A transcutaneous oxygen partial pressure measurement method based on near-infrared spectroscopy and residual network

Abstract

The present invention relates to the field of artificial intelligence technology, and specifically discloses a transcutaneous oxygen partial pressure measurement method based on near-infrared spectroscopy and residual network, the method comprising: obtaining optical signals and body temperature values ​​based on near-infrared spectroscopy; constructing a linear model based on the modified Lambert-Beer law; inputting the optical signal into the linear model and outputting an intermediate value; constructing a nonlinear model corresponding to the transcutaneous oxygen partial pressure value based on the intermediate value and the body temperature value; constructing a neural network model based on the residual network and the nonlinear model; generating a transcutaneous oxygen partial pressure result based on the neural network model, and the linear model can quickly capture the linear relationship in the data, while the nonlinear model is good at capturing complex nonlinear relationships. The cascade architecture can take advantage of the advantages of both at the same time to improve the performance and prediction accuracy of the overall model.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for measuring transcutaneous oxygen partial pressure based on near-infrared spectroscopy and residual network. Background Art

[0002] Transcutaneous oxygen pressure measurement technology is a medical method that can non-invasively, in real time, and continuously monitor the oxygen pressure value under the human skin surface. This technology can help clinicians evaluate the patient's blood circulation and tissue oxygen supply. At present, the methods for measuring transcutaneous oxygen pressure in tissues are mainly divided into two categories: invasive measurement and non-invasive measurement. Invasive measurement takes a long time to detect, is cumbersome to operate, and is only applicable to a single measurement, and cannot obtain detection data in real time and continuously. In contrast, non-invasive measurement methods mainly include transcutaneous oxygen pressure measurement technology based on gas sensing electrodes and traditional near-infrared spectroscopy (NIR). Transcutaneous oxygen pressure technology based on gas sensing electrodes has certain defects: (1) The measurement results of this technology are affected by local factors such as skin temperature, blood flow, and tissue vascular conditions. Therefore, there may be large variability between different parts and different patients, which reduces the consistency and reliability of the measurement results; (2) Skin damage or abnormal conditions such as wounds, scars, and blisters may affect the accuracy of the measurement; (3) Due to the lag between the change in tissue oxygen pressure and the change in arterial oxygen pressure, this technology may not be able to capture the rapid changes in time. Traditional near-infrared spectroscopy technology also has certain defects, mainly including: (1) light will be absorbed and scattered in tissues, so the measurement of deep tissues is limited; (2) the sensitivity of NIR is limited by the performance of the light source and detector; (3) NIR usually provides information on relative changes rather than absolute concentration or partial pressure values;

[0003] (4) Data processing and interpretation require complex algorithms and models to convert optical signals into physiological parameters; (5) NIR systems need to be calibrated frequently to ensure measurement accuracy. Summary of the invention

[0004] The object of the present invention is to provide a method for measuring transcutaneous oxygen partial pressure based on near-infrared spectroscopy and residual network to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: a method for measuring transcutaneous oxygen partial pressure based on near-infrared spectroscopy and residual network, the method comprising:

[0006] Obtain optical signals and body temperature values ​​based on near-infrared spectroscopy;

[0007] A linear model was constructed based on the modified Lambert-Beer law;

[0008] The optical signal is input into the linear model and the intermediate value is output;

[0009] A nonlinear model corresponding to the transcutaneous oxygen partial pressure value was constructed based on the intermediate value and the body temperature value;

[0010] Construct a neural network model based on residual network and nonlinear model;

[0011] Generate transcutaneous oxygen partial pressure results based on a neural network model.

[0012] As a further solution of the present invention, the step of acquiring optical signals and body temperature values ​​based on near-infrared spectroscopy includes: acquiring optical signals through a sensor with a 4-light source-6-detector array structure.

[0013] As a further solution of the present invention, the step of constructing a linear model based on the modified Lambert-Beer law specifically includes:

[0014] Get the modified Lambert-Beer law formula;

[0015] The modified Lambert-Beer law formula is further modified based on the 4-light source-6-detector array structure and the additive combination theorem;

[0016] A linear model was constructed based on the modified formula.

[0017] As a further solution of the present invention, the step of constructing a nonlinear model corresponding to the transcutaneous oxygen partial pressure value based on the intermediate value and the body temperature value specifically includes:

[0018] Determine the factors that influence transcutaneous oxygen tension;

[0019] The correlation analysis of impact factors was conducted based on the Pearson correlation coefficient method;

[0020] Determine independent variables and covariates based on the analysis results;

[0021] A nonlinear model is established based on independent variables and covariates.

[0022] As a further solution of the present invention, the step of constructing a neural network model based on a residual network and a nonlinear model specifically includes:

[0023] Collect sample data;

[0024] Preprocess the sample data;

[0025] A neural network model is constructed based on the preprocessed sample data.

[0026] As a further solution of the present invention, the step of preprocessing the sample data specifically includes:

[0027] Clean the sample data;

[0028] Perform Z-Score standardization on the cleaned sample data;

[0029] The Min-Max method is used to compress each dimensional component of the sample;

[0030] The preprocessed sample data is divided into training set, validation set and test set.

[0031] As a further solution of the present invention, the neural network model includes 1 convolutional layer, 3 residual blocks, 1 maximum pooling layer, 1 flattening layer, 2 fully connected layers and 1 inactivation layer.

[0032] Compared with the prior art, the beneficial effects of the present invention are: linear models can quickly capture linear relationships in data, while nonlinear models are good at capturing complex nonlinear relationships. The cascade architecture can take advantage of the advantages of both at the same time to improve the performance and prediction accuracy of the overall model. The linear model can first extract the basic features in the data, and then the nonlinear model further processes and optimizes these features, thereby improving the ability to capture complex data patterns. By first using the linear model for preliminary processing, the features of the input data are simplified, making the subsequent nonlinear model processing more efficient. This can reduce the computational complexity and training time of the nonlinear model. On the basis of the linear model, the nonlinear model further refines the features, which helps to deeply understand the complex relationships in the data. The cascade architecture can combine simple linear relationships with complex nonlinear relationships to provide a more comprehensive explanation. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention.

[0034] Figure 1 A flowchart of a method for measuring transcutaneous oxygen partial pressure based on near-infrared spectroscopy and residual network is provided in an embodiment of the present invention.

[0035] Figure 2 Array structure diagram of an asymmetric 4-light source-6-detector array provided in an embodiment of the present invention.

[0036] Figure 3 A cross-sectional view of a 4-light source-6-detector array structure provided in an embodiment of the present invention.

[0037] Figure 4 A flowchart of the steps of constructing a linear model based on the modified Lambert-Beer law provided in an embodiment of the present invention.

[0038] Figure 5 A flowchart of the steps of constructing a nonlinear model corresponding to a transcutaneous oxygen partial pressure value based on an intermediate value and a body temperature value provided in an embodiment of the present invention.

[0039] Figure 6 A schematic diagram of an oxygen dissociation curve provided in an embodiment of the present invention.

[0040] Figure 7 A flowchart of the steps of constructing a neural network model based on a residual network and a nonlinear model provided in an embodiment of the present invention.

[0041] Figure 8 A flowchart of the steps of preprocessing sample data provided by an embodiment of the present invention.

[0042] Fig. 9 A detailed structural diagram of the ResNet neural network provided in an embodiment of the present invention.

[0043] Fig.10 A structural diagram of a residual unit provided in an embodiment of the present invention.

[0044] Fig.11 A training curve diagram of a neural network model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0046] Figure 1 The flowchart of a method for measuring transcutaneous oxygen partial pressure based on near-infrared spectroscopy and residual network is shown in FIG. In an embodiment of the present invention, a method for measuring transcutaneous oxygen partial pressure based on near-infrared spectroscopy and residual network specifically includes the following steps:

[0047] Obtain optical signals and body temperature values ​​based on near-infrared spectroscopy;

[0048] A linear model was constructed based on the modified Lambert-Beer law;

[0049] The optical signal is input into the linear model and the intermediate value is output;

[0050] A nonlinear model corresponding to the transcutaneous oxygen partial pressure value was constructed based on the intermediate value and the body temperature value;

[0051] Construct a neural network model based on residual network and nonlinear model;

[0052] Generate transcutaneous oxygen partial pressure results based on a neural network model.

[0053] In this embodiment, traditional NIR data analysis methods often rely on simple linear models or hand-designed features to interpret complex optical signals, which have many limitations in accuracy and robustness. ResNet solves the problem of gradient vanishing and gradient exploding in deep networks by introducing jump connections, allowing deeper neural networks to effectively train and capture more complex nonlinear relationships. This breakthrough technology allows neural networks to learn more sophisticated features and patterns from large amounts of optical signal data, thereby achieving high-precision prediction of TcPO2 values.

[0054] With the help of these advanced deep learning methods, the application potential of near-infrared spectroscopy in medical monitoring and diagnosis has been greatly enhanced, providing more reliable and efficient tools for clinical practice.

[0055] At the same time, linear models can quickly capture linear relationships in data, while nonlinear models are good at capturing complex nonlinear relationships. The cascade architecture can take advantage of the advantages of both to improve the performance and prediction accuracy of the overall model. The linear model can first extract the basic features in the data, and then the nonlinear model can further process and optimize these features, thereby improving the ability to capture complex data patterns. By using the linear model for preliminary processing first, the features of the input data are simplified, making the subsequent nonlinear model processing more efficient. This can reduce the computational complexity and training time of the nonlinear model. On the basis of the linear model, the nonlinear model further refines the features, which helps to deeply understand the complex relationships in the data. The cascade architecture can combine simple linear relationships with complex nonlinear relationships to provide a more comprehensive explanation.

[0056] like Figure 2 , Figure 3 As shown, as a preferred embodiment of the present invention, the step of acquiring optical signals and body temperature values ​​based on near-infrared spectroscopy includes: acquiring optical signals through a sensor with a 4-light source-6-detector array structure.

[0057] In this embodiment, the body temperature value is acquired by a temperature sensor, and the optical signal is acquired using an array design of 4 light sources and 6 detectors. The detector group consisting of a light source and 6 photodetectors of the same model adopts a reflective design to form the front-end probe of the system, and 6 layers of dense black velvet are covered around the probe to eliminate the influence of natural light. The 6 photodetectors are arranged around the light source in an asymmetric hexagram shape.

[0058] See also Figure 2 , It represents the physical distance from the detector k to the light source. The unit of measurement is mm. The distance calculation formula is as follows:

[0059]

[0060] The relationship between the detector and the light source is obtained through empirical research, examining the distance-light intensity relationship curve. Since the linear region of the curve is the same when light of the same wavelength propagates in different individual tissues, and the linear regions are different when light of different wavelengths propagates in the same individual, the optimal distance interval between the light source and the detector is limited to the common linear region. At this time, the light propagation in deep tissues is also the most stable.

[0061] Figure 3 I0 represents the incident light intensity, I1 and I4 represent the outgoing light intensity received by the detector, and D1 and D4 represent the optical path of the light in the deep blood tissue. and Indicates the straight-line distance from detectors 1 and 4 to the light source on the skin tissue.

[0062] again DPF is the differential path length factor, which can be obtained in the following two ways:

[0063] DPF is measured by spectroscopy, and DPF factors of common human tissues have been reported in the literature;

[0064] The DPF is calculated by Monte Carlo simulation, that is, the relationship between the DPF factor and tissue thickness is simulated by using the characteristic that the energy of the emitted light increases as the radial distance of the receiving position decreases within a certain range.

[0065] All photodetectors share one operational amplifier to avoid errors caused by different parameters among multiple operational amplifiers.

[0066] like Figure 4 As shown, as a preferred embodiment of the present invention, the step of constructing a linear model based on the modified Lambert-Beer law specifically includes:

[0067] Get the modified Lambert-Beer law formula;

[0068] The modified Lambert-Beer law formula is further modified based on the 4-light source-6-detector array structure and the additive combination theorem;

[0069] A linear model was constructed based on the modified formula.

[0070] In this embodiment, the linear model is constructed based on the modified Beer-Lambert formula (i.e., Lambert-Beer formula) proposed by Mark Cope et al. in 1988. We obtain the HbO2, Hb and CO2 concentrations and the light attenuation factor in the plasma of biological tissues according to the modified Lambert-Beer formula:

[0071]

[0072] Where OD stands for optical density; Attenuation (OD) stands for absorbance, often abbreviated as A; I0 stands for the intensity of the incident light beam; I k Indicates the intensity of the outgoing beam; μ a represents the absorption coefficient; d p is the distance between the light source and the detector; DPF is the differential path length factor, which depends on the absorption and scattering coefficients (μ a and μ s ) and the scattering phase function, which can be calculated by Monte Carlo simulation or measured by spectroscopy; G represents the light attenuation factor, which is a constant loss factor related to the optical properties and geometric structure of outer tissues such as skin and fat.

[0073] When n wavelengths are used to measure the concentration of m components in biological tissue plasma, let the absorbance of the i-th wavelength be A i , the concentration of the jth component is C j , the absorption coefficient of the jth component at the i-th wavelength is The modified Lambert-Beer formula can be further modified according to the additive combination theorem:

[0074]

[0075] In short:

[0076]

[0077] When there are l inter-electrode paths, the above formula can be obtained:

[0078]

[0079] In short:

[0080]

[0081] in is the absorbance at the i-th wavelength and the k-th path; is the length of the kth path; G k represents the light attenuation factor on the kth path (1≤k≤l); express and The Kronecker product between .

[0082] When the light source emits a wavelength of λ1, ..., λ n The incident light passes through the superficial skin tissue of the organism and is reflected by the detectors T1, ..., T lCapture, and finally convert n×l absorbance values It is a known physical parameter specified at various wavelengths and can be obtained by referring to relevant literature. The concentrations of the components C1,…,C m and attenuation factors G1,…,G l If it is unknown, it can be solved using the above formula.

[0083] In the step of inputting the optical signal into the linear model and outputting the intermediate value, we mainly calculate the concentration values ​​of HbO2, Hb and CO2 in plasma (respectively C Hb , ). We give the reference wavelengths of 660nm, 805nm, 880nm, and 1620nm in the optical window under the conditions of 4-6 arrays in this embodiment, and use the direct solution method to calculate the wavelengths on different detectors. C Hb , and G.

[0084] According to the modified Lambert-Beer formula, the absorbance of biological tissue is related to HbO2, Hb and dissolved CO2 (mainly containing H03 - The relationship between the concentration of ions and free CO2 and the light attenuation factor G is as follows:

[0085]

[0086] The solution is:

[0087]

[0088] like Figure 5 , Figure 6 As shown in FIG. 1 , as a preferred embodiment of the present invention, the step of constructing a nonlinear model corresponding to the transcutaneous oxygen partial pressure value based on the intermediate value and the body temperature value specifically includes:

[0089] Determine the factors that influence transcutaneous oxygen tension;

[0090] The correlation analysis of impact factors was conducted based on the Pearson correlation coefficient method;

[0091] Determine independent variables and covariates based on the analysis results;

[0092] A nonlinear model is established based on independent variables and covariates.

[0093] In this embodiment, it is known that C Hb , Under the conditions of and G, the relationship between them and TcPO2 is established, that is, a nonlinear model is established.

[0094] Based on the Hill equation that describes the kinetic relationship between hemoglobin and O2, and the oxygen dissociation curve that shows the actual result of hemoglobin and O2 binding Figure 6 ), we can see that the influencing factors of TcPO2 mainly include: C Hb , Blood pH, 2,3-DPG (2,3-diphosphoglyceraldehyde) concentration, body temperature (T), etc. After correlation analysis of the above indicators using the Pearson correlation coefficient method, we ignored pH and 2,3-DPG and C Hb As the independent variable, T and light attenuation factor G were used as covariates to establish a nonlinear model as follows:

[0095]

[0096] The 4-light source-6-detector array sensor can obtain 6 groups of C Hb , and G, the temperature sensor collects a T value each time, and the dimension of the effective input X dim(X)=25, that is:

[0097]

[0098] In the actual data set construction process, the current mainstream TcPO2 monitor in clinical practice is used to synchronously collect TcPO2 values ​​as output data Y, so the above formula can also be expressed as: Y = F(X).

[0099] like Figure 7 As shown, as a preferred embodiment of the present invention, the step of constructing a neural network model based on a residual network and a nonlinear model specifically includes:

[0100] Collect sample data;

[0101] Preprocess the sample data;

[0102] A neural network model is constructed based on the preprocessed sample data.

[0103] In this embodiment, the TcPO2 measuring device can be applied to a variety of medical application scenarios under different oxygen environments, and changes in ambient oxygen concentration and pressure will directly affect the oxygen supply to the skin and local tissues, thereby affecting the value range of the TcPO2 value.

[0104]

[0105]

[0106] The above table shows the fluctuation range of TcPO2 values ​​under different oxygen environments. It can be seen that the upper and lower limits of TcPO2 values ​​under high pressure are much greater than those under normal pressure. If the sample contains experimental data under "high pressure oxygen" environment, it will obviously significantly expand the range of characteristic values, introduce more noise and complex nonlinear relationships, and make it more difficult for the model to accurately fit the data.

[0107] Therefore, we limit the sample data to those collected under normal pressure environments, such as "inhalation of air under normal pressure", "inhalation of pure oxygen under normal pressure", and "high-flow oxygen therapy under normal pressure". This makes the data set more homogeneous, reduces the complexity of the model, and thus reduces the risk of overfitting.

[0108] At this time, the range of variation of data features is relatively small and concentrated. The model is fitted in a smaller data range, and the prediction accuracy will be higher.

[0109] According to the above model applicability definition, the TcPO2 value interval in the sample data will be strictly limited to [0, 500]. In order to ensure that the TcPO2 values ​​are evenly distributed in the range of 0 to 500, we divide [0, 500] into [0, 50), [50, 100), ..., [450, 500], and adopt a certain data balancing strategy to make the number of samples in each equally divided interval almost the same.

[0110] We set the sample size required for model training to N = 100,000 and design the sampling parameters as follows:

[0111] Sampling frequency (F s ): 5 times / second;

[0112] Sampling time (TL s ): 1.5 hours = 5400 seconds;

[0113] Sampling time window length (WL s ): 5 seconds;

[0114] Sliding step length (SL s ): 2 seconds / time;

[0115] So the number of time windows (WF s ), and the number of inputs X and outputs Y in each time window (WX s , WY s )as follows:

[0116] WX s =WY s =WL s ×F s ;

[0117] Given dim(X)=25, WX s = 25. We reshape the input data (25×25) of each time window into a single-channel two-dimensional pseudo image data, which is regarded as the sample input and recorded as X 25×25×1 ; At the same time, calculate the mean Y of Y in the time window avg As a sample output, each subject produces Samples (X 25×25×1 , Y avg ).

[0118] In the sample definition (X 25×25×1 , Y avg ), the required number of subjects is: 100000 / 2698≈37.06, that is, 38 subjects.

[0119] To collect data and make the TcPO2 values ​​widely distributed in the range of 0-500mmHg, experiments can be designed to cover this range.

[0120] Recruit subjects of different ages, genders, and health conditions, including healthy individuals and patients with respiratory or cardiovascular diseases, to increase the diversity of the data. The following is the recommended distribution of subject characteristics: 20-30 years old: 20%; 31-40 years old: 20%; 41-50 years old: 20%; 51-60 years old: 20%; 61-70 years old: 20%.

[0121] Men: 50%; Women: 50%.

[0122] Healthy individuals: 50%; individuals with respiratory or cardiovascular disease: 50%.

[0123] Smoking history: 50% with a smoking history, 50% without a smoking history.

[0124] Body mass index (BMI): Normal, overweight, and obese each account for about 33%

[0125] Sampling strategy: Design a variety of normal pressure oxygen therapy experimental environments, arrange subjects to conduct experiments under different oxygen therapy conditions, and take appropriate rest and adjustments before each experiment. In each oxygen therapy state, adjust the oxygen concentration and flow rate to cover more TcPO2 values.

[0126] Data balancing strategy: During data collection, the TcPO2 value is monitored in real time to ensure that 10 equal intervals are covered. The oxygen therapy status and parameters can be dynamically adjusted according to the number of samples in the current interval; a certain number of samples are selected from each equal interval to ensure balanced distribution; if there are insufficient samples in some equal intervals, oversampling (such as SMOTE) or undersampling methods can be used to balance the data.

[0127] Environmental control: Ensure that the data collection process is carried out in a controlled laboratory environment, the ambient temperature and humidity remain stable, and the measurement instruments and parameter calibrations are kept consistent.

[0128] Repeat the experiment for each subject several times to ensure the stability and reliability of data collection. The time points recorded for each measurement should include baseline, start, different time points and end.

[0129] like Figure 8 As shown, as a preferred embodiment of the present invention, the step of preprocessing the sample data specifically includes:

[0130] Clean the sample data;

[0131] Perform Z-Score standardization on the cleaned sample data;

[0132] The Min-Max method is used to compress each dimensional component of the sample;

[0133] The preprocessed sample data is divided into training set, validation set and test set.

[0134] In this embodiment, the data cleaning uses Box Plot and Z-Score methods to remove outliers and obvious measurement errors. Secondly, linear interpolation is used to fill the missing values ​​of each component in the original sample to ensure the quality of the data and the reliability of the model.

[0135] Data standardization and normalization: Since the ResNet neural network model we designed does not introduce a batch normalization layer (Batch Normaliztion, BN) at the beginning, all features ( C Hb , G, T and TcPO2) were Z-Score standardized to present a distribution with a mean of 0 and a standard deviation of 1.

[0136] On this basis, we use the Min-Max method to compress each dimensional component of the sample to between 0 and 1. The normalization formula is as follows:

[0137]

[0138] in and are the minimum and maximum values ​​of the i-th dimension component of the sample population of X, respectively. (i) is the original value of the i-th dimension component, For X (i) The normalized result.

[0139] The dataset was divided into training set (70%), validation set (15%), and test set (15%), and the samples in each dataset were ensured to be evenly distributed, covering all subject characteristics and oxygen therapy status.

[0140] like Fig. 9 , Fig.10 , Fig.11 As shown, as a preferred embodiment of the present invention, the neural network model includes 1 convolution layer, 3 residual blocks, 1 maximum pooling layer, 1 flattening layer, 2 fully connected layers and 1 deactivation layer.

[0141] In this embodiment, in the process of building the ResNet neural network, neural network models with 2, 3, and 4 residual blocks were tried respectively. Through experimental training, it was found that the 3-Residual Block network model was better than other models. Therefore, the final selected model consisted of 1 convolutional layer (Conv2D), 3 residual blocks (each residual block (Residual Block) contains 2 residual units (Residual Unit), each residual unit contains 1 convolutional layer, 1 batch normalization layer and ReLU activation function), 1 maximum pooling layer (MaxPooling2D), 1 flattening layer (Flatten), 2 fully connected layers (Dense) and 1 deactivation layer (Dropout).

[0142] The detailed structure of the neural network architecture is as follows Fig. 9 As shown in Figure 2, ResNet is composed of three stacked residual blocks, and each residual block is composed of two stacked residual units. The residual unit is a Conv2D+BN+ReLU arrangement structure (see Fig.10 ), and skip some layers through shortcut connections, and directly add the output from the upper layer to the output of the current stacked block to solve the problem of deep network degradation. In addition, BN is introduced in the residual unit to solve the problem of gradient diffusion and reduced training speed during training.

[0143] The residual unit consists of a residual part ( Fig.10 left) and the direct-mapped portion ( Fig.10 The output of the residual part is expressed as F(x n ), which consists of convolution operation, batch normalization and ReLU activation, directly mapping x through Shortcut n is sent directly to the output position, so the output of the residual block is:

[0144] x n+1 =ReLU(F(x n )+x n );

[0145] This can avoid the gradient vanishing phenomenon to a certain extent. In addition, it can prevent the network from degrading when the network is too deep.

[0146] When x n With x n+1 When the number of channels and features of do not match, a 1×1 convolution operation can be used on the shortcut to make the number of channels and features of the two match.

[0147] In the above network structure, ResNet first inputs the preprocessed data into the convolution layer for spatial feature extraction, which can increase the sparsity of the network and reduce the training parameters. Then the extracted features are input into multiple Residual Blocks for feature extraction again, avoiding the situation where the gradient changes extremely with the increase of network depth and the delay of prediction cycle. Adding ReLU to the Residual Block can increase the nonlinearity of the network and improve the training speed.

[0148] In the present invention, a single-channel pseudo image sample X of 25×25×1 is 25×25×1 Input into the starting Conv2D of ResNet, the convolution kernel parameters of Conv2D are 7×7×64, edge padding=3, strides=1, activation function is ReLU, and the output feature map size is 25×25×64.

[0149] In the "Residual Block 1" residual block, the convolution kernels of its internal Conv2D are all 3×3×64, padding=3, strides=1, so the input is 25×25×64 and the output is 25×25×64.

[0150] In the "ResidualBlock 2" residual block, the convolution kernels of its internal Conv2D are all 3×3×128, so the input is 25×25×64 and the output is 13×13×128.

[0151] In the "Residual Block 3" residual block, the convolution kernels of its internal Conv2D are all 3×3×256, so the input is 13×13×128 and the output is 7×7×256.

[0152] In the MaxPooling2D layer, since the convolution kernel is 2×2×256, the input is 7×7×256 and the output is 4×4×256.

[0153] After the Flatten layer, the output is 4096×1.

[0154] After the first Dense layer, the output is 1024×1.

[0155] Finally, after Dropout and the second Dense layer, the predicted value is output

[0156] The following table shows the parameter configuration of the convolutional layer in this network.

[0157]

[0158]

[0159] In the model training of this embodiment, the optimizer selects the Adam optimizer, the loss selects the MSE, the dropout is set to 0.3, the batch_size is set to 128, and the epochs is set to 100. The calculation formula of MSE is as follows:

[0160]

[0161] In the above formula, N is the sample size of each training batch. is the model prediction value, Y avg is the actual measured value.

[0162] At the same time, at the end of each epoch, the performance on the validation set is used to monitor the training process, and the batch_size and epochs values ​​of the next epoch are adjusted as needed.

[0163] During training, the warmup strategy is used to optimize the learning rate. Here is the setting:

[0164] (1) Initial learning rate: initial_lr = 1e-6;

[0165] (2) Target learning rate: target_lr = 1e-3;

[0166] (3) Number of epochs in the warmup phase: warmup_epochs = 5.

[0167] And customize the learning rate scheduler after_warmup_scheduler, which dynamically adjusts the learning rate according to the validation set loss at the beginning of each epoch.

[0168] In the process of training the model using the training set and the validation set, the curve of the loss function MSE changing with epochs is as follows Fig.11As shown in the figure, the model converges slowly during the initial training, and the MSE decreases gently with the increase of epochs, with occasional repetitions; after 10 epochs, the rate of MSE decreases faster; after reaching the 40th epoch, the change of MSE gradually stabilizes; when the MSE loss drops to about 12, it tends to be stable, and its value remains basically unchanged. Therefore, the model after 100 epochs of iterative training is used as the final prediction model of TcPO2.

[0169] The evaluation indicators selected in the embodiment include mean square error (MSE), mean absolute error (MAE) and goodness of fit R 2 Among them, MAE and R 2 The calculation formula is as follows:

[0170]

[0171] in, is the actual measured value Y avg The performance evaluation of the trained ResNet model on the test set is shown in the following table:

[0172]

[0173] In summary, under the premise of using near-infrared spectroscopy to collect the absorbance of each component closely related to TcPO2 in blood tissue, the algorithm flow of real-time and non-invasive measurement of TcPO2 value based on ResNet is as follows:

[0174] Input: absorbance sequence of each component in blood Temperature value sequence {T};

[0175] Output: TcPO2 value sequence to be measured

[0176] (1) and T are input into the linear model in sequence, and the effective input value of the ResNet model is solved using formula (7)

[0177] (2) Preprocess the input sequence X to obtain a 25×25×1 pseudo image data sequence {X 25×25×1};

[0178] (3) {X 25×25×1} are sequentially input into the trained 3-Residual Block structure ResNet model (see Figure 5 ), and obtain the TcPO2 prediction value sequence

[0179] The ResNet model with a 3-Residual Block structure contains 13 convolutional layers, 1 MaxPooling layer, 2 Dense layers, 1 Dropout layer and a Flatten layer, with a total of more than 4 million weight parameters.

[0180] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for measuring transcutaneous oxygen partial pressure based on near-infrared spectroscopy and residual network, characterized in that: The method comprises: Obtain optical signals and body temperature values ​​based on near-infrared spectroscopy; A linear model was constructed based on the modified Lambert-Beer law; The optical signal is input into the linear model and the intermediate value is output; A nonlinear model corresponding to the transcutaneous oxygen partial pressure value was constructed based on the intermediate value and the body temperature value; Construct a neural network model based on residual network and nonlinear model; Generate transcutaneous oxygen partial pressure results based on the neural network model; The step of constructing a nonlinear model corresponding to the transcutaneous oxygen partial pressure value based on the intermediate value and the body temperature value specifically includes: Determine the factors that influence transcutaneous oxygen tension; The correlation analysis of impact factors was conducted based on the Pearson correlation coefficient method; Determine independent variables and covariates based on the analysis results; Establish nonlinear models based on independent variables and covariates; The steps of constructing a neural network model based on a residual network and a nonlinear model specifically include: Collect sample data; Preprocess the sample data; A neural network model is constructed based on the preprocessed sample data.

2. The method for measuring transcutaneous oxygen partial pressure based on near infrared spectroscopy and residual network according to claim 1, characterized in that: The steps of acquiring optical signals and body temperature values ​​based on near-infrared spectroscopy include: acquiring optical signals through a sensor with a 4-light source-6-detector array structure, and acquiring body temperature values ​​through a temperature sensor.

3. The method for measuring transcutaneous oxygen partial pressure based on near infrared spectroscopy and residual network according to claim 2, characterized in that: The step of constructing a linear model based on the modified Lambert-Beer law specifically includes: Get the modified Lambert-Beer law formula; The modified Lambert-Beer law formula is further modified based on the 4-light source-6-detector array structure and the additive combination theorem; A linear model was constructed based on the modified formula.

4. The method for measuring transcutaneous oxygen partial pressure based on near infrared spectroscopy and residual network according to claim 1, characterized in that: The step of preprocessing the sample data specifically includes: Clean the sample data; Perform Z-Score standardization on the cleaned sample data; The Min-Max method is used to compress each dimensional component of the sample; The preprocessed sample data is divided into training set, validation set and test set.

5. The method for measuring transcutaneous oxygen partial pressure based on near infrared spectroscopy and residual network according to claim 1, characterized in that: The neural network model includes 1 convolutional layer, 3 residual blocks, 1 maximum pooling layer, 1 flattening layer, 2 fully connected layers and 1 inactivation layer.

Images