Pulse oximetry method based on image processing

Abstract

The present application relates to the technical field of blood oxygen detection, and more particularly to a pulse blood oxygen saturation detection method based on image processing, which introduces a transfer learning strategy and a domain adaptation network, can be adjusted according to the individual physiological parameters of the user, thereby improving the applicability to different populations. In particular for dark-skinned population, the present application significantly reduces the measurement error, reduces the error rate by more than 30%. The present application designs a set of real-time calculation mechanism of current signal quality index Q, and combines the Bayesian filtering model for dynamic calibration, so as to ensure that reliable blood oxygen saturation estimation value can be obtained even under low signal-to-noise ratio condition. When the signal quality decreases, the historical data is compensated for the current detection value through the LSTM network prediction module, so as to further ensure the accuracy and continuity of the result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of blood oxygen detection, and particularly relates to a pulse blood oxygen saturation detection method based on image processing. BACKGROUND

[0002] At present, traditional blood oxygen saturation detection mainly relies on photoplethysmography (PPG), and red light / infrared light absorption changes are measured through a finger probe. Typical representatives include the Signal Extraction Technology of Masimo Company (patent US20150351697A1) and the optical sensor scheme of Apple Watch (patent US20200323435A1).

[0003] However, the above prior art has three major defects:

[0004] 1. Contact measurement limitation: poor comfort due to probe compression, and easy skin damage during long-term monitoring.

[0005] 2. Insufficient anti-interference ability: motion artifacts and environmental light noise seriously affect the measurement accuracy (see IEEE TBME research data in 2021, error up to ±5%).

[0006] 3. Lack of personalized adaptation: individual differences such as skin color and subcutaneous fat thickness are not considered (Nature Biomedical Engineering pointed out in 2022 that the error of deep-skinned people increased by 30%).

[0007] Therefore, a pulse blood oxygen saturation detection method based on image processing is needed to solve the above problems. SUMMARY

[0008] The present application provides a pulse blood oxygen saturation detection method based on image processing. The present application introduces a transfer learning strategy and a domain adaptation network, which can be adjusted according to the individual physiological parameters (such as skin color, age, body fat rate, etc.) of the user, thereby improving the applicability to different populations. In particular for deep-skinned people, the present application significantly reduces the measurement error, reducing the error rate by more than 30%.

[0009] The technical scheme adopted by the present application to solve the above technical problems is: a pulse blood oxygen saturation detection method based on image processing, comprising the following steps:

[0010] S1, acquiring a fingertip video stream of a measured person through a multispectral imaging device, and synchronously obtaining an initial blood oxygen saturation reference value;

[0011] S2, spatio-temporal domain preprocessing is performed on the fingertip video stream, including: using adaptive weighted guided filtering to eliminate ambient light interference, and constructing a motion artifact compensation model to eliminate micro-motion interference;

[0012] S3, double-channel feature parameters in the preprocessed video are extracted, including: time-frequency features of PPG signals are extracted through wavelet transform, and spatial features of skin microvessel distribution are extracted through principal component analysis;

[0013] S4, a double-flow convolutional neural network model is constructed, wherein a ResNet-50 is used in the first branch to extract time-frequency features, a U-Net is used in the second branch to extract spatial features, and attention mechanism is used to fuse the features of the two branches;

[0014] S5, a current signal quality index Q is calculated in real time, and the Q value is generated by fusing three parameters of signal-to-noise ratio, waveform periodicity and skin light transmittance through entropy weight method;

[0015] S6, a dynamic calibration module is established, and an initial blood oxygen saturation reference value, a real-time Q value and a neural network output value are input into a Bayesian filtering model for iterative optimization;

[0016] S7, when the signal quality index Q is lower than a threshold value, a time series prediction module based on LSTM network is triggered to compensate the current detection value by using historical data;

[0017] S8, a contribution degree distribution graph of each feature parameter is generated through an explainability analysis module, and the weight coefficients of each input parameter are dynamically adjusted;

[0018] S9, a transfer learning strategy is used to construct a domain adaptive network by using personalized physiological parameters, so as to realize online fine-tuning of model parameters;

[0019] S10, a final blood oxygen saturation value is output, and hierarchical prompt information including confidence interval and abnormal fluctuation warning is generated synchronously.

[0020] Further, the motion artifact compensation model in step S2 specifically includes:

[0021] S21, a hand motion trajectory prediction sub-model based on an inertial measurement unit is constructed;

[0022] S22, a spatio-temporal mapping relationship between video frame sequence and motion trajectory is established;

[0023] S23, a motion vector field is calculated by using an optical flow method, and reverse compensation is performed in combination with the trajectory prediction result;

[0024] S24, the quality of the compensated video sequence is verified by using a convolutional recurrent network.

[0025] Further, the establishment method of the attention mechanism fusion in step S4 specifically includes:

[0026] S41, channel attention is added to the time-frequency feature map to generate a feature heat map;

[0027] S42, spatial attention is added to the spatial feature map to generate a region attention distribution map;

[0028] S43, dynamically fuse the double-channel features through a gating mechanism, and the fusion weight is dynamically adjusted by the real-time Q value. Further, the specific establishment method of the Bayesian filtering model in step S6 is:

[0029] S61, a state space model containing device error terms and physiological fluctuation terms is established;

[0030] S62, the neural network output is used as an observation value, and the initial reference value is used as a prior distribution;

[0031] S63, adjust the process noise covariance matrix using the Q value;

[0032] S64, use a particle filter algorithm to estimate the posterior probability.

[0033] Further, the construction method of the time series prediction module in step S7 specifically includes:

[0034] S71, construct a bidirectional LSTM network architecture, and the input window contains time series features of the previous 60 seconds;

[0035] S72, introduce a time convolution network to extract multi-scale time series patterns;

[0036] S73, enhance the robustness of the prediction result through an adversarial training strategy;

[0037] S74, design a residual connection structure to compensate for prediction errors.

[0038] Further, the establishment method of the explainability analysis module in step S8 includes the following steps:

[0039] S81, use integral gradient method to calculate the contribution of input features;

[0040] S82, locate the feature area through class activation mapping;

[0041] S83, establish a dynamic association rule between contribution and Q value;

[0042] S84, automatically generate model optimization suggestions according to the contribution distribution.

[0043] Further, the construction method of the domain adaptation network in step S9 specifically includes:

[0044] S91, construct a domain conversion module based on a generative adversarial network;

[0045] S92, design a personalized feature encoder containing skin color, age, and body fat percentage;

[0046] S93, implement sample rapid adaptation using meta-learning strategy;

[0047] S94, enhance cross-domain feature representation capability through contrastive learning.

[0048] Further, the generation of hierarchical prompt information in step S10 includes:

[0049] S101, establish a multi-level early warning threshold system based on clinical guidelines;

[0050] S102, identify abnormal states by combining trend analysis and fluctuation pattern recognition;

[0051] S103, generate dynamic confidence intervals containing time evolution characteristics;

[0052] S104, integrate multi-modal alerting methods of voice, vision, and touch.

[0053] The advantages of the present application are:

[0054] 1. The present application uses a non-contact multi-spectral imaging device to collect fingertip video stream, without direct contact with the skin, greatly improving the comfort of the user, suitable for long-term continuous monitoring. Compared with the traditional finger probe, the present method avoids the problem of skin damage caused by compression.

[0055] 2. The present application eliminates environmental light interference through adaptive weighted guided filtering, and constructs a motion artifact compensation model to eliminate micro-motion interference, significantly improving the detection stability and accuracy in complex environments.

[0056] 3. The present application introduces a transfer learning strategy and a domain adaptation network, which can be adjusted according to the user's individual physiological parameters (such as skin color, age, body fat percentage, etc.), thereby improving the applicability to different populations. Especially for dark-skinned people, the present application significantly reduces the measurement error, reducing the error rate by more than 30%.

[0057] 4. The present application designs a mechanism for real-time calculation of the current signal quality index Q, and combines a Bayesian filtering model for dynamic calibration, ensuring that reliable blood oxygen saturation estimates can be obtained even under low signal-to-noise ratio conditions. When the signal quality decreases, the LSTM network prediction module uses historical data to compensate for the current detection value, further ensuring the accuracy and continuity of the results.

[0058] 5. In order to facilitate understanding and improve the performance of the algorithm, the contribution distribution diagram of each feature parameter can be generated by the explainability analysis module, and the input parameter weight is automatically adjusted accordingly. This transparent processing method helps to quickly locate the problem and optimize the model, thereby continuously improving the overall performance of the system.

[0059] 6. The present application not only provides accurate blood oxygen saturation value output, but also synchronously generates graded prompt information containing confidence interval and abnormal fluctuation warning. Users can not only understand their health status, but also receive timely warning of potential risks. DETAILED DESCRIPTION

[0060] The technical solutions of the present application will be described below in a clear and complete manner. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0061] Embodiment 1:

[0062] The present application provides a pulse blood oxygen saturation detection method based on image processing,

[0063] A pulse blood oxygen saturation detection method based on image processing comprises the following steps:

[0064] S1, acquiring the fingertip video stream of the measured person by a multispectral imaging device, and synchronously obtaining the initial blood oxygen saturation reference value. Specifically, the multispectral imaging device uses a customized multispectral camera (component list see Table 1), which includes: 8 groups of narrowband LED arrays (center wavelength: 530nm / 660nm / 810nm / 940nm); a global shutter CMOS sensor (frame rate 60fps, resolution 1280x960); a six-axis IMU (MPU-6050, sampling rate 200Hz).

[0065] Table 1 Parameters of multispectral imaging device (multispectral camera)

[0066]

[0067] S2, performing spatiotemporal domain preprocessing on the fingertip video stream, including: using adaptive weighted guided filtering to eliminate environmental light interference, and constructing a motion artifact compensation model to eliminate micro-motion interference. Specifically, the adaptive weighted guided filtering includes:

[0068] The light intensity distribution model is established according to the following formula: (x, y, t) = a(t)·R(x, y, t) + b(t)·L env(x, y, t), where R is the reflection component, representing the intensity of light reflected by the object surface at position x, y and time t, L env is the ambient light component, representing the intensity of background or scattered light present in the environment at position x, y and time t, the weight coefficients a, b are solved by iterative optimization; where x is the horizontal position coordinate, y is the vertical position coordinate, these two variables are used to describe the position of a certain point in the image or scene. The variable t represents time. Specifically, t represents the time point at a certain time. In video streams, lighting conditions may change over time, so the time variable t is introduced to describe the lighting intensity distribution at different time points. In this way, the most appropriate weight coefficients a, b at different times and positions can be obtained, so that the lighting intensity distribution can be modeled more accurately.

[0069] Further, the motion artifact compensation model in step S2 specifically includes:

[0070] S21, a hand motion trajectory prediction sub-model based on an inertial measurement unit is constructed, specifically including:

[0071] S211, data acquisition: use an inertial measurement unit (IMU) device (such as an accelerometer, gyroscope, etc.) fixed on the hand of the measured person to record the motion data of the hand in real time. The data includes acceleration, angular velocity and other parameters, which will be used to predict the motion trajectory of the hand.

[0072] S212, feature extraction: extract key features from IMU data, such as acceleration rate of change, angular velocity trend, etc. These features will be used as input to train the hand motion trajectory prediction model.

[0073] S213, model training: use machine learning algorithms (such as long short-term memory network LSTM or convolutional neural network CNN) to train the extracted features to establish a sub-model that can predict the hand motion trajectory.

[0074] S214, prediction output: in actual application, use the trained model to predict the hand motion trajectory in the future period of time according to the current IMU data.

[0075] S22, establish the spatio-temporal mapping relationship between video frame sequence and motion trajectory, specifically including:

[0076] S223, time synchronization: ensure that the timestamps of IMU data and video frames are completely synchronized, so as to accurately correspond each frame of video and the corresponding hand motion state.

[0077] S224, spatial mapping: According to the hand motion trajectory prediction result, calculate the position change of each frame of video corresponding to the fingertip. Space-time mapping matrix: Construct a space-time mapping matrix, which records the offset of the fingertip position in each frame of video relative to the reference position, which is used to adjust the video frame to make it more consistent with the real fingertip position.

[0078] S23, calculate the motion vector field using the optical flow method, and combine the trajectory prediction result for reverse compensation. Specifically, it includes:

[0079] S231, optical flow calculation: use optical flow method (Optical Flow) to calculate the motion vector field between adjacent frames, determine the moving direction and speed of each pixel point in each frame. Optical flow method can be realized by Lucas-Kanade algorithm or Farneback algorithm.

[0080] S232, motion vector correction: combine the motion vector field calculated by the optical flow method with the hand motion trajectory prediction result. For each pixel point, according to its motion vector and predicted motion trajectory, calculate the reverse displacement amount that should be compensated.

[0081] S233, reverse compensation: apply reverse compensation algorithm to adjust each pixel point in each frame of video according to the calculated reverse displacement amount, so as to eliminate the artifacts caused by hand motion.

[0082] S24, verify the quality of the compensated video sequence through convolutional recurrent network, specifically including:

[0083] S241, construct convolutional recurrent network (ConvRNN): design a convolutional recurrent network (such as an architecture combining convolutional layer and LSTM layer) for evaluating the quality of the compensated video sequence. The input is the compensated video frame sequence, and the output is the quality score.

[0084] S242, train the network: use high-quality (no motion artifacts, high definition and uniform illumination) video frame sequence as positive sample, low-quality (with artifacts, low definition and non-uniform illumination) video frame sequence as negative sample, train the convolutional recurrent network. The goal is to let the network be able to distinguish between high-quality and low-quality video frame sequences and give corresponding scores.

[0085] S243, quality evaluation: in practical application, input the compensated video frame sequence into the trained convolutional recurrent network to get its quality score. If the score is lower than the pre-set threshold, it is considered that the compensation effect is not good, and the compensation algorithm needs to be further optimized; if the score is higher than the pre-set threshold, it is considered that the compensation is effective, and the subsequent processing can continue.

[0086] S3, extract dual-channel feature parameters in pre-processed video, including: extracting time-frequency features of PPG signal through wavelet transform, and extracting spatial features of skin microvessel distribution through principal component analysis; specifically, extracting time-frequency features of PPG signal through wavelet transform includes:

[0087] S31, data preparation: extracting image sequences of fingertip regions from pre-processed video frame sequences.

[0088] S32, wavelet transform: using Morlet wavelet or other suitable wavelet basis functions to perform time-frequency analysis on image sequences to extract time-frequency features of PPG signal. These features include frequency components, amplitude changes, etc. The frequency range is set to 0.5Hz to 3Hz, and continuous wavelet transform (CWT) is used to extract these frequency components and generate time-frequency graphs.

[0089] wherein the spatial features of the skin microvessel distribution extracted by principal component analysis include:

[0090] S33, data preparation: similarly, extracting image sequences of fingertip regions from pre-processed video frame sequences.

[0091] S34, PCA analysis: applying principal component analysis (PCA) method to extract main spatial features in image sequences, which can reflect the distribution of skin microvessels.

[0092] S4, constructing a dual-stream convolutional neural network model, wherein the first branch uses ResNet-50 to extract time-frequency features, uses ResNet-50 as the basic network structure, and the input is the time-frequency graph generated after wavelet transform, then pre-trains the ResNet-50 model on a large-scale dataset, fine-tunes it for specific tasks, and generates time-frequency feature maps. The second branch uses U-Net to extract spatial features, uses U-Net as the basic network structure, and the input is the spatial feature map extracted by principal component analysis (PCA). Pre-train the U-Net model on a large-scale dataset, fine-tune it for specific tasks to generate spatial feature maps, then fuse the dual-branch features through attention mechanism, wherein the establishment method of attention mechanism fusion includes:

[0093] S41, channel attention weighting is performed on the time-frequency feature map to generate a feature heat map;

[0094] S42, spatial attention weighting is performed on the spatial feature map to generate a region attention distribution map;

[0095] S43, dynamically fusing dual-channel features through a gating mechanism, and the fusion weight is dynamically adjusted by real-time Q value.

[0096] S5, calculate the current signal quality index Q in real time, the Q value is generated by the entropy weight method from the fusion of three parameters of signal-to-noise ratio, waveform periodicity and skin light transmittance, specifically, including:

[0097] S51, calculate the signal-to-noise ratio (SNR), which is calculated by the following formula: Where P signal is the signal power, P noise is the noise power, the power ratio of the signal part and the noise part is calculated by frequency domain analysis of the signal.

[0098] S52, calculate the waveform periodicity: determine the main period of the waveform by using the existing autocorrelation function or Fourier transform.

[0099] S53, calculate the skin light transmittance: according to the data collected by the multispectral imaging device, calculate the light transmittance under different wavelengths; use different wavelengths (such as 530nm, 660nm, 810nm, 940nm) of narrowband LED array to calculate the light transmittance under each wavelength.

[0100] S54, entropy weight fusion: use entropy weight method to weight and fuse the three parameters of signal-to-noise ratio, waveform periodicity and skin light transmittance to get the final Q value. Calculate the entropy value of each parameter, and then determine its weight, and finally get the Q value by weighted summation.

[0101] S6, establish a dynamic calibration module, input the initial blood oxygen saturation reference value, real-time Q value and neural network output value into the Bayesian filtering model for iterative optimization, wherein the specific establishment method of the Bayesian filtering model is:

[0102] S61, establish a state space model containing device error items and physiological fluctuation items, specifically, the formula is: x t =A xt-1 +B ut +w t , wherein x t is a state vector, A is a state transition matrix, B is a control input matrix, ut is a control input vector, and wt is a process noise.

[0103] S62, take the neural network output as the observation value and the initial reference value as the prior distribution.

[0104] S63, adjust the process noise covariance matrix using the Q value, and dynamically adjust the noise matrix according to the change of the Q value to reflect the change of the signal quality.

[0105] S64, use particle filtering algorithm to estimate the posterior probability, and get the optimal state estimation value by resampling and updating particles.

[0106] S7, triggering the LSTM network-based time series prediction module to compensate for the current detection value using historical data when the signal quality index Q is below the threshold, wherein the construction method of the time series prediction module in step S7 specifically includes:

[0107] S71, constructing a bidirectional LSTM network architecture, and the input window contains time series features of the previous 60 seconds;

[0108] S72, introducing a time convolution network to extract multi-scale time series patterns;

[0109] S73, enhancing the robustness of the prediction result through an adversarial training strategy;

[0110] S74, designing a residual connection structure to compensate for prediction errors, specifically including setting a basic bidirectional LSTM (Bi-LSTM) unit for processing input time series data, setting the input time series length as T and the feature dimension as d, and implementing it through Python:

[0111]

[0112]

[0113] In combination with the above components, a complete time series prediction model can be constructed, which includes a bidirectional LSTM layer, a residual connection, and a final prediction layer. In a deep network, the gradient may become very small during the backpropagation process, making it difficult to effectively update the parameters of the earlier layers. By introducing a residual connection, the gradient can be directly passed to the earlier layers, alleviating the gradient vanishing problem.

[0114] S8, generating a contribution distribution map for each feature parameter through an explainability analysis module, wherein the feature parameters include: physiological parameters: heart rate, blood pressure, respiratory rate, body temperature. Environmental factors: temperature, humidity. Signal quality related indicators: signal strength, noise level, signal-to-noise ratio. Time series data: time stamp of data collection, periodic change of data (such as daily or weekly change pattern), and then dynamically adjusting the weight coefficients of each input parameter, specifically including the following steps:

[0115] S81, using the integrated gradient method (Integrated Gradients) to calculate the contribution of each input feature to the model output. By differentiating the input features, the impact on the output is calculated.

[0116] S82, locate the feature area through class activation mapping: use the class activation mapping (CAM) technology to locate the feature area that has the greatest impact on the classification result. By weighted averaging the output of the convolution layer, a class activation map is generated.

[0117] S83, Establish dynamic association rules between contribution degree and Q value: According to the distribution of contribution degree, establish dynamic association rules with Q value. Through statistical analysis, determine the relationship between contribution degree and Q value. Specifically, it includes:

[0118] S831, Calculate the contribution of each feature to the model prediction result by integral gradient method

[0119] S832, Select a baseline input (such as all-zero vector), interpolate between baseline input and actual input to generate a series of intermediate inputs.

[0120] S833, Calculate the gradient of each intermediate input, and integrate all gradients. The final contribution is the result of integrated gradient.

[0121] S834, Calculate the Pearson correlation coefficient between the contribution of each feature and Q value to measure the linear relationship between them.

[0122] S835, Use existing linear regression models to fit the relationship between contribution degree and Q value, which can help understand which features have a significant impact on Q value.

[0123] S836, Use clustering algorithms (such as K-means, hierarchical clustering) to divide samples into different groups, and then analyze the relationship between contribution degree and Q value within each group.

[0124] Through the above steps, dynamic association rules between contribution degree and Q value can be established, so as to better understand and optimize the model in the blood oxygen saturation detection system.

[0125] S84, Automatically generate model optimization suggestions based on contribution degree distribution: According to the distribution of contribution degree, automatically adjust model parameters and propose optimization suggestions. Provide interpretable analysis results to help understand the working principle of the model and optimize it.

[0126] S9, Adopt transfer learning strategy, use personalized physiological parameters to build domain adaptive network, realize online fine-tuning of model parameters, specifically including:

[0127] S91, Construct domain conversion module based on generative adversarial network;

[0128] S92, Design personalized feature encoder containing skin color, age, and body fat rate;

[0129] S93, Use meta-learning strategy to realize fast adaptation of samples;

[0130] S94, Enhance cross-domain feature representation ability through contrastive learning.

[0131] S10, output the final blood oxygen saturation value, and synchronously generate hierarchical prompt information containing a confidence interval and an abnormal fluctuation warning, wherein the generation of the hierarchical prompt information comprises:

[0132] S101, a multi-level warning threshold system based on clinical guidelines is established, specifically, different levels of warning thresholds are set according to clinical guidelines. For example, a first-level warning is issued when SpO2 is lower than 90%, and a second-level warning is issued when SpO2 is lower than 85%.

[0133] S102, an abnormal fluctuation mode is identified through time series analysis. ARIMA model or other existing time series analysis methods are used to detect abnormal fluctuations.

[0134] S103, a dynamic confidence interval is calculated to reflect the time evolution characteristics. The confidence interval is generated by Bayesian inference method.

[0135] S104, through voice prompt, visual alarm (such as flashing light), tactile feedback (such as vibration) and other ways, remind the user to pay attention to the potential health risk.

[0136] The adaptive weighted guided filtering of the application improves the ambient light noise suppression rate to 92.3% (compared with 68.5% of the traditional morphological filtering); the motion artifact compensation model improves the signal-to-noise ratio (SNR) from 4.2dB to 12.7dB under micro-motion interference. Breakthrough in measurement accuracy: the dual-flow neural network feature fusion reduces the SpO2 estimation error by 42%; the LSTM time series prediction module still maintains 85% prediction accuracy when the signal is interrupted. Personalized adaptation enhancement: the domain adaptation network reduces the measurement deviation across skin color groups from ±3.1% to ±

[0137] 1.5%; the explainability analysis module automatically generates 98.2% effective model optimization suggestions.

[0138] The detection method of the application is deployed on Huawei Watch 5Pro hardware platform: processor: HiSilicon Hi3861V100, special NPU acceleration dual-flow network inference; power consumption optimization: dynamic Q value control reduces the average power consumption from 12.3mW to 7.8mW; user interface: real-time display of SpO2 trend chart (sampling interval 1 second), abnormal fluctuation triggers vibration warning. Deployment test in hospital ICU: data acquisition: 50 postoperative patients, continuous monitoring for 72 hours;

[0139] Performance indicators are:

[0140]

[0141] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions recorded in the above embodiments can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for detecting pulse oximetry based on image processing, characterized in that, The method comprises the following steps: S1, acquiring a fingertip video stream of a measured person by a multispectral imaging device, and synchronously obtaining an initial blood oxygen saturation reference value; S2, performing spatio-temporal domain preprocessing on the fingertip video stream, including: eliminating environmental light interference by using adaptive weighted guided filtering, and constructing a motion artifact compensation model to eliminate micro-motion interference; S3, extracting double-channel feature parameters in the preprocessed video, including: extracting time-frequency features of a PPG signal by wavelet transform, and extracting spatial features of skin microvessel distribution by principal component analysis; S4, constructing a double-flow convolutional neural network model, wherein a first branch extracts time-frequency features by using ResNet-50, a second branch extracts spatial features by using U-Net, and double-branch features are fused by an attention mechanism; S5, calculating a current signal quality index Q in real time, wherein the Q value is generated by fusing three parameters of signal-to-noise ratio, waveform periodicity and skin light transmittance by an entropy weight method; S6, establishing a dynamic calibration module, inputting the initial blood oxygen saturation reference value, the real-time Q value and the neural network output value into a Bayesian filtering model for iterative optimization; S7, when the signal quality index Q is lower than a threshold value, triggering a time series prediction module based on an LSTM network, and compensating the current detection value by using historical data; S8, generating a contribution degree distribution diagram of each feature parameter by an explainability analysis module, and dynamically adjusting weight coefficients of each input parameter; S9, using a transfer learning strategy, constructing a domain adaptive network by using individual physiological parameters, and realizing online fine-tuning of model parameters; S10, outputting a final blood oxygen saturation value, and synchronously generating graded prompt information containing a confidence interval and an abnormal fluctuation warning.

2. The image processing based pulse oximetry detection method of claim 1, wherein, The motion artifact compensation model in step S2 specifically comprises: S21, constructing a hand motion trajectory prediction sub-model based on an inertial measurement unit; S22, establishing a spatio-temporal mapping relationship between a video frame sequence and a motion trajectory; S23, calculating a motion vector field by using an optical flow method, and performing reverse compensation combined with a trajectory prediction result; S24, verifying the quality of the compensated video sequence by a convolutional recurrent network.

3. The image processing based pulse oximetry detection method of claim 1, wherein, The establishment method of the attention mechanism fusion in step S4 specifically comprises: S41, performing channel attention weighting on a time-frequency feature map, and generating a feature heat map; S42, performing spatial attention weighting on a spatial feature map, and generating a regional attention degree distribution map; S43, dynamically fusing double-channel features by a gating mechanism, and dynamically adjusting a fusion weight by a real-time Q value.

4. The image processing based pulse oximetry detection method of claim 1, wherein, The specific establishment method of the Bayesian filtering model in step S6 is as follows: S61, establishing a state space model containing a device error term and a physiological fluctuation term; S62, taking a neural network output as an observation value, and taking an initial reference value as a prior distribution; S63, adjusting a process noise covariance matrix by using a Q value; S64, performing posterior probability estimation by using a particle filtering algorithm.

5. The image processing based pulse oximetry detection method of claim 1, wherein, The construction method of the time series prediction module in step S7 specifically comprises: S71, constructing a bidirectional LSTM network architecture, and inputting a window containing time series features of the previous 60 seconds; S72, introducing a time convolution network to extract multi-scale time series patterns; S73, enhancing the robustness of a prediction result by using an adversarial training strategy; S74, designing a residual connection structure to compensate for prediction errors.

6. The image processing based pulse oximetry detection method of claim 1, wherein, The method for establishing the explainability analysis module in step S8 includes the following steps: S81, calculating the contribution degree of the input features by using integral gradient method; S82, locating the feature region through class activation mapping; S83, establishing dynamic association rules between the contribution degree and Q value; S84, automatically generating model optimization suggestions according to the contribution degree distribution.

7. The image processing based pulse oximetry detection method of claim 1, wherein, The method for constructing the domain adaptive network in step S9 specifically includes: S91, constructing a domain conversion module based on a generative adversarial network; S92, designing a personalized feature encoder containing skin color, age, and body fat rate; S93, realizing sample rapid adaptation by using meta-learning strategy; S94, enhancing cross-domain feature representation ability through contrastive learning.

8. The image processing based pulse oximetry detection method of claim 1, wherein, The generation of hierarchical prompt information in step S10 includes: S101, establishing a multi-level early warning threshold system based on clinical guidelines; S102, identifying abnormal states by combining trend analysis and fluctuation pattern recognition; S103, generating dynamic confidence intervals containing time evolution characteristics; S104, integrating multi-modal warning methods of voice, vision, and touch.

Images