relative error rate between the predicted value and the true value.

[0015] Furthermore, the morphological and optical features corresponding to each target red blood cell are extracted to form a feature matrix, including: Feature extraction is performed on each target red blood cell to obtain the corresponding morphological and optical features of each target red blood cell; An original feature matrix is ​​constructed based on the morphological and optical features corresponding to each target red blood cell;

[0016] in, express ×7-dimensional original feature matrix; Indicates the first The area of ​​each target red blood cell; Indicates the first The equivalent diameter of a target red blood cell; Indicates the first The roundness of each target red blood cell; Indicates the first The percentage of lightly stained central regions in each target red blood cell; Indicates the first The average gray value of each target red blood cell; Indicates the first The standard deviation of grayscale values ​​for each target red blood cell; Indicates the first The absorbance of each target red blood cell; Based on a clinical normal red blood cell sample database, the statistical distribution of each feature is calculated to generate a conventional baseline vector; the conventional baseline vector is then expanded to generate a feature conventional baseline matrix. Calculate the deviation of each feature in each original feature matrix from the conventional benchmark;

[0017] in, Represents the original feature matrix. The first target red blood cell Deviation of each feature; Represents the original feature matrix. The first target red blood cell The original values ​​of each feature; Indicates the first The mean of each feature in the feature standard reference matrix; Indicates the first The standard deviation of each feature in the feature benchmark matrix; Indicates the dynamic decay factor; Indicates the first The weights of each feature; The original feature matrix is ​​corrected based on the deviation to obtain the target feature matrix;

[0018] in, Represents the original feature matrix. The first target red blood cell The corrected values ​​of each feature; This represents the first number of normal red blood cells. The standard baseline mean of each feature.

[0019] Furthermore, the methods for constructing the mapping relationship model include: Calculate the Pearson correlation coefficient between each feature in the feature matrix and the hemoglobin content, and filter the feature matrix based on the correlation coefficient to obtain the key feature matrix; Based on the key feature matrix, a feature-hemoglobin dynamic correlation function is constructed, and each feature in the key feature matrix is ​​adaptively adjusted to obtain an enhanced feature matrix. Physical prior equations are constructed based on clinical physiological principles; these physical prior equations include the Lambert-Beer law extension equation, the central light stained area constraint equation, and the morphological-optical co-constraint equation. A physical information neural network is constructed, comprising an input layer, a first hidden layer, a second hidden layer, a third hidden layer, a physical constraint layer, and an output layer. The enhanced feature matrix is ​​input into the physical information neural network, which outputs a physical mapping feature vector. The physical residual loss and feature reconstruction loss of the physical constraint equation are used as optimization objectives to pre-train the physical information neural network and obtain a stable physical mapping feature vector. Using the stable physical mapping feature vector as input, the training is performed through the deep learning regression branch, the physical analysis branch and the feature correction branch respectively, to obtain the output value of each branch and the branch loss corresponding to each branch; The sample-level confidence of each branch is calculated based on the branch loss corresponding to the deep learning regression branch, physical analysis branch, and feature correction branch. Based on the output values ​​and confidence levels of each branch, the final predicted hemoglobin value is determined, and a mapping relationship model is obtained.

[0020] Compared with the prior art, the beneficial effects of the present invention are: This statistical method, based on morphological-optical combined features, acquires multi-band images of labeled samples using a multi-band microscope imaging module. The image acquisition and preprocessing module performs denoising, illumination correction, and image alignment. The feature extraction module then extracts the morphological and optical features of red blood cells. Finally, the model training and prediction module correlates the feature matrix with the actual content, constructing a mapping relationship between each feature and hemoglobin content. The data is then presented through the results output and verification module. By fusing multiple features, this method reduces errors from single features, adapts to complex samples with slight cell overlap and morphological abnormalities, and captures the complex correlation between features and content. Compared to single methods, it offers higher accuracy and stronger anti-interference capabilities, providing more reliable hemoglobin detection data for clinical use. Attached Figure Description

[0021] Figure 1 This is a flowchart of the hemoglobin content analysis method based on magnified digital micrographs of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] To address the technical issues of existing technologies that rely solely on Beer-Lambert's law calculations based on optical features, neglecting the intrinsic relationship between red blood cell morphology and hemoglobin content, and the susceptibility of single-feature detection to interference from cell overlap and morphological abnormalities, leading to insufficient detection accuracy, please refer to [link to relevant documentation]. Figure 1 This embodiment provides the following technical solution: The method for analyzing hemoglobin content based on magnified digital microscopic images includes the following steps: Image acquisition: A magnified digital image of the labeled sample is acquired using a multi-band imaging device, and the acquired magnified digital image is preprocessed, including noise reduction, illumination correction and multi-band image alignment. The labeled sample is a blood sample whose true hemoglobin content has been pre-determined by the ferric cyanide method, and the magnified digital image includes no less than 3 different characteristic bands. In this embodiment, the three different characteristic wavelength bands in image acquisition are 380-440nm, 530-550nm, and 580-600nm. The multi-band imaging device includes a switchable LED light source, a high-magnification optical microscope, and a CMOS image sensor with a resolution of no less than 2 million pixels. The light intensity stability error of the LED light source in each characteristic wavelength band does not exceed ±5%, and the image acquisition frame rate is no less than 15fps, ensuring that the acquired multi-band images have consistent brightness, clear images, and no obvious motion blur. In this embodiment, the 380-440nm band is used for hemoglobin-specific absorption, the 530-550nm band is the traditional Hb derivative band for cross-validation, and the 580-600nm band helps exclude non-erythrocyte regions. Simultaneously, the differences in the response of different bands to cell structures are utilized to assist in extracting morphological features, improving cell recognition accuracy, and reducing interference from non-target substances. For example, hemoglobin content is not only related to the 418nm absorbance but also to the "percentage of the central lightly stained area" of the cell (reflecting biconcave morphology), and errors can be reduced through feature fusion. Multi-feature extraction: Identify red blood cell regions from the preprocessed magnified digital microscopic image, screen the identified red blood cell regions for target red blood cells, output the target red blood cell region image, and simultaneously extract the morphological and optical features corresponding to each target red blood cell to form a feature matrix; Model construction: Construct a mapping model between the feature matrix and the actual hemoglobin content of the labeled sample. Based on the mapping model, calculate the feature matrix of the sample to be tested to obtain the predicted value of the hemoglobin content of the sample to be tested. Prediction and Validation: Output the predicted value of hemoglobin content of the sample to be tested. At the same time, compare the predicted value with the measured true value to verify and calculate the error rate.

[0024] In this embodiment, a statistical method based on morphological-optical joint features is used to identify red blood cell regions from preprocessed magnified digital microscopic images. Target cells are screened from the identified red blood cell regions, and images of the target red blood cell regions are output. Simultaneously, the morphological and optical features corresponding to each target red blood cell are extracted. The morphological features reflecting the physical structure of red blood cells and the optical features reflecting the optical properties of hemoglobin are jointly analyzed and statistically integrated to form a feature matrix. This achieves a synergistic correlation between morphological and optical features, annotating the correlation between the two features of the sample and the actual hemoglobin content, thereby reducing errors from a single feature and adapting to complex samples such as those with slight cell overlap and morphological abnormalities.

[0025] In this embodiment, the acquired magnified digital microscopic images are preprocessed, specifically including: Image denoising: The acquired multi-band magnified digital images were smoothed using Gaussian filtering (standard deviation set to 1.0) to remove Gaussian noise from the magnified digital images, and median filtering with a 3×3 window was used to eliminate salt-and-pepper noise from the magnified digital images. Illumination correction: An image grayscale histogram equalization algorithm is used to adjust the grayscale distribution of the denoised magnified digital image and correct the grayscale value deviation caused by local brightness deviation. Image alignment: Using fixed markers in the magnified digital images as a reference, such as tiny scales on the sample slide or specific impurity particles, the spatial position of the magnified digital images of each band is calibrated by a feature point matching algorithm.

[0026] In this embodiment, image denoising preserves image edge information while effectively suppressing noise. Since microscope illumination may be uneven, leading to significant brightness differences in different areas of the image, illumination correction ensures uniform overall image brightness, preventing the impact of uneven illumination on subsequent feature extraction. During multi-band image acquisition, slight sample movement and minor positional deviations caused by light source switching may result in spatial misalignment between different band images. Precise alignment of multi-band images ensures that the position of the same red blood cell remains consistent across different band images.

[0027] In this embodiment, multi-feature extraction includes: Red blood cell recognition: Based on a pre-trained red blood cell recognition model, the target red blood cells in the pre-processed magnified digital microscopic image are identified and the target red blood cell bounding box is generated. At the same time, the interference of non-target red blood cells such as white blood cells and platelets and impurities is excluded, and the target red blood cell region image containing only the red blood cell region is output. Morphological feature extraction: Based on the bounding box information of the target red blood cells, the morphological features of each target red blood cell are calculated, specifically including: Area: The number of pixels in the target red blood cell region image within the target red blood cell bounding box is obtained, and the actual area of ​​the target red blood cell is calculated based on the magnification of the microscope objective. In this embodiment, the microscope objective is a 100x objective lens, and the pixel-to-actual-size conversion ratio is 0.1 μm / pixel. (Unit: μm) 2 ; Equivalent diameter: Based on the area formula and calculation, the equivalent diameter of the target red blood cell is obtained, in μm. The area formula is as follows:

[0028] Where S is the actual area of ​​the target red blood cell, and d is the equivalent diameter; In this embodiment, the equivalent diameter is calculated assuming the target red blood cells are circular; Roundness: The roundness is calculated by obtaining the perimeter of the target red blood cell through an edge detection algorithm (such as Canny edge detection) and combining it with the actual area of ​​the target red blood cell, as shown in the following formula:

[0029] Where C is the roundness value and L is the circumference of the target red blood cell. The closer the roundness value is to 1, the closer the red blood cell shape is to the standard round shape. Central light-stained area proportion: The Otsu threshold segmentation algorithm is used to determine the grayscale threshold of the central light-stained area and the edge dark area of ​​the target red blood cell. The number of pixels in the central light-stained area and the entire red blood cell area are counted respectively, and the proportion of the central light-stained area is calculated. This feature can reflect the distribution of hemoglobin in red blood cells. Optical feature extraction: Based on the magnified digital images of the target red blood cells after alignment of each band, the optical features are calculated, specifically including: Average gray value: Traverse all pixels in each red blood cell region, count the pixel gray values ​​in each band, and obtain the average gray value of each target red blood cell in the corresponding band. This value reflects the degree of light transmission of the red blood cell in that band and is negatively correlated with the hemoglobin concentration. Gray standard deviation: Based on the average gray value of each band, the deviation of the gray value of all pixels in the target red blood cell area from the average gray value is calculated as the gray standard deviation. It can reflect the uniformity of gray distribution in the red blood cell area. The smaller the gray standard deviation, the more uniform the distribution of hemoglobin in the red blood cell. Conversely, the greater the distribution difference, the smaller the gray standard deviation. It can help judge the influence of red blood cell morphological integrity on hemoglobin content. Absorbance: In the magnified digital images of each wavelength band, the average gray value of the cell-free blank area is extracted as the gray value corresponding to the incident light intensity. The average gray value of the target red blood cells in the same wavelength band is retrieved as the gray value corresponding to the transmitted light intensity. The absorbance of the target red blood cells in each wavelength band is then calculated.

[0030] In this embodiment, a pre-trained model is used to accurately generate red blood cell bounding boxes, eliminating non-target interference and avoiding misjudgment of white blood cells and impurities, thus improving the accuracy of red blood cell recognition. Combining morphological feature extraction and optical feature extraction, the differences in the physical structure of red blood cells are quantified, which is intuitively reflected by the reduction in area and the increase in the proportion of the central light-stained area. This provides morphological dimension support for the correlation of hemoglobin content, correlates the light absorption characteristics of hemoglobin, and cross-validates multi-band data to reduce interference from light source fluctuations and differences in slide light transmission, thereby achieving accurate mapping between optical properties and hemoglobin concentration.

[0031] In this embodiment, the pre-trained red blood cell recognition model is trained by collecting more than 5,000 multi-band microscopic images covering different blood samples, labeling red blood cell regions, and forming a model training set containing bounding box coordinates and cell morphology labels. At the same time, interference samples such as white blood cells, platelets, and impurities are mixed in to enhance the model's anti-interference ability. The attention mechanism is combined to strengthen the extraction of red blood cell contour and texture features and suppress background noise.

[0032] In this embodiment, prediction and verification include: Predicted value calculation: The extracted target red blood cell morphological features and optical features of each band are integrated to form a feature matrix, which is then input into a pre-constructed mapping relationship model. The mapping relationship model outputs the predicted value of hemoglobin content of the sample to be tested based on the correlation logic between the feature matrix and the hemoglobin content. Accurate value acquisition: The ferric cyanide method was used to simultaneously detect the hemoglobin content of the test samples and obtain the accurate value of the hemoglobin content of the test samples as a verification benchmark; Error calculation and judgment: Calculate the relative error rate between the predicted value of hemoglobin content and the actual value of hemoglobin content. If the relative error rate is higher than the preset error threshold, trigger the anomaly investigation process. After eliminating operational errors, if the error still exceeds the standard, supplement the labeled data corresponding to the sample to be tested into the model training set and re-optimize the mapping relationship model. Abnormal Sample Adaptation: For complex samples with slight cell overlap (overlap area ≤30%) and morphological abnormalities (such as acanthocytic red blood cells and target red blood cells), the contours of individual cells are separated by edge detection algorithm and then morphological features are extracted. The morphological and optical features of morphologically abnormal red blood cells are recorded separately and assigned corresponding feature weights when predicting in the mapping relationship model. This ensures that the prediction error of hemoglobin content for complex samples is still controlled within ±3%, meeting the diverse clinical testing needs.

[0033] In this embodiment, by integrating the morphology-optics joint feature matrix input mapping relationship model, the synergistic correlation between the physical structure and optical properties of red blood cells is fully explored, effectively improving the correlation between the predicted and actual values ​​of hemoglobin content, reducing the basic prediction error, and quickly locating the source of error through an anomaly investigation process. Supplementing labeled data to optimize the model solves the prediction drift caused by batch differences, resulting in a lower error over-standard rate of the iterative model. Furthermore, it breaks through the dependence of traditional models on standard morphological red blood cells, reducing the prediction error of complex samples, significantly improving the detection accuracy, and meeting the diverse sample detection needs of clinical practice.

[0034] In this embodiment, the morphological and optical features corresponding to each target red blood cell are extracted to form a feature matrix, including: Feature extraction is performed on each target red blood cell to obtain the corresponding morphological and optical features of each target red blood cell; An original feature matrix is ​​constructed based on the morphological and optical features corresponding to each target red blood cell;

[0035] in, express ×7-dimensional original feature matrix; Indicates the first The area of ​​each target red blood cell; Indicates the first The equivalent diameter of a target red blood cell; Indicates the first The roundness of each target red blood cell; Indicates the first The percentage of lightly stained central areas in each target red blood cell; Indicates the first The average gray value of each target red blood cell; Indicates the first The standard deviation of grayscale values ​​for each target red blood cell; Indicates the first The absorbance of each target red blood cell; Based on a clinical normal red blood cell sample database, the statistical distribution of each feature is calculated to generate a conventional baseline vector; the conventional baseline vector is then expanded to generate a feature conventional baseline matrix. Calculate the deviation of each feature in each original feature matrix from the conventional benchmark;

[0036] in, Represents the original feature matrix. The first target red blood cell Deviation of each feature; Represents the original feature matrix. The first target red blood cell The original values ​​of each feature; Indicates the first The mean of each feature in the feature standard reference matrix; Indicates the first The standard deviation of each feature in the feature benchmark matrix; Indicates the dynamic decay factor; Indicates the first The weights of each feature; The original feature matrix is ​​corrected based on the deviation to obtain the target feature matrix;

[0037] in, Represents the original feature matrix. The first target red blood cell The corrected values ​​of each feature; This represents the first number of normal red blood cells. The standard baseline mean of each feature.

[0038] In this embodiment, the feature conventional reference matrix is: ×7 dimensions; used to represent the characteristic statistical benchmarks of normal red blood cells; each row element is a 7-dimensional conventional benchmark vector. ;in, This represents the mean area of ​​normal red blood cells; This represents the mean of the equivalent diameter of normal red blood cells; This represents the mean roundness of normal red blood cells; This represents the average percentage of the lightly stained central region of normal red blood cells; This represents the mean of the average gray values ​​of normal red blood cells. This represents the mean of the standard deviation of grayscale values ​​for normal red blood cells. This represents the average absorbance of normal red blood cells.

[0039] The working principle and beneficial effects of the above technical solution are as follows: By calculating the deviation of features from conventional benchmarks and correcting the original feature matrix, feature bias caused by individual differences, measurement errors, and other factors can be reduced, making the features more accurately reflect the true state of red blood cells. In clinical diagnosis, accurate red blood cell features are crucial for disease diagnosis and monitoring. The corrected feature matrix can provide doctors with more reliable diagnostic evidence, improving the accuracy and reliability of diagnosis. The introduction of a dynamic decay factor allows the deviation calculation to be adaptively adjusted according to the degree of difference between the feature values ​​and conventional benchmarks, avoiding over-correction or under-correction. Through comparison and correction with conventional benchmarks, features between different samples have better comparability, which helps to standardize features and facilitates subsequent data analysis and model training.

[0040] In this embodiment, the method for constructing the mapping relationship model includes: Calculate the Pearson correlation coefficient between each feature in the feature matrix and the hemoglobin content, and filter the feature matrix based on the correlation coefficient to obtain the key feature matrix; Based on the key feature matrix, a feature-hemoglobin dynamic correlation function is constructed, and each feature in the key feature matrix is ​​adaptively adjusted to obtain an enhanced feature matrix. Physical prior equations are constructed based on clinical physiological principles; these physical prior equations include the Lambert-Beer law extension equation, the central light stained area constraint equation, and the morphological-optical co-constraint equation. A physical information neural network is constructed, comprising an input layer, a first hidden layer, a second hidden layer, a third hidden layer, a physical constraint layer, and an output layer. The enhanced feature matrix is ​​input into the physical information neural network, which outputs a physical mapping feature vector. The physical residual loss and feature reconstruction loss of the physical constraint equation are used as optimization objectives to pre-train the physical information neural network and obtain a stable physical mapping feature vector. Using the stable physical mapping feature vector as input, the training is performed through the deep learning regression branch, the physical analysis branch and the feature correction branch respectively, to obtain the output value of each branch and the branch loss corresponding to each branch; The sample-level confidence of each branch is calculated based on the branch loss corresponding to the deep learning regression branch, physical analysis branch, and feature correction branch. Based on the output values ​​and confidence levels of each branch, the final predicted hemoglobin value is determined, and a mapping relationship model is obtained.

[0041] In this embodiment, the feature matrix is ​​filtered based on the correlation coefficient to obtain the key feature matrix. That is, features with correlation coefficients greater than or equal to the correlation coefficient threshold are selected, and low-correlation features with correlation coefficients less than the correlation coefficient threshold are combined to obtain the key feature matrix. For example, the low-correlation features of roundness and the proportion of lightly stained central area are combined to make up for the insufficient discriminative power of a single feature.

[0042] In this embodiment, a feature-hemoglobin dynamic correlation function is constructed based on the key feature matrix;

[0043] in, This represents the value of the j-th feature of the i-th sample in the enhanced feature matrix; This represents the value of the j-th feature of the i-th sample in the key feature matrix; Indicates the adjustment factor; The Pearson correlation coefficient between the j-th feature and hemoglobin is represented. This represents the mean of the j-th column of the key feature matrix; This represents the standard deviation of the j-th column of the key feature matrix.

[0044] In this embodiment, the extended equation of Lambert-Beer law is: ,in, Indicates hemoglobin content; , Represents the learnable parameters and initial values. =120, =50, A represents the absorbance of red blood cells; constraint equation for the central lightly stained region: ,in, , Indicates learnable parameters, =200, =30; Indicates the proportion of lightly stained central regions of erythrocytes; Morphological-optical co-constraint equation: , where the initial value =0.05, =40, , Indicates learnable parameters; Indicates the area of ​​red blood cells; This represents the average gray value of red blood cells.

[0045] In this embodiment, the physical information neural network includes an input layer, a first hidden layer, a second hidden layer, a third hidden layer, a physical constraint layer, and an output layer. The input layer contains 6 neurons, corresponding to the 6 feature dimensions of the enhanced feature matrix. The first hidden layer contains 64 neurons, employing the ReLU activation function and setting BatchNorm. The second hidden layer contains 32 neurons, employing the ReLU activation function and setting BatchNorm. The third hidden layer contains 16 neurons, employing the ReLU activation function and setting BatchNorm. The physical constraint layer embeds the three physical constraint equations and calculates the physical residuals. The output layer contains 16 neurons, employing the Linear activation function, and outputs a physical mapping feature vector. Simultaneously, a decoding layer is added to reconstruct the physical mapping feature vector back into the key feature matrix.

[0046] In this embodiment, the optimization objective is... ;in, This represents the total loss function during the pre-training phase of a physical information neural network. Indicates physical residual loss; Represents the feature reconstruction loss; , , , These represent the predicted values ​​of the three physical equations for the i-th sample; This represents the true hemoglobin content of the i-th sample; ,in, This represents the value of the j-th feature of the i-th sample in the key feature matrix; This represents the key feature dimension; in this embodiment, d=6. This represents the j-th dimension value in the key feature vector (6-dimensional) reconstructed after processing the physical mapping feature vector (16-dimensional) of the i-th sample through the decoder network; This represents the 16-dimensional physical mapping feature vector obtained after the i-th sample is processed by the physical information neural network.

[0047] In this embodiment, training is performed through a deep learning regression branch, a physical analysis branch, and a feature correction branch, respectively, to obtain the output value of each branch and the corresponding branch loss, including: The deep learning regression branch employs a two-layer LSTM network combined with a one-layer fully connected layer. The fully connected layer contains eight neurons, outputting a predicted value H1 and calculating the MSE loss. ; The physical analysis branch, based on the three physical constraint equations mentioned above, performs a weighted summation of the results of each equation using learnable weights (initial weights w1=0.5, w2=0.3, w3=0.2, outputting the predicted value H2 and calculating the MAE loss). ; The feature correction branch constructs a correction function based on the predicted values ​​H1 from the deep learning regression branch and H2 from the physical analysis branch. ; This represents the predicted value of the feature correction branch for the i-th sample; This represents the prediction value of the deep learning regression branch for the i-th sample; This represents the predicted value of the physical analysis branch for the i-th sample; , Indicates learnable parameters; The key feature-based correction function is typically implemented as a small multilayer perceptron (MLP), with the key feature vector of the i-th sample as input. Calculate the predicted value H3 and the weighted MSE loss. .

[0048] In this embodiment, the confidence level of each branch is calculated based on the branch loss corresponding to the deep learning regression branch, physical analysis branch, and feature correction branch. Based on the loss of each branch , and Calculate according to the formula The sample-level confidence scores for each branch, where, This represents the loss value of the m-th branch for the i-th sample; This represents the confidence level of the m-th branch with respect to the i-th sample; Indicates the adjustment parameter; In this embodiment, based on the output values ​​and confidence levels of each branch, the final predicted hemoglobin value is determined, resulting in a mapping relationship model, including:

[0049] in, This represents the predicted value of the m-th branch for the i-th sample; This represents the final predicted hemoglobin value for the i-th sample; In this embodiment, the total loss of the mapping relationship model is:

[0050] in, This represents the final predicted hemoglobin value for the i-th sample; This represents the true hemoglobin value of the i-th sample; This represents the confidence level of the m-th branch (m=1,2,3) for the i-th sample; Represents the regularization coefficient. =0.1; Represents the total loss function; This represents the total number of samples.

[0051] The working principle and beneficial effects of the above technical solution are as follows: By calculating and filtering the Pearson correlation coefficient between each feature in the feature matrix and hemoglobin content, key features highly correlated with hemoglobin content can be accurately identified from numerous features; a feature-hemoglobin dynamic correlation function is constructed based on the key feature matrix, and the key features are adaptively adjusted to obtain an enhanced feature matrix; this adaptive adjustment considers the dynamic relationship between features and hemoglobin, highlighting the parts of the features that play an important role in hemoglobin prediction; physical prior equations are constructed based on clinical physiological laws, including the Lambert-Beer law extension equation, the central light stained area constraint equation, and the morphological-optical co-constraint equation; these physical prior equations integrate medical and physiological expertise into the model, enabling the model to follow the human body... Physiological patterns are used for prediction, enhancing the model's interpretability and reliability. The constructed physical information neural network contains multiple hidden layers, enabling it to automatically learn complex nonlinear relationships in the feature matrix. The multi-layered hidden layer structure can progressively extract high-level representations of features, thereby better capturing the potential relationship between features and hemoglobin content and improving the model's prediction accuracy. Training is conducted through deep learning regression, physical analysis, and feature correction branches to predict hemoglobin from different perspectives. The deep learning regression branch can learn complex patterns and laws in the data; the physical analysis branch performs analytical calculations based on prior physical knowledge; and the feature correction branch further corrects and optimizes the features. The multi-branch design allows the model to comprehensively utilize the advantages of different methods, improving prediction accuracy and robustness.

[0052] To better demonstrate the hemoglobin content analysis method based on magnified digital microscopic images, this invention provides a hemoglobin content analysis system based on magnified digital microscopic images, comprising: The multi-band microscope imaging module includes a switchable LED light source, a high-magnification optical microscope, and a CMOS image sensor, which is used to acquire magnified digital images of labeled samples and test samples according to preset image acquisition parameters. The image acquisition and preprocessing module is used to acquire magnified digital images and sequentially perform noise reduction, illumination correction and multi-band image alignment operations. The feature extraction module is used to identify red blood cells in the preprocessed magnified digital microscopic image, filter out complete target red blood cells and output the target region image, and extract the morphological and optical features of each target red blood cell, integrating them to form a feature matrix. The model training and prediction module is used to perform statistical correlation analysis between the feature matrix and the actual hemoglobin content based on the statistical method of morphological-optical joint features, eliminate dimensional differences, construct a mapping relationship model, optimize parameters such as the number of decision trees and the maximum tree depth through 5-fold cross-validation, input the feature matrix corresponding to the sample to be tested into the mapping relationship model, and calculate and output the predicted value of hemoglobin content. The results output and verification module is used to determine the true hemoglobin content of the sample to be tested and to calculate the relative error rate between the predicted value and the true value.

[0053] In this embodiment, by utilizing multi-band imaging technology to jointly extract the morphological and optical features of red blood cells, a morphological-optical joint feature matrix is ​​constructed. Based on this matrix, a high-precision mapping relationship model with the actual hemoglobin content is established through statistical correlation analysis. After cross-validation and parameter optimization, the accuracy, objectivity, and robustness of the detection results are significantly improved, effectively eliminating the interference of single features or subjective interpretation. Furthermore, by automatically identifying and analyzing a large number of target red blood cells, high-throughput and standardized data processing is achieved. Finally, by inputting the feature matrix of the sample to be tested into the model, the predicted value of hemoglobin content can be quickly output, which is more representative and makes the detection results highly objective and repeatable.

[0054] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for analyzing hemoglobin content based on magnified digital microscopic images, characterized in that, Includes the following steps: Image acquisition: A magnified digital image of the labeled sample is acquired using a multi-band imaging device, and the acquired magnified digital image is preprocessed. The labeled sample is a blood sample with a predetermined true hemoglobin content, and the magnified digital image includes no less than three different characteristic bands. Multi-feature extraction: Identify red blood cell regions from the preprocessed magnified digital microscopic image, screen the identified red blood cell regions for target red blood cells, output the target red blood cell region image, and simultaneously extract the morphological and optical features corresponding to each target red blood cell to form a feature matrix; Model construction: Construct a mapping model between the feature matrix and the actual hemoglobin content of the labeled sample. Based on the mapping model, calculate the feature matrix of the sample to be tested to obtain the predicted value of the hemoglobin content of the sample to be tested. Prediction and Validation: Output the predicted value of hemoglobin content of the sample to be tested. At the same time, compare the predicted value with the measured true value to verify and calculate the error rate.

2. The method for analyzing hemoglobin content based on magnified digital microscopic images as described in claim 1, characterized in that, The three different characteristic wavelength bands in image acquisition are 380-440nm, 530-550nm, and 580-600nm. The multi-band imaging device includes a switchable LED light source, a high-magnification optical microscope, and an image sensor. The light intensity stability error of the LED light source in each characteristic wavelength band does not exceed ±5%, and the image acquisition frame rate is not less than 15fps.

3. The method for analyzing hemoglobin content based on magnified digital microscopic images as described in claim 2, characterized in that, Preprocessing of the acquired magnified digital microscopic images includes: Image denoising: The acquired multi-band magnified digital images are smoothed to remove Gaussian noise and eliminate salt-and-pepper noise. Illumination correction: Adjusting the grayscale distribution of the denoised magnified digital image to correct grayscale value deviations caused by local brightness deviations; Image alignment: Spatial position calibration of digital images of each band is performed using fixed marker points in the magnified microscopic image as a reference.

4. The method for analyzing hemoglobin content based on magnified digital microscopic images as described in claim 3, characterized in that, Multi-feature extraction, including: Red blood cell recognition: Based on a pre-trained red blood cell recognition model, the target red blood cells in the pre-processed magnified digital microscopic image are identified, and the target red blood cell bounding box is generated. At the same time, the interference of non-target red blood cells and impurities is eliminated, and the target red blood cell region image containing only the red blood cell region is output. Morphological feature extraction: Calculate the morphological features of each target red blood cell based on the target red blood cell bounding box information; Optical feature extraction: Based on the magnified digital images of each band after alignment, the optical features of the target red blood cells are calculated.

5. The method for analyzing hemoglobin content based on magnified digital microscopic images as described in claim 4, characterized in that, Morphological feature extraction, specifically including: Area: The number of pixels in the target red blood cell region image within the target red blood cell bounding box is obtained, and the actual area of ​​the target red blood cell is calculated based on the magnification of the microscope objective. Equivalent diameter: Based on the area formula and calculation, the actual area of ​​the target red blood cell is obtained to obtain the equivalent diameter of the target red blood cell; Roundness: The perimeter of the target red blood cell is obtained through an edge detection algorithm, and the roundness is calculated by combining it with the actual area of ​​the target red blood cell; Central light-stained area proportion: Determine the grayscale thresholds of the central light-stained area and the edge dark area of ​​the target red blood cell, count the number of pixels in the central light-stained area and the entire red blood cell area respectively, and calculate the proportion of the central light-stained area.

6. The method for analyzing hemoglobin content based on magnified digital microscopic images as described in claim 5, characterized in that, Optical feature extraction, specifically including: Average gray value: Traverse all pixels within each red blood cell region, count the pixel gray values ​​in each band, and obtain the average gray value of each target red blood cell in the corresponding band; Gray-scale standard deviation: The degree of deviation between the gray-scale values ​​of all pixels within the target red cell region and the average gray-scale value is calculated based on the average gray-scale value of each band, and is used as the gray-scale standard deviation; Absorbance: In the magnified digital images of each wavelength band, the average gray value of the cell-free blank area is extracted as the gray value corresponding to the incident light intensity. The average gray value of the target red blood cells in the same wavelength band is retrieved as the gray value corresponding to the transmitted light intensity. The absorbance of the target red blood cells in each wavelength band is then calculated.

7. The method for analyzing hemoglobin content based on magnified digital microscopic images as described in claim 6, characterized in that, Prediction and verification, including: Predicted value calculation: The extracted target red blood cell morphological features and optical features of each band are integrated to form a feature matrix, which is then input into a pre-constructed mapping relationship model. The mapping relationship model outputs the predicted value of hemoglobin content of the sample to be tested based on the correlation logic between the feature matrix and the hemoglobin content. Accurate value acquisition: The ferric cyanide method was used to simultaneously detect the hemoglobin content of the test samples and obtain the accurate value of the hemoglobin content of the test samples as a verification benchmark; Error calculation and judgment: Calculate the relative error rate between the predicted value of hemoglobin content and the actual value of hemoglobin content. If the relative error rate is higher than the preset error threshold, trigger the anomaly investigation process. After eliminating operational errors, if the error still exceeds the standard, supplement the labeled data corresponding to the sample to be tested into the model training set and re-optimize the mapping relationship model. Abnormal sample adaptation: For complex samples with slight cell overlap and morphological abnormalities, the contours of individual cells are separated by edge detection algorithm and then morphological features are extracted. The morphological and optical features of morphologically abnormal red blood cells are recorded separately and assigned corresponding feature weights when predicting in the mapping relationship model.

8. A hemoglobin content analysis system based on magnified digital microscopic images, employing the hemoglobin content analysis method based on magnified digital microscopic images as described in claim 7, characterized in that, include: The multi-band microscope imaging module is used to acquire magnified digital images of labeled samples and test samples according to preset image acquisition parameters; The image acquisition and preprocessing module is used to acquire magnified digital images and sequentially perform noise reduction, illumination correction and multi-band image alignment operations. The feature extraction module is used to identify red blood cells in the preprocessed magnified digital microscopic image, filter out complete target red blood cells and output the target region image, and extract the morphological and optical features of each target red blood cell, integrating them to form a feature matrix. The model training and prediction module is used to perform statistical correlation analysis between the feature matrix and the actual hemoglobin content based on the statistical method of morphological-optical joint features, construct a mapping relationship model, input the feature matrix corresponding to the sample to be tested into the mapping relationship model, and calculate and output the predicted value of hemoglobin content. The results output and verification module is used to determine the true hemoglobin content of the sample to be tested and to calculate the relative error rate between the predicted value and the true value.

9. The method for analyzing hemoglobin content based on magnified digital microscopic images as described in claim 1, characterized in that, The morphological and optical features corresponding to each target red blood cell are extracted to form a feature matrix, including: Feature extraction is performed on each target red blood cell to obtain the corresponding morphological and optical features of each target red blood cell; An original feature matrix is ​​constructed based on the morphological and optical features corresponding to each target red blood cell; ； in, express ×7-dimensional original feature matrix; Indicates the first The area of ​​each target red blood cell; Indicates the first The equivalent diameter of a target red blood cell; Indicates the first The roundness of each target red blood cell; Indicates the first The percentage of lightly stained central areas in each target red blood cell; Indicates the first The average gray value of each target red blood cell; Indicates the first The standard deviation of grayscale values ​​for each target red blood cell; Indicates the first The absorbance of each target red blood cell; Based on a clinical normal red blood cell sample database, the statistical distribution of each feature is calculated to generate a conventional baseline vector; the conventional baseline vector is then expanded to generate a feature conventional baseline matrix. Calculate the deviation of each feature in each original feature matrix from the conventional benchmark; ； in, Represents the original feature matrix. The first target red blood cell Deviation of each feature; Represents the original feature matrix. The first target red blood cell The original values ​​of each feature; Indicates the first The mean of each feature in the feature standard reference matrix; Indicates the first The standard deviation of each feature in the feature benchmark matrix; Indicates the dynamic decay factor; Indicates the first The weights of each feature; The original feature matrix is ​​corrected based on the deviation to obtain the target feature matrix; ； in, Represents the original feature matrix. The first target red blood cell The corrected values ​​of each feature; This represents the first number of normal red blood cells. The standard baseline mean of each feature.

10. The method for analyzing hemoglobin content based on magnified digital microscopic images as described in claim 1, characterized in that, Methods for constructing mapping relationship models include: Calculate the Pearson correlation coefficient between each feature in the feature matrix and the hemoglobin content, and filter the feature matrix based on the correlation coefficient to obtain the key feature matrix; Based on the key feature matrix, a feature-hemoglobin dynamic correlation function is constructed, and each feature in the key feature matrix is ​​adaptively adjusted to obtain an enhanced feature matrix. Physical prior equations are constructed based on clinical physiological principles; these physical prior equations include the Lambert-Beer law extension equation, the central light stained area constraint equation, and the morphological-optical co-constraint equation. A physical information neural network is constructed, comprising an input layer, a first hidden layer, a second hidden layer, a third hidden layer, a physical constraint layer, and an output layer. The enhanced feature matrix is ​​input into the physical information neural network, which outputs a physical mapping feature vector. The physical residual loss and feature reconstruction loss of the physical constraint equation are used as optimization objectives to pre-train the physical information neural network and obtain a stable physical mapping feature vector. Using the stable physical mapping feature vector as input, the training is performed through the deep learning regression branch, the physical analysis branch and the feature correction branch respectively, to obtain the output value of each branch and the branch loss corresponding to each branch; The sample-level confidence of each branch is calculated based on the branch loss corresponding to the deep learning regression branch, physical analysis branch, and feature correction branch. Based on the output values ​​and confidence levels of each branch, the final predicted hemoglobin value is determined, and a mapping relationship model is obtained.

Images