Abdominal fat distribution analysis device and method based on near-infrared multi-band detection

Abstract

The application discloses a kind of based on near infrared multiband detection abdominal fat distribution analysis device and method, it is related to fat analysis field, the spectrum of the information of subcutaneous adipose tissue (SAT) mainly reflected by close-range probe acquisition, accurate estimation is combined with pre-training model individualized subcutaneous fat profile parameter. Then, these parameters are used to drive an optical forward propagation model, dynamically predict and generate the SAT contribution spectrum of the individual at the remote probe. Finally, subtract this accurate prediction SAT contribution signal from the mixed spectrum collected by the remote probe, thereby effectively stripping the strong background noise. The signal-to-noise ratio of the residual spectrum obtained is significantly improved, and the visceral fat index can be accurately analyzed through a special correction model, solving the core technical problem that VAT cannot be measured due to SAT signal interference.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of fat analysis, and more particularly, to an abdominal fat distribution analysis device and method based on near-infrared multi-band detection. BACKGROUND

[0002] Excessive accumulation of abdominal fat, especially the increase of visceral adipose tissue (VAT), is a key risk factor for many metabolic diseases. Therefore, being able to accurately distinguish and quantify subcutaneous adipose tissue (SAT) and visceral fat has important clinical value for health assessment and disease prevention. Traditional detection methods such as magnetic resonance imaging (MRI) are accurate but costly and inconvenient, and are difficult to meet the needs of daily monitoring. Near-infrared spectroscopy (NIRS) technology provides a feasible way for portable analysis of abdominal fat due to its non-invasive and rapid characteristics.

[0003] However, using near-infrared spectroscopy technology to accurately analyze abdominal fat distribution, especially to measure visceral fat, faces great technical challenges. The core difficulty lies in the extraction of visceral fat signals. Light propagates in human tissue with sharp attenuation, and needs to penetrate the skin, subcutaneous fat and muscle layers to reach the visceral fat area, resulting in extremely weak effective signals returning to the detector. At the same time, since the light source and detector are both located on the body surface, the vast majority of signals received by the detector come from shallow subcutaneous fat with shorter path and stronger contribution. This strong signal from SAT completely overwhelms the weak target signal of deep VAT, like a huge background noise. Existing technologies attempt to use different source-probe distances to distinguish signals at different depths, for example, assuming that a long-distance probe can mainly receive signals from deep tissues. But this assumption is too simplistic, because the detection path of the long-distance probe still covers a large amount of subcutaneous fat tissue, and the light spectrum it receives is essentially a mixture of a large amount of SAT contribution + a small amount of VAT contribution. If this mixed signal is not effectively processed, the interference signal of subcutaneous fat cannot be accurately stripped, and the information reflecting the visceral fat condition cannot be accurately extracted.

[0004] Therefore, how to separate and extract the weak visceral fat characteristic signal from the strong background noise is a technical problem that needs to be solved in this field. SUMMARY

[0005] To solve the existing problems, according to an aspect of the present application, a kind of abdominal fat distribution analysis method based on near infrared multi-band detection is provided, it includes: obtaining the first original spectrum data and the second original spectrum data collected by first photodetector and second photodetector, wherein the distance between first photodetector and light source is less than the distance between second photodetector and light source;First original spectrum data and second original spectrum data are preprocessed and converted to absorption spectrum to obtain first spectrum vector and second spectrum vector;SAT layer optical parameter estimation is carried out to the first spectrum vector to obtain subcutaneous fat profile parameter;The dynamic prediction and stripping of SAT contribution signal is carried out to the second spectrum vector based on the subcutaneous fat profile parameter to obtain residual spectrum vector;The residual spectrum vector is input into pre-trained multivariate correction model to obtain final visceral fat index.

[0006] According to another aspect of the present application, a kind of abdominal fat distribution analysis device based on near infrared multi-band detection is provided, it includes: original spectrum data acquisition module, for obtaining the first original spectrum data and the second original spectrum data collected by first photodetector and second photodetector, wherein the distance between first photodetector and light source is less than the distance between second photodetector and light source;Original spectrum data preprocessing conversion module, for first original spectrum data and second original spectrum data are preprocessed and converted to absorption spectrum to obtain first spectrum vector and second spectrum vector;Optical parameter estimation module, for SAT layer optical parameter estimation is carried out to the first spectrum vector to obtain subcutaneous fat profile parameter;Dynamic prediction and stripping module, for the dynamic prediction and stripping of SAT contribution signal is carried out to the second spectrum vector based on the subcutaneous fat profile parameter to obtain residual spectrum vector;Final visceral fat index generation module, for the residual spectrum vector is input into pre-trained multivariate correction model to obtain final visceral fat index.

[0007] Compared with the prior art, the application provides an abdominal fat distribution analysis device and method based on near-infrared multi-waveband detection, which proposes a dynamic stripping strategy to solve the extraction problem of visceral fat signals. First, a double-probe system is used to collect the spectrum mainly reflecting the subcutaneous fat (SAT) information by the close-range probe, and the individualized subcutaneous fat profile parameters are accurately estimated by combining the pre-trained model. Then, the parameters are used to drive an optical forward propagation model to dynamically predict and generate the SAT contribution spectrum of the individual at the long-range probe. Finally, the accurate predicted SAT contribution signal is subtracted from the mixed spectrum collected by the long-range probe, so as to effectively strip the strong background noise. The signal-to-noise ratio of the visceral fat (VAT) signal of the residual spectrum obtained in this way is significantly improved, and the visceral fat index can be accurately analyzed through a special correction model, thereby solving the core technical problem that VAT cannot be measured due to the interference of SAT signals. BRIEF DESCRIPTION OF DRAWINGS

[0008] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description of embodiments of the present application taken in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of embodiments of the present application and are incorporated in and constitute a part of this specification, illustrate embodiments of the present application and serve to explain the present application, but do not limit the present application. In the drawings, the same reference numerals refer to the same components or steps throughout the drawings.

[0009] Figure 1 A flowchart of the abdominal fat distribution analysis method based on near-infrared multi-waveband detection according to the embodiments of the present application.

[0010] Figure 2 A data flow schematic diagram of the abdominal fat distribution analysis method based on near-infrared multi-waveband detection according to the embodiments of the present application.

[0011] Figure 3 A flowchart of step S2 in the abdominal fat distribution analysis method based on near-infrared multi-waveband detection according to the embodiments of the present application.

[0012] Figure 4 A flowchart of step S4 in the abdominal fat distribution analysis method based on near-infrared multi-waveband detection according to the embodiments of the present application.

[0013] Figure 5 A block diagram of the abdominal fat distribution analysis device based on near-infrared multi-waveband detection according to the embodiments of the present application. DETAILED DESCRIPTION

[0014] Embodiments of the present disclosure will be described below in greater detail with reference to the accompanying drawings. While certain embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be embodied in various forms and should not be interpreted as being limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly and completely understood.

[0015] To address the problems in the foregoing background art description, the present application proposes an abdominal fat distribution analysis method based on near-infrared multi-band detection. Figure 1 A flowchart of the abdominal fat distribution analysis method based on near-infrared multi-band detection according to an embodiment of the present application. Figure 2 A data flow diagram of the abdominal fat distribution analysis method based on near-infrared multi-band detection according to an embodiment of the present application. As shown in Figure 1 and Figure 2 As shown in the drawings, the abdominal fat distribution analysis method based on near-infrared multi-band detection according to an embodiment of the present application includes: step S1, acquiring first original spectral data and second original spectral data collected by a first photodetector and a second photodetector, wherein the distance between the first photodetector and a light source is less than the distance between the second photodetector and the light source; step S2, performing data preprocessing and absorption spectrum conversion on the first original spectral data and the second original spectral data to obtain a first spectral vector and a second spectral vector; step S3, performing SAT layer optical parameter estimation on the first spectral vector to obtain a subcutaneous fat profile parameter; step S4, performing dynamic prediction and stripping of the SAT contribution signal based on the subcutaneous fat profile parameter to obtain a residual spectral vector; and step S5, inputting the residual spectral vector into a pre-trained multivariate correction model to obtain a final visceral fat index.

[0016] In step S1, first raw spectral data and second raw spectral data collected by a first photodetector and a second photodetector are acquired, wherein the distance between the first photodetector and the light source is less than the distance between the second photodetector and the light source. It should be understood that, when using near-infrared spectroscopy to detect abdominal fat, the propagation path and depth of photons in the tissue are closely related to the distance between the light source and the detector (source-detector distance). When the source-detector distance is short, the detector mainly receives photons that are scattered back in the shallow tissue (mainly the subcutaneous fat layer) through a short path, and the spectral information can highly represent the characteristics of the subcutaneous fat. When the source-detector distance is long, the photons have a greater probability of penetrating to deeper tissue levels, and although the detection light path covers a deeper area, a large amount of shallow subcutaneous fat is still contained in the signal. Therefore, in order to effectively distinguish and ultimately separate the interference signal from the subcutaneous fat, the present application adopts a differentiated detection strategy. By setting a first photodetector at a close distance and a second photodetector at a long distance, two sets of spectral data with different depth information sensitivity can be acquired simultaneously. The first raw spectral data collected by the close-range detector provides a key basis for subsequent accurate evaluation of the characteristics of the subcutaneous fat, and the second raw spectral data collected by the long-range detector serves as a mixed signal containing information of the target visceral fat, laying a data foundation for subsequent signal separation and analysis.

[0017] In one specific embodiment of the present application, the specific process of step S1 is as follows: First, a detection device containing a specific hardware configuration needs to be built. The device integrates a wide-spectrum near-infrared light source, such as a light-emitting diode array, with an emission wavelength range covering 700 nanometers to 1100 nanometers. Around the light source, two high-sensitivity photodetectors, such as avalanche photodiodes or silicon photomultipliers, are arranged concentrically or linearly. In one specific embodiment of the present application, the distance between the first photodetector and the light source is 1.5 cm, and the distance between the second photodetector and the light source is 4.0 cm. This distance configuration has undergone a large number of optical simulations and experimental verifications, aiming to ensure that the first detector can maximize the capture of signals from the subcutaneous fat layer, and the second detector can effectively detect deep signals containing visceral fat information.

[0018] When measuring, the operator tightly attaches the probe end face of the detection device to a specific position of the abdomen of the measured person, such as the side of the navel. After starting the measurement program, the light source emits near-infrared light according to the preset timing and power. After entering the abdominal tissue, the light is absorbed and scattered at different depths, and is finally received by the first photodetector and the second photodetector, respectively. The detector converts the received light signal into a corresponding electrical signal. For example, at a certain time, the light source emits light with a wavelength of 850 nanometers, the light intensity received by the first detector corresponds to a voltage value V1, and the light intensity received by the second detector corresponds to another voltage value V2.

[0019] To acquire the complete spectral data, the light source scans a number of discrete wavelength points within its working wavelength range, for example, at 5 nanometer intervals, from 700 nanometers to 1100 nanometers. At each wavelength point, the two detectors simultaneously collect the light intensity. This process continues for a short measurement period, for example, 2 seconds, to improve the signal-to-noise ratio by repeating sampling multiple times. The collected raw data are two sets of time series, each of which contains the light intensity values at each wavelength point over time.

[0020] Finally, the collected data are processed. For the first photodetector, the time series light intensity data collected at all wavelength points are combined to form the first raw spectral data. Similarly, for the second photodetector, the time series light intensity data collected at all wavelength points are combined to form the second raw spectral data.

[0021] In step S2, the first raw spectral data and the second raw spectral data are preprocessed and converted into absorption spectra to obtain the first spectral vector and the second spectral vector. Accordingly, the raw spectral data directly obtained from the detectors are inevitably disturbed by various physical and physiological factors. For example, the electronic noise of the instrument itself, environmental light interference, and the small physiological activities (such as heartbeat and breathing) of the subject can all cause signal fluctuations. In addition, the raw light intensity data are not linearly related to the concentration of tissue components and are difficult to be directly used for chemometric modeling. Therefore, in order to eliminate these random noises and systematic drifts, extract stable and directly related spectral information of the tissue absorption characteristics, and convert them into a standardized format suitable for subsequent multivariate correction model analysis, a series of preprocessing and conversion are required to purify the raw measurement values into clean and standardized spectral vectors that accurately reflect the optical properties of the abdominal tissue, thereby providing a high-quality data basis for subsequent accurate subcutaneous fat parameter estimation and visceral fat signal stripping.

[0022] In one specific embodiment of the present application, Figure 3 The flowchart of step S2 in the abdominal fat distribution analysis method based on near-infrared multi-band detection according to the embodiment of the present application is shown in FIG. 2. As shown in FIG. 2, Figure 3As shown, in step S2, data preprocessing and absorption spectrum conversion are performed on the first and second original spectrum data to obtain the first and second spectrum vectors, including: step S21, time series signal filtering and mean value processing are performed on the first and second original spectrum data to obtain first and second stable light intensity vectors; step S22, light intensity-absorbance conversion is performed on the first and second stable light intensity vectors to obtain first and second absorption spectrum vectors; and step S23, smoothing filtering and standard normal transformation are performed on the first and second absorption spectrum vectors to obtain the first and second spectrum vectors.

[0023] In the above specific embodiments, the specific process of step S2 is as follows: first, step S21 is performed, time series signal filtering and mean value processing. It can be understood that in actual near-infrared spectrum collection, the original signal output by the detector is not an ideal stable value, but a dynamic time series signal containing various noise components. These noises are derived from the thermal noise of the internal electronic elements of the instrument, the shot noise in the photoelectric conversion process, and the inevitable electromagnetic interference (such as power frequency interference) in the measurement environment. In addition, the physiological activities of the measured person, such as the heartbeat and the abdominal micro-displacement and blood flow changes caused by breathing, also superimpose low-frequency fluctuations on the signal. These random and periodic interferences can seriously affect the quality of the spectrum data, reduce the signal-to-noise ratio, and if not processed, will directly affect the accuracy of subsequent analysis. Therefore, in order to extract stable and reliable light intensity information that can represent the true optical characteristics of the tissue from the original measurement stream full of noise, the original time series data needs to be filtered and mean value processed.

[0024] Specifically, first, the first raw spectral data is processed. This data is a data set containing the light intensity value sequence collected continuously in a measurement time window, for example, 2 seconds, at a plurality of preset wavelength points, for example, 81 wavelength points with a step of 5 nanometers from 700 nanometers to 1100 nanometers. Taking the data at a wavelength of 850 nanometers as an example, the original data can be a time sequence light intensity value containing 200 data points. Since the sequence can mix high-frequency electronic noise and 50 Hz power frequency interference, a digital low-pass filter is first applied to it. Specifically, a fourth-order Butterworth low-pass filter can be used, and the cutoff frequency thereof is set to 10 Hz. The selection of this cutoff frequency is based on the analysis of physiological signals (such as heartbeat usually at 1-2 Hz, and respiration at 0.2-0.5 Hz) and main noise frequencies, aiming to retain useful physiological rhythm information while effectively filtering out random noise and power frequency interference of higher frequencies. Applying the filter to the 200 time series data points at the 850 nanometer wavelength point, a filtered and smoother time sequence is obtained. Then, all data points in the filtered sequence are calculated by arithmetic mean, thereby obtaining a single value representing the stable light intensity response at this wavelength. For example, the 200 filtered light intensity values are added and divided by 200, obtaining a mean value, which is the element value corresponding to the 850 nanometer wavelength in the first stable light intensity vector. The filtering and averaging process described above is applied to the time series data of all 81 wavelength points contained in the first raw spectral data in turn. After processing all wavelength points, the first stable light intensity vector is obtained. The vector is a one-dimensional array with a length of 81, and each element corresponds to the stable light intensity mean at a wavelength of 700 nanometers, 705 nanometers,..., 1100 nanometers, respectively. The second raw spectral data is processed using the same method and parameter settings. That is, the time series light intensity data collected by the second detector at each wavelength point is also filtered by a fourth-order Butterworth low-pass filter with a cutoff frequency of 10 Hz, and then the arithmetic mean of the time sequence is calculated. Finally, a second stable light intensity vector of 81 dimensions is obtained.

[0025] Next, step S22 is performed to convert the light intensity to absorbance. Accordingly, the stable light intensity vector obtained after filtering and averaging eliminates most of the random noise, but it is still in physical light intensity units and has a complex nonlinear relationship with the concentration of chemical components (such as fat and water) in the tissue to be measured. Directly using light intensity data for quantitative analysis will be disturbed by many factors such as light source intensity fluctuations, detector response differences, and different scattering characteristics between samples, making it difficult to establish a stable and accurate quantitative model. According to the Beer-Lambert law, under ideal conditions, the absorbance of light is proportional to the concentration of the absorbing substance. Therefore, in order to convert the measurement data from the physical domain to the chemical domain and establish a more direct and linear correlation with the concentration of tissue components, the light intensity needs to be converted to absorbance.

[0026] In one specific embodiment of the present application, step S22, converting the first stable light intensity vector and the second stable light intensity vector to a first absorption spectrum vector and a second absorption spectrum vector, includes: based on the reference light intensity vector, converting each element in the first stable light intensity vector and the second stable light intensity vector to the first absorption spectrum vector and the second absorption spectrum vector according to the following formula: ; wherein, is the light intensity value of each element in the first stable light intensity vector and the second stable light intensity vector, is each reference light intensity in the reference light intensity vector, is the absorbance of each element in the first absorption spectrum vector and the second absorption spectrum vector. Specifically, first, a reference, i.e. a reference light intensity vector, needs to be obtained. It is worth noting that the acquisition of this reference light intensity vector is a key calibration step. Before the actual measurement of the human body, the probe end face of the detection device is tightly attached to a standard reference body with known optical properties. The reference body is preferably a material with high Lambertian reflection characteristics, such as a medical-grade polytetrafluoroethylene (PTFE) standard plate with sufficient thickness and smooth surface. Then, the same measurement process as in step S1 is performed, i.e. the light source is driven to scan all wavelength points, and the light intensity is collected by the first and second photodetectors. Since the reflection characteristics of the standard reference body are constant and known during the measurement, the collected light intensity can be regarded as a reference signal representing the maximum response capability of the instrument. The raw spectral data measured on the standard reference body is also subjected to the time series signal filtering and averaging process in S21, and finally one or more reference stable light intensity vectors are obtained. In this embodiment, only the data measured on the standard reference body by the first photodetector can be used to generate a unified reference light intensity vector, which has the same dimension as the number of measurement wavelength points, e.g. 81 dimensions. After obtaining the reference light intensity vector, the stable light intensity vector from the human body can be converted. Taking the first stable light intensity vector as an example, it contains 81 elements, each corresponding to the stable light intensity value at different wavelengths. The conversion process follows the formula. Specifically, for the first element in the first stable light intensity vector corresponding to 700 nanometer wavelength, its value is taken, and the reference light intensity value corresponding to 700 nanometer wavelength is taken from the reference light intensity vector, and the calculation is performed by the above formula to obtain the absorbance at 700 nanometer wavelength. Then, the point-by-point calculation process is repeated for all 81 elements in the first stable light intensity vector. All the calculation results are combined to form the first absorption spectrum vector, each element of which represents the absorbance of the human tissue at the corresponding wavelength. The second stable light intensity vector is processed in the same way, and the reference light intensity vector obtained before is used to calculate each element in the second stable light intensity vector by the formula, and finally the second absorption spectrum vector is generated.

[0027] Finally, the step S23 is performed to smooth the absorption spectrum vector and apply a standard normal variate transformation. It should be appreciated that the absorption spectrum vector after the conversion is theoretically related to the chemical concentration, but still contains some residual interference information in the actual measurement. For example, there can be high-frequency random noise in the spectrum curve that is introduced by the instrument and is unrelated to the chemical information, which can affect the accurate identification of the spectral features. More importantly, due to physical factors such as uneven contact pressure of the probe with the skin, differences in the surface state of the skin, and the like, the total scattering path length of the light in the tissue changes between different individuals or even between different times of the same measurement. This change introduces an additive baseline drift or multiplicative scale change in the entire spectrum, which masks the real spectral changes caused by the absorption differences of the chemical components of the tissue. Therefore, in order to further purify the spectral signal and eliminate the interference caused by these physical effects, and to make the spectra of different samples comparable, the absorption spectrum vector needs to be finally processed in detail.

[0028] Specifically, first, the first absorption spectrum vector is smoothed. This vector is a one-dimensional array with the same dimension as the number of wavelength points, e.g. 81 dimensions, but there can be small high-frequency noises between adjacent data points. To filter out these noises while preserving the main spectral absorption peak features, Savitzky-Golay smoothing filter algorithm is adopted. This algorithm can effectively smooth the data without causing peak shape broadening or peak position shift by least square fitting a low-order polynomial to the data within a moving window. The specific parameter setting is: polynomial order 2, window width 9 data points. This means that for each point in the first absorption spectrum vector, itself and the previous and next 4 points (total 9 points) are taken to fit a quadratic polynomial to these 9 points, and then the value of the fitted polynomial at this point is used to replace the original value. This process starts from the 5th point of the spectrum and ends at the 5th point from the end, and special treatment or keeping unchanged can be adopted for the boundary points. After SG smoothing, a first smoothed absorption spectrum vector with lower noise level is obtained. Next, the first smoothed absorption spectrum vector is subjected to standard normal variate (SNV) transformation. SNV is a common row processing method used to correct baseline drift and scale changes caused by light scattering. The calculation process consists of two steps: first, calculate the arithmetic mean and standard deviation of all 81 elements of the spectrum vector. For example, the mean of the first smoothed absorption spectrum vector is calculated to be 0.8, and the standard deviation is 0.15. Second, subtract the arithmetic mean 0.8 from each element in the vector, and then divide the result by the standard deviation 0.15. That is, for each element Ai in the vector, perform the transformation: A'i=(Ai-0.8) / 0.15. After SNV processing, the resulting spectrum vector has a mean of 0 and a standard deviation of 1. This one-dimensional array generated finally is the first spectrum vector. The same processing flow and parameter setting are used to operate on the second absorption spectrum vector. That is, SG smoothing filter with window width 9 and polynomial order 2 is first applied to obtain the second smoothed absorption spectrum vector. Then, the mean and standard deviation of this vector itself are calculated, and SNV transformation is performed. Finally, the second spectrum vector with a mean of 0 and a standard deviation of 1 is obtained.

[0029] In step S3, the first spectrum vector is subjected to SAT layer optical parameter estimation to obtain subcutaneous fat profile parameters. It can be understood that the background art has explained that the spectrum collected by the remote detector is a mixture of subcutaneous fat (SAT) and visceral fat (VAT) signals, in which the SAT signal occupies the dominant position and constitutes a strong interference. In order to accurately extract the weak VAT signal, the contribution of the SAT signal needs to be accurately quantified and stripped off. However, the SAT thickness, composition (such as water content, collagen content) and optical properties (absorption coefficient, scattering coefficient) of each person have significant individual differences. Such differences result in the dynamic change of the contribution of SAT to the remote spectrum, which cannot be deducted by a fixed mode. Therefore, in order to realize the individualization and dynamic stripping of SAT interference, the SAT layer of the current subject needs to be accurately imaged first. That is, by analyzing the near-infrared spectrum most sensitive to SAT information, the key parameters characterizing the SAT layer characteristics of the current subject are estimated in real time.

[0030] In a specific embodiment of the present application, step S3, the first spectrum vector is subjected to SAT layer optical parameter estimation to obtain subcutaneous fat profile parameters, comprising: inputting the first spectrum vector into a trained multivariate correction model dedicated to SAT analysis to obtain the subcutaneous fat profile parameters, wherein the training data of the trained multivariate correction model dedicated to SAT analysis is [near-infrared spectrum, SAT thickness / composition measured by MRI] data pair.

[0031] In the above specific embodiment, the specific process of step S3 is as follows: first, the dedicated multivariate correction model needs to be constructed and trained. The model is preferably a partial least squares regression (PLSR) model. It is worth mentioning that the PLSR model has significant advantages in dealing with chemometrics problems, especially in spectral analysis. Near-infrared spectrum data usually has two typical characteristics: one is high dimension, for example, the spectrum vector in this embodiment contains 81 wavelength points, i.e. 81 variables; the second is that there is a high degree of collinearity between variables, i.e. the absorption values of adjacent wavelengths are highly correlated. The PLSR algorithm can effectively cope with these challenges. It maximizes the covariance of the latent variables (LVs) between the decomposition of the spectral data matrix (X) and the response variable matrix (Y), so as to extract a few comprehensive variables (i.e. LVs) most related to the target parameters (such as SAT thickness) from the spectrum with a large amount of noise and redundant information, and establish a stable regression model based on these LVs, avoiding the instability and overfitting risk of traditional multivariate linear regression due to the collinearity problem.

[0032] The training process of the PLSR model is a data-driven process, the core of which is to establish the mapping relationship between the near-infrared spectral information and the SAT physicochemical parameters measured by the gold standard. The training data come from a large sample dataset, for example, 200 volunteers with different body types, ages, genders, and fat distributions are recruited. For each volunteer, two measurements are performed successively at the same precise location on their abdomen. First, the probe device of the present application is used to collect the spectrum through the close-range probe (source-probe distance of 1.5 cm), and the first spectrum vector is processed into an 81-dimensional standardized vector by strictly following the aforementioned steps S1 and S2, which constitutes the input part (X) of the training data pair. Subsequently, immediately use the high-resolution magnetic resonance imaging (MRI) equipment to perform a cross-sectional scan on the exact same location. Through professional medical image analysis software, the subcutaneous fat layer is accurately segmented from the MRI image, and its vertical thickness is measured, for example, for a certain volunteer, the measured value is 15.3 mm, and the relative content ratio of fat and water in the layer (for example, fat accounts for 80%) can also be further analyzed, and even the equivalent absorption coefficient and reduced scattering coefficient spectrum can be obtained by combining the diffusion optical theory. These parameters obtained from the MRI, which are considered to be the most accurate, constitute the target output part (Y) of the training data pair. Input these 200 sets of [first spectrum vector, SAT parameters measured by MRI] data pairs into the PLSR training algorithm. The algorithm determines the optimal number of latent variables (for example, 5 LVs are determined by cross-validation method) through iterative optimization, and calculates the weight matrix required to project the original 81-dimensional spectrum vector into this 5-dimensional latent variable space, and the regression coefficient matrix required to map the coordinates in this 5-dimensional latent variable space to the final SAT parameter output. These weight matrix and regression coefficient matrix are fixed in the model, marking the completion of the training of the model.

[0033] In actual measurement applications, when this step is performed, the first spectrum vector is directly input into the PLSR model that has been trained. The calculation process inside the model is to perform matrix multiplication operation on the input first spectrum vector and the regression coefficient vector stored in the model, and add a bias term. This operation process is essentially to invert the corresponding physical and chemical properties according to the features in the spectrum such as the absorption peak intensity and shape of water and fat at a specific wavelength. The output of the model is the subcutaneous fat profile parameter. This parameter can be a vector containing multiple key indicators. For example, the model may output a vector containing three elements: [18.2, 0.65, 0.9]. Among them, the first element 18.2 represents the estimated subcutaneous fat thickness of 18.2 mm; the second element 0.65 may represent the volume percentage of fat and water; and the third element 0.9 may represent a comprehensive index related to scattering characteristics.

[0034] In actual measurement application, when a new subject is detected, the first spectrum vector obtained by pre-processing is sent into the trained PLSR model. The model first multiplies the 81-dimensional input vector with the weight matrix to convert it into a 5-dimensional latent variable score vector. Then, the score vector is multiplied with the regression coefficient matrix and the bias term determined in training is added. The output of this calculation process is the subcutaneous fat profile parameters. For example, for a specific input spectrum, the output of the model can be a vector containing three elements: [18.2, 0.85, 0.9]. Each element in this vector has a clear physical meaning: the first element 18.2 is the predicted subcutaneous fat thickness of 18.2 mm, the second element 0.85 is the predicted fat content percentage, and the third element 0.9 can be a dimensionless index representing the average scattering ability of the tissue.

[0035] In step S4, the dynamic prediction and stripping of the SAT contribution signal from the second spectrum vector is performed based on the subcutaneous fat profile parameters to obtain a residual spectrum vector. It should be understood that the previous steps have successfully accurately personalized the subcutaneous fat layer (SAT) of the current subject, obtaining its thickness, composition, and other key profile parameters. However, the ultimate goal is to detect and quantify the visceral fat (VAT) covered by the SAT layer. The second spectrum vector collected by the remote detector, although with a deeper detection depth, is still mostly contributed by the propagation path of photons in the SAT layer, and the signal characteristics of VAT are severely overwhelmed by this strong SAT background signal. Simply assuming that the remote spectrum represents deep information is a false assumption. Therefore, in order to reveal the weak VAT information from this mixed signal, an active and accurate signal stripping strategy is needed. Thus, the personalized SAT profile parameters obtained in the previous step are used to dynamically and quantitatively predict the specific contribution of the SAT layer to the second spectrum vector, and then subtract it from the mixed spectrum to separate the residual signal that truly originates from the deep tissue.

[0036] In a specific embodiment of the present application, Figure 4 The flowchart of step S4 in the abdominal fat distribution analysis method based on near-infrared multi-band detection according to the embodiment of the present application is shown in FIG. 4. As shown in FIG. 4, step S4, the dynamic prediction and stripping of the SAT contribution signal from the second spectrum vector based on the subcutaneous fat profile parameters to obtain a residual spectrum vector, includes: step S41, inputting the subcutaneous fat profile parameters into the trained optical forward propagation model to obtain a predicted subcutaneous fat contribution spectrum; and step S42, calculating the residual between the second spectrum vector and the predicted subcutaneous fat contribution spectrum to obtain the residual spectrum vector. Figure 4

[0037] ​In the above specific embodiments, the specific procedure of step S4 is as follows: first, step S41 is performed to construct and train the optical forward propagation model. The optical forward propagation model in the present application is preferably a deep neural network (DNN), the function of which can be understood as a highly specialized nonlinear function for accurately mapping a set of parameters describing the physical properties of subcutaneous adipose tissue (SAT) to the spectral signal that the adipose layer should produce under specific detection conditions. The construction and training process of the DNN aims to encode and solidify the complex physical laws of photon propagation in biological tissue into a series of numerical matrices (i.e. weights) and vectors (i.e. biases), thereby achieving fast prediction from tissue parameters to spectral signals.

[0038] The training process of the DNN model begins with large-scale, physics-based computer simulation. First, a digital two-layer abdominal tissue model is constructed, which only contains the skin layer and the SAT layer. By systematically and finely stepping through the key parameters of the SAT layer, such as varying its thickness in the physiological range of 5 to 40 millimeters, adjusting the volume percentage of fat and water in the interval of 60% to 95%, and assigning corresponding wavelength-dependent absorption coefficient and reduced scattering coefficient values to each component ratio according to literature or experimental data. For each set of parameters, such as a virtual SAT sample with a thickness of 18.2 millimeters and a fat content of 85%, a rigorous forward simulation is performed using Monte Carlo photon propagation simulation software. In the simulation, millions of photons enter the tissue from the source position, and their every scattering and absorption event in the tissue is tracked. Finally, the number and wavelength distribution of photons escaping the tissue surface at a source-probe distance of 4.0 centimeters are statistically recorded, resulting in a corresponding theoretically accurate long-distance spectrum. By repeating this process thousands of times, a large training dataset containing rich physical information is generated, with each record being a [input SAT profile parameters, simulated long-distance spectrum] data pair.

[0039] Subsequently, this dataset is used to specifically construct and train a DNN with a certain topology. The architecture of this network is precisely defined: its input layer is provided with 3 neurons, each of which is used to receive one of the three values in the subcutaneous fat profile parameter vector, namely the thickness, the fat content and the scattering index. The core of the network is four fully connected hidden layers, each of which contains 128 neurons, and the layers are connected through weight matrices and bias vectors. The activation function of each neuron is the rectified linear unit (ReLU), which effectively introduces nonlinearity and enables the network to fit the complex physical process of photon propagation. The output layer of the network is provided with 81 neurons, the number of which is exactly the same as the number of wavelength points in the spectrum to be predicted, and each neuron corresponds to the spectral absorbance value at a specific wavelength. During the training phase, the training dataset is repeatedly input into the network through the standard backpropagation algorithm and the Adam optimizer. According to the mean square error between the predicted spectrum output by the network and the corresponding true simulation spectrum in the dataset, the algorithm automatically and iteratively fine-tunes the connection weights and bias terms of all hidden layers and the output layer in the network. This process continues until the error converges to a minimum value, at which point all the weights and bias parameters in the network are finally determined and fixed, forming a trained, efficient prediction model that accurately reflects the physical relationship between the SAT parameters and the spectrum.

[0040] In the actual measurement process, when this step is performed, the subcutaneous fat profile parameters, such as [18.2, 0.85, 0.9], are input into the input layer of this trained DNN model. This input vector is first multiplied by the weight matrix of the first hidden layer and added to the bias vector of this layer, and the result is processed by the ReLU activation function to generate the output of the first hidden layer. This output is then input into the second hidden layer, and the same multiplication, addition and activation operations are repeated, and so on, layer by layer, so that the information flows through all four hidden layers and is transformed by a complex nonlinearity at each layer. Finally, the output of the fourth hidden layer is input into the output layer, multiplied by the weight matrix of the output layer and added to the bias, and the 81 neurons in the output layer will each generate a value. These 81 values combined form an 81-dimensional vector, which is the predicted subcutaneous fat contribution spectrum, which accurately reproduces the spectral shape that a SAT layer with a thickness of 18.2 mm and a fat content of 85% should exhibit at a source-probe distance of 4.0 cm.

[0041] Then, step S42 is performed to calculate the residual. The second spectrum vector is subtracted from the predicted subcutaneous fat contribution spectrum. That is, the corresponding elements of the two 81-dimensional vectors are subtracted one by one. The physical meaning of this subtraction operation is to accurately remove the SAT contribution part predicted by the model from the true mixed signal containing the SAT contribution + VAT contribution. The final result of the calculation is a new 81-dimensional vector, which is the residual spectrum vector.

[0042] In particular, in the conventional processing flow, the multivariate correction model and the optical forward propagation model dedicated to SAT analysis are two independent modules, each of which is in charge of its own task. The former is dedicated to accurately estimating the subcutaneous fat profile parameters from the spectrum, and the latter is dedicated to perfectly reconstructing the spectrum according to the parameters. Although this independent training method is convenient to operate, its performance upper limit is inevitably limited by the local optimal solution of each subtask. That is, when the SAT analysis model pursues the accuracy of parameter estimation, it may inadvertently discard or distort some weak spectral features that are crucial for the final visceral adipose tissue (VAT) prediction. Therefore, in order to break this inherent performance bottleneck, achieve global optimization, and maximize the prediction accuracy of the final visceral fat index, the present application introduces an end-to-end joint training strategy to co-optimize the two models as a whole.

[0043] Based on this, in one specific embodiment of the present application, training the multivariate correction model and the optical forward propagation model dedicated to SAT analysis includes: based on the first spectrum vector, the second spectrum vector and the residual spectrum vector, calculating a cycle consistency loss function item by model forward propagation and comparison, that is: ; wherein, and is the first spectrum vector and the second spectrum vector, is the multivariate correction model dedicated to SAT analysis, is the optical forward propagation model, is the residual spectrum vector, is the two-norm of the calculation vector, is the position point subtraction, is the cycle consistency loss function item. It should be understood that, since the input and output of the multivariate correction model dedicated to SAT analysis and the optical forward propagation model are relative to each other, they can be regarded as a self-consistent pair of shared physical knowledge that constrains each other, and therefore, based on the bidirectional constraint and unity of the encoder-decoder, they can be constructed as an enhanced autoencoder, that is, the multivariate correction model dedicated to SAT analysis (as an encoder) and the optical forward propagation model (as a decoder) are constructed as an enhanced autoencoder architecture, and the physical constraint is applied by using the spectrum-parameter-spectrum closed loop. The cycle consistency loss function item constrains the residual distribution between the spectrum reconstructed by the decoder from the parameters estimated by the encoder from the first spectrum vector and the true second spectrum vector to be consistent with the distribution of the target residual spectrum vector by model forward propagation and comparison. In this way, it forces the parameter representation learned by the encoder to be not only abstract numerical values, but also parameter representations that are useful for the reconstruction of the spectrum by the decoder and have clear physical meaning, thereby ensuring the self-consistency and physical reality of the entire information flow process.

[0044] Based on the first spectrum vector and the residual spectrum vector, a feedback loss function term based on the residual spectrum vector is constructed, i.e., ; wherein, represents the KL divergence value between the calculation vectors, is the feedback loss function term based on the residual spectrum vector. Accordingly, during joint training, if the model excessively focuses on the perfection of parameter estimation and the purification of the distribution of the residual spectrum, it may cause a negative effect, i.e., while removing the SAT component, the extremely weak VAT features mixed therein are also ignored or eliminated. Therefore, the feedback loss function term is designed to establish the reverse feedback correlation of the residual spectrum vector to different dimension parameters and spectrum vectors. That is, it allows to produce a small, strategic error on the SAT parameter estimation, as long as this error can help to achieve more accurate prediction of VAT features. By introducing a differentiable measure such as KL divergence, the loss term ensures the differentiability under the condition of numerical difference, avoids the problem of gradient disappearance or instability in the training process, and ensures that the model will not sacrifice the prediction accuracy of the final VAT while pursuing SAT stripping.

[0045] The weighted sum of the cycle consistency loss function term and the feedback loss function term based on the residual spectrum vector is calculated as the final loss function, and the multi-element correction model and the optical forward propagation model dedicated to SAT analysis are trained end-to-end by joint training through gradient descent back propagation, i.e., ; wherein, and are the weights of the cycle consistency loss and the feedback loss, respectively, which are determined by experimental optimization on the validation set, is the final loss function. That is, by setting the preset weight coefficient, for example, setting the weight of the cycle consistency loss as 0.2 and the weight of the feedback loss as 0.8, the relative importance between the physical consistency constraint and the final target feedback during training can be accurately regulated. By minimizing this comprehensive final loss function, the two models can be effectively and cooperatively trained end-to-end, so that they evolve together towards the global optimal goal of improving the final prediction accuracy of the visceral fat index, thereby obtaining a performance beyond that of independent training methods.

[0046] In step S5, the residual spectrum vector is input into the pre-trained multivariate calibration model to obtain the final visceral fat index. Accordingly, through a series of precise signal processing and stripping in the preceding steps, a residual spectrum vector has been successfully extracted from the remote mixed spectrum seriously contaminated by subcutaneous fat (SAT). This residual spectrum has been theoretically maximally purified from the interference of SAT, and the spectral characteristics it contains mainly reflect the results of photon interaction with deeper tissues, especially visceral fat (VAT) after penetrating the SAT layer. However, this residual spectrum vector itself is still a complex set of spectral data, not a directly readable clinical index. There is a complex, nonlinear quantitative relationship between it and the specific quantity of VAT (such as volume, area, or some index). Therefore, in order to complete the conversion from a purified spectral signal to a final quantitative result with clear clinical significance, a special quantitative analysis model is needed to interpret the residual spectrum and calculate the final visceral fat index from it.

[0047] In a specific embodiment of the present application, the specific process of step S5 is as follows: the multivariate calibration model for VAT analysis is also preferably a partial least squares regression (PLSR) model. The essence of this model is to mathematically encode and solidify the complex, implicit quantitative relationship between the residual spectrum vector purified in the previous steps, which is rich in deep tissue information, and the VAT values measured by the widely accepted gold standard magnetic resonance imaging (MRI). The PLSR model can intelligently extract the core information most relevant to the target variable (VAT value) from spectral data full of noise and highly correlated between wavelengths, and construct a stable and reliable prediction function, which is crucial for interpreting the weak VAT signal characteristics after multiple processing.

[0048] The training process of this model is similar to the model in S3, but its target and data source are essentially different. Training this model requires a special data set containing VAT gold standard measurement values. Specifically, a group of volunteers is recruited, and for each volunteer, the residual spectrum vector of the abdomen is first calculated using the device of the present application and the complete processing flow of the preceding S1 to S4, which constitutes the input part (X) of the training data pair. Then, the same abdominal region is scanned using magnetic resonance imaging (MRI), and the cross-sectional area of visceral fat is accurately calculated through professional image post-processing software (for example, the VAT area of a volunteer at the fourth lumbar vertebra level is 120 square centimeters)

[0049] or total volume. This precise VAT measurement from MRI constitutes the true value label (Y) in the training data pair. By collecting, for example, 200 such [residual spectrum vector, MRI measured VAT value] data pairs, a training dataset for the VAT model is constructed. With this dataset, the PLSR algorithm is trained. The algorithm will dig out the weak features in the residual spectrum that are related to VAT changes, for example, absorption patterns in certain wavelength regions, and establish a robust mapping from the 81-dimensional residual spectrum space to the final VAT index value. This process will generate a set of regression coefficients (weight vector) and bias term that are specific to VAT analysis, which are optimized and fixed in this multivariate calibration model, marking the completion of the training of the VAT quantitative model.

[0050] In the actual measurement process, when this final step is performed, the residual spectrum vector is directly input into this PLSR model that has been trained and is specific to VAT analysis. Inside the model, an efficient calculation process is performed: the input residual spectrum vector is multiplied by the VAT regression coefficient vector stored in the model, and then the corresponding bias term is added. The result of this operation is a single numerical value. For example, for a specific residual spectrum input, the output of the model may be 125.5. This numerical value is the final visceral fat index, which represents that the model predicts the visceral fat area to be 125.5 square centimeters according to the input spectrum. This final output index can be used for clinical evaluation, providing a convenient, non-invasive, and quantitative scientific basis for the user's health management and disease risk warning.

[0051] In summary, the abdominal fat distribution analysis method based on near-infrared multi-band detection according to the embodiments of the present application is illustrated, which proposes a dynamic stripping strategy to solve the problem of extracting visceral fat signals. First, using a dual-probe system, the spectrum collected by the close-range probe, which mainly reflects the subcutaneous fat (SAT) information, is combined with the pre-trained model to accurately estimate the individualized subcutaneous fat profile parameters. Then, these parameters are used to drive an optical forward propagation model to dynamically predict and generate the SAT contribution spectrum that the individual should have at the long-range probe. Finally, the accurately predicted SAT contribution signal is subtracted from the mixed spectrum collected by the long-range probe, thereby effectively stripping the strong background noise. The residual spectrum obtained in this way has a significantly improved signal-to-noise ratio of the visceral fat (VAT) signal, and the visceral fat index can be accurately analyzed through a special correction model, thereby solving the core technical problem that VAT cannot be measured due to SAT signal interference.

[0052] Figure 5 A block diagram of an abdominal fat distribution analysis device based on near-infrared multi-band detection according to embodiments of the present application is shown in FIG. 1. As shown in FIG. 1, the device includes a near-infrared multi-band detector 100, a data processing unit 200, and a display unit 300. Figure 5As shown, the abdominal fat distribution analysis apparatus 100 based on near-infrared multi-band detection according to the embodiment of the present application comprises: a raw spectrum data acquisition module 110, configured to acquire first raw spectrum data and second raw spectrum data collected by a first photodetector and a second photodetector, wherein the distance between the first photodetector and a light source is less than the distance between the second photodetector and the light source; a raw spectrum data preprocessing and conversion module 120, configured to perform data preprocessing and absorption spectrum conversion on the first raw spectrum data and the second raw spectrum data to obtain a first spectrum vector and a second spectrum vector; an optical parameter estimation module 130, configured to perform SAT layer optical parameter estimation on the first spectrum vector to obtain a subcutaneous fat profile parameter; a dynamic prediction and stripping module 140, configured to perform dynamic prediction and stripping of a SAT contribution signal on the second spectrum vector based on the subcutaneous fat profile parameter to obtain a residual spectrum vector; and a final visceral fat index generation module 150, configured to input the residual spectrum vector into a pre-trained multivariate correction model to obtain a final visceral fat index.

[0053] As described above, the abdominal fat distribution analysis apparatus 100 based on near-infrared multi-band detection according to the embodiment of the present application can be implemented in various wireless terminals, such as a server with an abdominal fat distribution analysis algorithm based on near-infrared multi-band detection, etc. In a possible implementation manner, the abdominal fat distribution analysis apparatus 100 based on near-infrared multi-band detection according to the embodiment of the present application can be integrated into a wireless terminal as a software module and / or a hardware module. For example, the abdominal fat distribution analysis apparatus 100 based on near-infrared multi-band detection can be a software module in an operating system of the wireless terminal, or can be an application program developed for the wireless terminal; of course, the abdominal fat distribution analysis apparatus 100 based on near-infrared multi-band detection can also be one of many hardware modules of the wireless terminal.

[0054] Alternatively, in another example, the abdominal fat distribution analysis apparatus 100 based on near-infrared multi-band detection and the wireless terminal can also be separate devices, and the abdominal fat distribution analysis apparatus 100 based on near-infrared multi-band detection can be connected to the wireless terminal through a wired and / or wireless network, and transmit interactive information in an agreed data format.

[0055] Here, those skilled in the art can understand that the specific operations of each step in the above-described abdominal fat distribution analysis apparatus based on near-infrared multi-band detection have been described in detail above with reference to the description of the abdominal fat distribution analysis method based on near-infrared multi-band detection of the present application Figures 1 to 3 , and therefore, the repeated description thereof will be omitted.

Claims

1. A method for analyzing abdominal fat distribution based on near-infrared multi-band detection, characterized in that, include: First and second raw spectral data are acquired by a first photodetector and a second photodetector, wherein the distance between the first photodetector and the light source is less than the distance between the second photodetector and the light source; data preprocessing and absorption spectral conversion are performed on the first and second raw spectral data to obtain a first spectral vector and a second spectral vector; SAT layer optical parameter estimation is performed on the first spectral vector to obtain subcutaneous fat contour parameters. Based on the subcutaneous fat contour parameters, the SAT contribution signal of the second spectral vector is dynamically predicted and stripped to obtain the residual spectral vector; the residual spectral vector is then input into a pre-trained multivariate correction model to obtain the final visceral fat index.

2. The method for abdominal fat distribution analysis based on near-infrared multi-band detection according to claim 1, characterized in that, The distance between the first photodetector and the light source is 1.5 cm, and the distance between the second photodetector and the light source is 4.0 cm.

3. The method for abdominal fat distribution analysis based on near-infrared multi-band detection according to claim 1, characterized in that, The data preprocessing and absorption spectrum transformation of the first and second original spectral data to obtain the first and second spectral vectors includes: performing time-series signal filtering and mean averaging on the first and second original spectral data to obtain the first and second stable light intensity vectors; performing light intensity-absorbance transformation on the first and second stable light intensity vectors to obtain the first and second absorption spectral vectors; and performing smoothing filtering and standard normal transformation on the first and second absorption spectral vectors to obtain the first and second spectral vectors.

4. The method for abdominal fat distribution analysis based on near-infrared multi-band detection according to claim 3, characterized in that, Performing an intensity-absorbance conversion on the first stable light intensity vector and the second stable light intensity vector to obtain a first absorption spectrum vector and a second absorption spectrum vector includes: based on a reference light intensity vector, performing an intensity-absorbance conversion on each element of the first stable light intensity vector and the second stable light intensity vector using the following formula to obtain the first absorption spectrum vector and the second absorption spectrum vector, wherein the formula is: ;in, Let be the light intensity values ​​of each element in the first stable light intensity vector and the second stable light intensity vector. For each reference light intensity in the reference light intensity vector, It represents the absorbance of each element in the first absorption spectral vector and the second absorption spectral vector.

5. The method for abdominal fat distribution analysis based on near-infrared multi-band detection according to claim 1, characterized in that, Estimating the SAT layer optical parameters of the first spectral vector to obtain subcutaneous fat contour parameters includes: inputting the first spectral vector into a trained multivariate correction model specifically for SAT analysis to obtain the subcutaneous fat contour parameters, wherein the training data of the trained multivariate correction model specifically for SAT analysis is a data pair of [near-field spectroscopy, SAT thickness / composition measured by MRI].

6. The method for abdominal fat distribution analysis based on near-infrared multi-band detection according to claim 1, characterized in that, Dynamic prediction and stripping of the SAT contribution signal of the second spectral vector based on the subcutaneous fat contour parameters to obtain the residual spectral vector includes: inputting the subcutaneous fat contour parameters into a trained optical forward propagation model to obtain the predicted subcutaneous fat contribution spectrum; and calculating the residual between the second spectral vector and the predicted subcutaneous fat contribution spectrum to obtain the residual spectral vector.

7. A device for analyzing abdominal fat distribution based on near-infrared multi-band detection, characterized in that, include: The raw spectral data acquisition module is used to acquire the first raw spectral data and the second raw spectral data collected by the first photodetector and the second photodetector, wherein the distance between the first photodetector and the light source is smaller than the distance between the second photodetector and the light source; the raw spectral data preprocessing and conversion module is used to perform data preprocessing and absorption spectral conversion on the first raw spectral data and the second raw spectral data to obtain the first spectral vector and the second spectral vector. An optical parameter estimation module is used to estimate the SAT layer optical parameters of the first spectral vector to obtain subcutaneous fat contour parameters; a dynamic prediction and stripping module is used to dynamically predict and strip the SAT contribution signal of the second spectral vector based on the subcutaneous fat contour parameters to obtain a residual spectral vector; and a final visceral fat index generation module is used to input the residual spectral vector into a pre-trained multivariate correction model to obtain the final visceral fat index.

Images