A non-invasive hemoglobin detection method based on atlas combination

Abstract

The present application relates to a kind of non-invasive hemoglobin detection methods based on atlas combination, belong to physiological signal detection field.The present application is by using white light source to irradiate the skin area (finger) containing human body, simultaneously using hyperspectral camera and industrial camera simultaneously to the image acquisition of skin area, by training spectral super-resolution model, realize hyperspectral reconstruction, then to hyperspectral data de-redundancy while retaining the features of image spatial information and spectral information, further according to the result of Monte Carlo simulation determines the channel attention module for specific wave band, finally establishes regression prediction model and predicts human hemoglobin concentration.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of human health detection, relates to the fields of image processing, deep learning and optical technology detection, and particularly relates to a method for non-invasive detection of hemoglobin concentration. BACKGROUND

[0002] As one of the most common clinical laboratory tests, blood hemoglobin detection is usually used as a preliminary screening for red blood cell reduction, and its content is an important basis for clinical diagnosis of anemia and other blood diseases. Recent research has found that blood hemoglobin or hematocrit levels are highly correlated with acute hemorrhage. Therefore, real-time monitoring of human hemoglobin is of great significance to human health.

[0003] Unlike pulse oximetry, blood hemoglobin measurement is not direct, and the current reliable biological detection method is to extract blood by needle method, usually using a portable blood analysis instrument for blood extraction, which is seriously dependent on short shelf life and environmentally sensitive color bands, and has the problems of high price and is not conducive to popularization; With the development of telemedicine technology, simpler and safer non-invasive hemoglobin detection technology is gradually replacing blood-based detection, and the equipment for non-invasive detection of blood hemoglobin also has the problem of high cost, and often causes doubts in accuracy, and still needs a large number of experimental verification to become a reliable health monitoring product. At present, the non-invasive monitoring of hemoglobin is still based on optical measurement method, mainly using the intensity information, phase information, spectrum information, polarization information of the transmitted light and the absorption and scattering information of the human skin tissue to establish the relationship with the hemoglobin, and the optical non-invasive hemoglobin detection mainly includes photoacoustic spectroscopy, photoelectric volume pulse wave method and spectral analysis method. The photoacoustic spectroscopy needs a large photoacoustic spectroscopy system to realize the hemoglobin detection, which is expensive and not suitable for real-time monitoring in a family environment; The photoelectric volume pulse wave method combines the pulse signal caused by the beating of the heart, and uses the time domain information of the human body to realize the physiological parameter monitoring, and is mainly applied to heart rate monitoring, and is one of the mainstream directions of the current non-invasive hemoglobin monitoring, but this method is greatly affected by individual differences, and is generally used for predicting the overall trend change of the hemoglobin concentration, and the accuracy needs to be improved; The spectral analysis method combines the advantages of different analysis methods, and mainly includes color analysis method, spectral imaging, near-infrared spectral analysis and the like. The color analysis method utilizes the color characteristics of hemoglobin, and realizes the hemoglobin concentration detection through RGB color analysis, but the imaging equipment is restricted by factors such as environmental light, shooting distance and shooting angle when acquiring features; The spectral imaging method can directly display the vessel diameter, and also has problems such as complex and expensive equipment; The near-infrared spectral analysis has good penetration to human skin and the like, and is an ideal waveband for non-invasive detection, but the human blood composition is complex, and the absorption spectrum of other components overlaps with the absorption spectrum of hemoglobin, which affects the detection precision; In the face of this problem, the application provides a new non-invasive hemoglobin concentration detection method, provides a practical valuable technical scheme, and realizes non-invasive hemoglobin concentration detection.

[0004] The working principle of the method based on the atlas combined non-invasive hemoglobin concentration detection method is to utilize the absorption and scattering of light by hemoglobin molecules in blood, to perform RGB imaging on a skin area through a visible light imaging device, to perform hyperspectral image reconstruction through a spectral super-resolution algorithm, to perform feature extraction through a designed algorithm while retaining spatial information and spectral information of multidimensional data, and to finally establish a regression relationship between feature values and hemoglobin concentration, so as to obtain a hemoglobin concentration value. SUMMARY

[0005] The application provides a detection method for measuring the hemoglobin concentration of a human body quickly and non-invasively under the illumination of a visible light band light source by using a visible light imaging device. After removing the interference of stray light, the RGB image of the skin of a finger of the measured object is collected first, and then the collected image is reconstructed into a hyperspectral image by a spectral super-resolution algorithm. Then, feature extraction is performed by atlas combination, and a regression algorithm is established to predict the hemoglobin concentration. The application can realize non-invasive hemoglobin concentration detection, meets the daily detection needs of ordinary people, has wide development space and application prospect, and details are described below: the detection method is realized by a non-invasive hemoglobin detection method based on atlas combination, and the method is characterized in that: the spectral super-resolution algorithm data collection is completed by an ordinary industrial camera, a hyperspectral camera and a lens; and the illumination light source is a white light source. The method first collects the RGB image of a finger but not limited to a finger and thirty-one channel hyperspectral images as a data set to train a spectral super-resolution network. Then, the hyperspectral image is subjected to data dimension reduction and feature extraction, while the spatial information and spectral information of the data are retained. Finally, the characteristic values are input into a hemoglobin regression prediction model, so that the hemoglobin concentration value of the measured object can be obtained.

[0006] The imaging device is located at a distance of 25-40 cm from the finger, so that the non-contact non-invasive measurement of the hemoglobin concentration value is realized, and the risk of infection caused by the contact invasive measurement is avoided.

[0007] The light source is a white light source.

[0008] The spectral super-resolution algorithm includes but is not limited to a multi-scale hybrid domain attention mechanism network re-model for skin images.

[0009] The hyperspectral image data dimension reduction mode includes but is not limited to a generalized low-rank model for removing redundancy of high-dimensional data.

[0010] The hemoglobin feature information includes but is not limited to spectral information and image spatial information, and the feature extraction is performed by 3D convolution. The channel attention module and 3D convolution are adjusted by Monte Carlo simulation simulation data, and the characteristic values contain the basic information contained in the image.

[0011] The hemoglobin regression prediction model includes but is not limited to a multivariate linear regression and a classical algorithm fused by double quadratic weighting.

[0012] The hemoglobin detection method comprises the following steps.

[0013] Step 1, collect a data set to train a spectral super-resolution network model;

[0014] 1), start the white light source, and emit visible light to the finger or other parts of the skin.

[0015] 2) Start the industrial camera and hyperspectral camera, ensure that no other light enters to affect the data collection of the imaging device.

[0016] 3), the fingers are placed at the position of the angle between the light source and the imaging device, and kept still for 10s, and the RGB image and hyperspectral image are collected.

[0017] Step 2, create a feature value extraction model based on atlas combination;

[0018] 1), design a skin diffuse reflection experiment to determine the wave band weight.

[0019] 2), reduce the dimensionality of the hyperspectral data and remove redundancy.

[0020] 3), 3D convolution is carried out in the experiment to extract features, while the spatial information and spectral information are reserved.

[0021] Step 3, predict the human hemoglobin concentration according to the feature information;

[0022] 1), establish a hemoglobin concentration regression prediction model based on a classic algorithm.

[0023] 2), the feature value is input into the regression model to predict the human hemoglobin concentration.

[0024] 3), compare a plurality of predicted values with true values to calculate the error range.

[0025] Preferably, the hyperspectral image reconstruction adopts a multi-scale hybrid domain attention mechanism network model for skin images.

[0026] Preferably, Monte Carlo simulation experiment is adopted to determine the key wave band of hemoglobin.

[0027] Preferably, the data dimensionality reduction adopts a generalized low-rank model for high-dimensional data de-redundancy.

[0028] Preferably, the regression prediction model adopts an algorithm based on multiple linear regression and double quadratic weighted fusion.

[0029] Beneficial effects

[0030] 1. The non-invasive hemoglobin concentration detection method based on atlas combination does not need to collect blood from the measured object, reduces the pain of the subject, and avoids the risk of wound infection caused by multiple measurements.

[0031] 2. The non-invasive hemoglobin detection method based on atlas combination makes full use of the basic information of hyperspectral data, and the de-redundancy operation on the data can effectively reduce the running time and improve the accuracy and rapidity of predicting human hemoglobin.

[0032] 3. The non-invasive hemoglobin detection method based on atlas combination of the present application only needs a common industrial camera as an input device, without additional equipment, and has low cost. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is a schematic diagram of the method of the present application.

[0034] Figure 2 is a flowchart of the non-invasive hemoglobin detection method based on atlas combination of the present application.

[0035] Among them, 1-imaging device, 2-test object, 3-white light source. DETAILED DESCRIPTION

[0036] In order to make the purpose, advantages and characteristics of the present application clearer, a non-invasive hemoglobin detection method based on atlas combination of the present application is further described in detail below in combination with the drawings and specific embodiments. It should be noted that: the drawings are all greatly simplified and use non-precise proportions, only to facilitate and clearly assist the purpose of explaining the embodiments of the present application, the structures shown in the drawings are part of the actual structures; the acquisition part of the image signal of the present application is not limited to the human finger part, and other parts are also applicable. The hyperspectral image reconstruction algorithm used in the present application is not limited to a certain algorithm. The regression prediction algorithm used in the present application is also not limited to a certain algorithm.

[0037] The embodiment of the present application takes a common healthy person as a subject.

[0038] The non-invasive hemoglobin detection method based on atlas combination disclosed in the present embodiment is shown in the image acquisition schematic diagram as shown in Figure 1 , the execution process is shown in the attached Figure 2 , and the specific content is as follows.

[0039] Step 1, simultaneously use a hyperspectral camera and a common industrial camera to collect hyperspectral images and RGB images of the same finger skin tissue to train a model;

[0040] Step 1-1, after isolating stray light, start the white light source, and the tester remains stationary to determine that the light source uniformly irradiates on the human finger skin.

[0041] Step 1-2, simultaneously start the industrial camera and the hyperspectral camera, focus to the clearest place, and calibrate the hyperspectral camera with a whiteboard.

[0042] Step 1-3, place the finger at the specified position, keep a distance of 25-40 cm from the hyperspectral camera and the industrial camera, remain stationary for 10-15 seconds, and collect hyperspectral images and RGB images.

[0043] Step 1-4, the collected RGB image and hyperspectral image are taken as the training set, and are taken as the input of the multi-scale hybrid domain attention mechanism network model, wherein the mathematical description of the reconstruction process is:

[0044] x 3×1 =S 3×N y N×1 +e 3×1 (1)

[0045] Wherein x corresponds to the 3x1 vector of RGB three channels (x=[R, G, B] T =[I(R),I(G),I(B)] T ), S is the 3xN matrix corresponding to the spectral response of the RGB camera, y is the Nx1 vector (y=[I(λ1),I(λ2),···,I(λ N ] T ) represents the spectral intensity, and e is a 3x1 vector of system noise with zero mean. Expanding it to m pixels, it can be expressed as:

[0046] x 3×m =S 3×N y N×m (2)

[0047] (2) can be expanded in detail as:

[0048]

[0049] Wherein S is the spectral response function of the RGB channel of the industrial camera, and the hyperspectral data matrix corresponding to m pixels can be obtained as:

[0050] y N×m =[S 3×N ] -1 x 3×m (4)

[0051] The reconstructed hyperspectral data is compared with the source data for evaluation, and the calculation method of three evaluation indexes can be represented as:

[0052]

[0053]

[0054]

[0055] Wherein G(x) is the generated hyperspectral image, y is the original image, and R is the response function of the camera.

[0056] Step 2 extracts the feature points of the hyperspectral image while retaining the spatial information and spectral information;

[0057] Step 2-1, the hyperspectral data is reduced in dimension using a generalized low-rank model, and the generated hyperspectral data is decomposed into the product of two sub-matrices, namely a basis matrix spanning a subspace and a corresponding coefficient matrix, and the sub-matrix product is maximized to approximate the original data matrix using a minimum error estimation method, and the coefficient matrix can cover most of the information of the original matrix.

[0058] Step 2-2, the coefficient matrix is subjected to convolution processing, and 3D convolution simultaneously extracts features in space and spectrum, which can include most of the information related to the concentration of hemoglobin.

[0059] Step 2-3, the Monte Carlo simulation method is used to perform light transmission experiments in skin tissue under different hemoglobin concentrations, and the skin diffuse reflectance spectrum under different hemoglobin concentrations is obtained, and a channel attention module is set for the diffuse reflectance spectrum to obtain more comprehensive feature information.

[0060] Step 3, a regression prediction model for human hemoglobin characteristic values and true values is established, and the measured human hemoglobin concentration is predicted after inputting the characteristic values;

[0061] Step 3-1, a regression relationship between the hemoglobin characteristic values and the true values is established using an algorithm based on multiple linear regression and double quadratic weighted fusion.

[0062] Step 3-2, the hemoglobin characteristics are input into the correction model to predict the hemoglobin concentration.

[0063] Step 3-3, the hemoglobin prediction data and the true value are processed, and the accuracy of this method is evaluated using an evaluation function.

[0064]

[0065] In the formula, y i represents the true value, R represents the predicted value of the model. 2 The value range of R is (0, 1), when R 2 is closer to 1, it represents that the fitting degree of the model is better.

[0066] In summary, the above is only a preferred embodiment provided by the present application, and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application shall be included in the protection scope of the present application.

Claims

1. A non-invasive hemoglobin detection method based on spectral fusion, used to achieve non-invasive, non-contact hemoglobin concentration detection, characterized in that, This method consists of a spectral super-resolution model, a combined spectral and hyperspectral hemoglobin feature extraction model, and a regression prediction model. The spectral super-resolution model is a multi-scale hybrid domain attention mechanism prediction model for skin. The feature extraction model employs an algorithm combining spectral and hyperspectral features with a generalized low-rank structure. The regression prediction model uses an algorithm based on multiple linear regression and biquadratic weighted fusion. The method first acquires RGB and hyperspectral images of the same area of ​​the finger in an environment free from stray light interference. Then, the reconstructed hyperspectral data is deredundanted, and features are extracted using a 3D convolution algorithm based on Monte Carlo simulation results, while preserving the spatial integrity of the image. The spectral information is used to input feature values ​​into a regression prediction model to predict the human hemoglobin concentration. The imaging device is located 25-40cm from the fingertip, enabling non-invasive measurement of hemoglobin concentration and avoiding the risks of pain and infection associated with invasive measurements. The light source includes a white light source. The spectral super-resolution model includes a multi-scale hybrid domain attention mechanism network. The data redundancy removal algorithm includes a generalized low-rank model. The feature extraction model includes a 3D convolution-based graph-spectral joint feature extraction algorithm. The regression prediction model includes an algorithm based on multiple linear regression and biquadratic weighted fusion. The hemoglobin detection method includes the following steps. Step 1: Collect datasets to train a network model to reconstruct hyperspectral images; 1) Turn on the white light source to emit visible light to your finger; 2) Start the industrial camera and hyperspectral camera, perform focusing and whiteboard calibration to ensure that no interfering light affects the image acquisition by the camera; 3) Place your finger at a fixed angle to the acquisition device and hold it still for 10 seconds to acquire RGB and hyperspectral images; Step 2: Establish a graph-based joint feature extraction model based on 3D convolution; 1) Design Monte Carlo simulation experiments to determine the diffuse reflectance spectrum of the skin under different hemoglobin concentrations; 2) Perform matrix decomposition on the hyperspectral data within the minimum error range, decomposing it into a basis matrix and the coefficient matrix corresponding to the basis matrix; 3) Use 3D convolution to extract features from hyperspectral data while preserving spatial and spectral information; Step 3: Establish a regression prediction model to predict the concentration of human hemoglobin; 1) Establish a hemoglobin prediction algorithm based on multiple linear regression and biquadratic weighted fusion; 2) Input hemoglobin features into the calibration model to predict hemoglobin concentration; 3) Process the predicted hemoglobin data and the true values, and use the evaluation function to assess the accuracy of the hemoglobin prediction.

Images