De-noising model training method, de-noising method and electronic device
By simulating the generation of training sample sets and training a denoising model using a staged loss function, the problems of signal damage and insufficient accuracy in fNIRS signal denoising methods are solved, achieving high-fidelity and robust denoising effects.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HEBEI GEO UNIVERSITY
- Filing Date
- 2026-04-29
- Publication Date
- 2026-07-10
AI Technical Summary
In existing technologies, functional near-infrared spectral signal (fNIRS) denoising methods are prone to damaging the real signal, and deep learning methods rely on large-scale paired training sample sets which are difficult to obtain, resulting in insufficient model accuracy.
Multiple pure hemodynamic response signals were generated by simulation and noise was added to form a training sample set. The denoising model was trained in stages using different loss functions, including a first loss function for initial signal feature restoration and a second loss function for further optimization of signal smoothness and noise suppression. A denoising autoencoder with a U-Net structure was used for training.
It achieves the preservation of signal amplitude, morphology, and timing characteristics while suppressing noise, improving the signal fidelity and robustness after denoising and significantly reducing residual artifacts.
Smart Images
Figure CN122364918A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of biomedical signal processing technology, and in particular to a denoising model training method, a denoising method, and an electronic device. Background Technology
[0002] Functional near-infrared spectroscopy (fNIRS) technology enables non-invasive monitoring of brain activity by measuring changes in hemoglobin concentration in the cerebral cortex. Its portability and high temporal resolution have led to its widespread application in neuroscience, rehabilitation medicine, and human-computer interaction. However, fNIRS signals are highly susceptible to motion artifacts (MA) generated by the subject's body movements (especially head movements). The amplitude of these motion artifacts is typically much larger than the hemodynamic response (HRF) signal caused by genuine brain activity, often completely obscuring the useful signal and severely impacting the accuracy of subsequent data analysis. Therefore, denoising of fNIRS signals is necessary.
[0003] Existing technologies typically employ bandpass filtering, principal component analysis (PCA), and wavelet denoising methods for noise reduction. However, these methods can damage the true HRF signal while removing noise, leading to a decrease in signal fidelity. Deep learning methods can also be used for noise reduction, but these methods heavily rely on large-scale, paired training sample sets of noisy and clean signals. Obtaining such training sample sets for fNIRS signals is difficult, significantly impacting model accuracy. Summary of the Invention
[0004] This invention provides a denoising model training method, a denoising method, and an electronic device to solve the problems of signal damage or insufficient denoising accuracy in existing denoising methods.
[0005] In a first aspect, embodiments of the present invention provide a method for training a denoising model, comprising: Multiple pure hemodynamic response signals are simulated and generated, and noise is added to each pure hemodynamic response signal to form multiple noisy functional near-infrared spectral signals; each pure hemodynamic response signal corresponds one-to-one with each noisy functional near-infrared spectral signal to form a training sample set. Using the training sample set, the initial denoising model is trained with the first loss function to obtain the intermediate denoising model; The intermediate denoising model is trained using the training sample set and the second loss function to obtain the target denoising model; The first loss function is different from the second loss function.
[0006] The second aspect provides a functional near-infrared spectral signal denoising method, including: The functional near-infrared spectral signal to be denoised is input into the target denoising model provided in the first aspect embodiment above to obtain the denoised functional near-infrared spectral signal.
[0007] Thirdly, embodiments of the present invention provide an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the denoising model training method in the first aspect or any possible implementation of the first aspect, and the functional near-infrared spectral signal denoising method in the second aspect or any possible implementation of the second aspect.
[0008] This invention provides a denoising model training method, a denoising method, and an electronic device. The denoising model training method includes: simulating and generating multiple pure hemodynamic response signals, and adding noise to each pure hemodynamic response signal to form multiple noisy functional near-infrared spectral signals; wherein each pure hemodynamic response signal corresponds one-to-one with each noisy functional near-infrared spectral signal to form a training sample set; using the training sample set, training an initial denoising model with a first loss function to obtain an intermediate denoising model; using the training sample set, training the intermediate denoising model with a second loss function to obtain a target denoising model; wherein the first loss function and the second loss function are different. This invention generates training sample sets based on data-driven methods, which can generate unlimited, diverse, and highly realistic training data. Furthermore, by training the denoising model using different loss functions in stages, this application enables the model to effectively suppress noise in functional near-infrared spectral signals while better preserving the amplitude, morphology, and temporal characteristics of the signal, thus improving the fidelity and robustness of the denoised signal. Attached Figure Description
[0009] Figure 1 This is a flowchart illustrating the implementation of a denoising model training method provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the generation of noisy functional near-infrared spectral signals by superposition provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of a denoising autoencoder with a U-Net structure provided in an embodiment of the present invention; Figure 4 This is a denoising effect diagram of the functional near-infrared spectral signal denoising method provided in the embodiments of the present invention; Figure 5This is a comparison of the average residual artifacts of the functional near-infrared spectral signal after denoising using the functional near-infrared spectral signal denoising method provided in the embodiments of the present invention, after denoising using the wavelet denoising model, and the original functional near-infrared spectral signal to be denoised. Figure 6 This is a schematic diagram of the structure of the denoising model training device provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation
[0010] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0011] See Figure 1 The diagram illustrates the implementation flowchart of a denoising model training method provided by an embodiment of the present invention, which is described in detail below: refer to Figure 1 The above-mentioned denoising model training method includes: S101: Simulate and generate multiple pure hemodynamic response signals, and add noise to each pure hemodynamic response signal to form multiple noisy functional near-infrared spectral signals; wherein each pure hemodynamic response signal corresponds one-to-one with each noisy functional near-infrared spectral signal to form a training sample set; This application is data-driven and uses mathematical modeling and linear superposition to automatically generate an infinite number of pairs of pure HRF signals (label Y) and noisy fNIRS signals (input X) training samples. No labeling or real paired data is required. It can generate an infinite number of highly realistic and diverse training sets and can quickly build large-scale training sample sets.
[0012] In one possible implementation, S101 may include: S1011: For any pure hemodynamic response signal, randomly generate motion artifacts within a physiologically reasonable range; wherein, motion artifacts include: spike artifacts and baseline drift artifacts; extract the resting-state functional near-infrared spectral signal from the true functional near-infrared spectral signal, and generate physiological noise based on the resting-state functional near-infrared spectral signal; superimpose the motion artifacts, physiological noise, and the pure hemodynamic response signal to obtain the noisy functional near-infrared spectral signal corresponding to the pure hemodynamic response signal.
[0013] In the core application scenarios of fNIRS (neuroscience, rehabilitation medicine, monitoring of children's brain development, human-computer interaction, etc.), subjects (especially patients, infants, and subjects in motion) cannot keep their heads absolutely still for a long time. Head shaking, relative displacement of the probe and scalp, and body movement will inevitably produce motion artifacts. These are interferences that cannot be completely avoided by physical means in the actual data collection process, and are also the primary cause of data failure in clinical and scientific research scenarios.
[0014] The signal amplitude of motion artifacts is typically tens to hundreds of times greater than that of the hemodynamic response (HRF) signal induced by real brain activity. Specifically, spike artifacts cause abrupt signal jumps, directly truncating the temporal characteristics of the effective signal; baseline drift artifacts cause overall low-frequency shifts in the signal, completely altering the baseline level and relative amplitude of the HRF signal. The superposition of these two artifacts completely drowns out the weak brain activity signals of interest, directly leading to complete distortion of subsequent brain region activation analysis and statistical tests. Therefore, this application considers the impact of motion artifacts by incorporating motion artifacts into the pure hemodynamic response signal, taking into account both spike and baseline drift artifacts, to make the training samples closer to real samples.
[0015] Physiological noise originates from the body's own non-brain-derived physiological activities, including heartbeat, respiration, vasoconstriction and vasodilation, Mayer waves, etc. It is a background interference that inevitably exists in all real fNIRS signals and cannot be eliminated by physical means.
[0016] The HRF signal of real brain activity is concentrated in the low frequency band of 0.01-0.1Hz, while the core frequency band of physiological noise (0.2-0.3Hz for respiration, about 0.1Hz for Mayer waves, and heartbeat harmonics) highly overlaps with the frequency band of HRF signal. Traditional bandpass filtering cannot achieve accurate separation between the two, which can easily lead to false positive / false negative results in subsequent brain activation analysis. This is the core hidden interference affecting the accuracy of data statistics.
[0017] Based on the above analysis, motion artifacts (spike artifacts and baseline drift artifacts) and physiological noise are the two most unavoidable, most destructive to the effective signal, and directly determine the model's generalization ability in real fNIRS signals. Therefore, referring to... Figure 2 In this application, HRF signals are superimposed with motion artifacts and physiological noise, enabling noisy functional near-infrared spectroscopy signals to simulate the statistical characteristics and temporal features of real fNIRS signals, balancing thorough denoising with high fidelity of the effective signal. Simultaneously, linear superposition does not destroy the morphology and amplitude of the HRF signal, ensuring complete label accuracy and allowing the model to precisely learn the characteristics of real brain activity signals.
[0018] In one possible implementation, S1011 may include: 1. An autoregressive model was used to fit the resting-state functional near-infrared spectral signal to obtain noise parameters; the noise parameters include: autoregressive model coefficients and residual standard deviation. The resting-state fNIRS signal is a pure HRF brain activity signal without task-induced noise. Its core components are the human body's inherent non-brain-derived physiological noise (heartbeat, respiration, Mayer waves, vasomotor activity, etc.), which are natural physiological noise sample sources without additional interference, providing a real and unbiased learning object for model fitting.
[0019] This application uses real, unlabeled, resting-state fNIRS signals as natural physiological noise samples. It uses an autoregressive (AR) model to mathematically and reproducibly fit the intrinsic temporal correlation and statistical distribution characteristics of physiological noise, thus obtaining a "noise recipe" (autoregressive model coefficients and residual standard deviations) that is entirely derived from real data.
[0020] The autoregressive (AR) model coefficients can accurately quantify the temporal correlation, power spectrum distribution, and frequency band characteristics of physiological noise, characterizing the intrinsic laws governing the changes of physiological noise over time. The residual standard deviation is used to quantify the random fluctuation amplitude and overall energy distribution of physiological noise, reflecting the intensity characteristics of real physiological noise. Together, these two sets of parameters constitute a reproducible and transferable physiological noise "noise recipe," which is entirely derived from real collected data and free from any subjective bias.
[0021] For example, input real, unlabeled fNIRS experimental data (e.g., .snirf file), load the file using the mne-nirs library, and convert it into optical density (OD) signal; apply short-channel regression algorithm to initially remove systematic noise; based on the event annotations in the file, automatically identify and extract all resting-state signal segments located between two task stimuli with a duration greater than a preset threshold (e.g., 10 seconds), and splice them together to obtain the resting-state fNIRS signal; use the statsmodels library to fit a fifth-order autoregressive (AR(5)) model, and save the fitted autoregressive model coefficients and residual standard deviations as the "noise recipe" corresponding to the file.
[0022] 2. Generate physiological noise based on noise parameters.
[0023] Based on the "noise recipe" obtained above, a simulation signal that is highly homologous to real human physiological noise is generated, which completely solves the problems of mismatch between artificially simulated physiological noise and real scene noise characteristics and poor generalization in the existing technology, and provides a high-fidelity supervised data foundation for model training.
[0024] Specifically, generating physiological noise based on noise parameters can include: (1) Generate a Gaussian white noise sequence that matches the standard deviation of the fitted residuals; (2) Based on the autoregressive model coefficients, physiological noise with time-series correlation, statistical distribution and power spectrum characteristics that are completely consistent with the resting state functional near-infrared spectral signal is generated by linear weighted recursion of historical values.
[0025] The generated noise is neither random white noise nor artificially set simple colored noise, but completely replicates the noise characteristics brought about by real human physiological activities. It is highly homologous to the physiological noise in real fNIRS acquisition scenarios, ensuring the consistency between the synthesized noisy fNIRS signal and the real application scenario from the source.
[0026] In one possible implementation, S101 may further include: S1012: Generate multiple inactive samples; The non-activation sample corresponds to the resting state in the fNIRS experiment where there is no task stimulus and no brain activation. At this time, there is no HRF response, which is a real "no signal" situation. It is used to let the model learn "what is a state without effective signal".
[0027] This application introduces inactive samples, allowing the model to learn true signals, noise, and no-signal states, which can effectively avoid misjudging noise as brain activation and reduce the false positive rate.
[0028] S1013: Based on the dual gamma function, generate multiple positive samples whose amplitude, shape, and delay all vary randomly within a preset range; The dual gamma function is the recognized standard HRF model in the field of fNIRS, which can accurately simulate the rise, peak, fall, and slight overshoot of hemoglobin concentration changes after brain stimulation.
[0029] This application is based on a dual gamma function, which randomly perturbs the amplitude, morphology, and delay within a physiologically reasonable range, and can generate positive samples with rich morphology that closely resemble real brain activity.
[0030] S1014: Multiple inactive samples and multiple positive samples form multiple pure hemodynamic response signals.
[0031] By merging the inactive samples corresponding to brainless activation with the positive samples corresponding to brain activation, we can obtain all types of data covering the real fNIRS data. This not only ensures the physiological authenticity of the signal but also improves the diversity of training samples and the robustness of the model. It provides a reliable, standardized, and scalable data foundation for subsequent synthesis of noisy fNIRS signals and the realization of high-precision fNIRS signal denoising.
[0032] This application is based on a data-driven simulation method that can generate an infinite, diverse and highly realistic training sample set without any manual labeling.
[0033] For example, the proportion of non-activated samples in multiple pure hemodynamic response signals can be 15%.
[0034] Based on the above training samples, this application can adopt a two-stage training method. The first stage establishes basic restoration capabilities, and the second stage fine-tunes and optimizes the output quality. This retains the real signal amplitude, peak value, and timing position learned in the first stage, while removing high-frequency jitter and residual artifacts, resulting in a cleaner HRF waveform that is more in line with physiological laws.
[0035] S102: Using the training sample set, the initial denoising model is trained with the first loss function to obtain the intermediate denoising model; First, the initial denoising model is trained using the first loss function, enabling the model to learn to accurately reconstruct the amplitude, shape, and time position of the HRF signal, thus establishing a reliable basic denoising capability.
[0036] In one possible implementation, the first loss function can be the mean squared error loss function.
[0037] Mean Squared Error (MSE) is smooth and easy to optimize. Combined with the ADAM optimizer, it allows the model to quickly learn the basic shape, peaks, and baseline of the HRF, avoiding divergence in the early stages of training. Simultaneously, the point-by-point constraint of MSE ensures the output matches the true value, preserving the amplitude, waveform, and temporal structure of real brain activity signals to the greatest extent possible, thus preventing signal damage at its source. Furthermore, the training samples are strictly paired samples, and MSE, as a standard supervised loss, is best suited for learning the mapping relationship from "noisy" to "clean".
[0038] It should be noted that normalization can also be used to enable MSE to stably handle the problem of large amplitude differences in fNIRS signals, thereby improving the model's adaptability to signals of different intensities.
[0039] In one possible implementation, the initial denoising model can be a denoising autoencoder with a U-Net structure.
[0040] The HRF of fNIRS is a low-frequency, slowly varying, waveform-sensitive timing signal with small amplitude, easily smoothed or lost. Net is the structure that best guarantees high-fidelity waveforms.
[0041] Specifically, the structural reference of the U-Net denoising autoencoder. Figure 3 . refer to Figure 3The input data is a noisy fNIRS signal, and the output data is a denoised fNIRS signal. The structure includes an encoder path and a decoder path, fusing deep and shallow features through skip connections. The encoder consists of multiple Conv1d layers and MaxPooling layers, while the decoder consists of Upsampling layers and Conv1d layers. All convolutional layers are followed by the LeakyReLU activation function to enhance the model's non-linear expressiveness and prevent gradient vanishing.
[0042] The encoder is used to extract deep denoising features, and the decoder is used to accurately recover the signal waveform. By using skip connections, it preserves the temporal details and morphological features of brain activity signals to the greatest extent. It can achieve high-fidelity restoration of effective signals under strong motion artifacts and physiological noise interference. It also has the advantages of stable training, fast convergence and strong generalization. It is particularly suitable for denoising tasks of biomedical temporal signals such as fNIRS, which are weak in amplitude, high in interference and sensitive to waveform.
[0043] For example, the first stage of training can be 100 epochs to preserve the model weights with the best performance.
[0044] S103: Using the training sample set, the intermediate denoising model is trained with the second loss function to obtain the target denoising model; The first loss function is different from the second loss function.
[0045] Based on the intermediate denoising model that has achieved high-fidelity restoration in the first stage, the model is trained again by changing the second loss function to be different from that in the first stage. This allows the model to further improve the smoothness of the output signal, remove small jitters, and suppress residual noise while maintaining the accuracy of signal amplitude and shape, ultimately obtaining the target denoising model.
[0046] In one possible implementation, the second loss function can be:
[0047] in, For the second loss function, For the first loss function, , , These are the signal variance penalty term, the first-order difference penalty term, and the change amplitude penalty term, respectively. , , For hyperparameter weights.
[0048] The first loss function ensures that the model output can approximate the true HRF amplitude, shape, and timing point by point. The second loss function is a composite loss function, which introduces penalty terms for signal variance, first-order difference, and amplitude variation.
[0049] (1) The signal variance penalty term constrains the overall fluctuation energy of the output signal from a global statistical perspective, suppressing excessive dispersion of the entire signal caused by residual physiological noise and baseline drift; the signal variance penalty term can be defined as the variance of the model output signal, and the calculation formula is as follows:
[0050] (2) The first-order difference penalty term constrains the jump amplitude between adjacent sampling points from the perspective of the local neighborhood, improves the point-to-point smoothness of the output signal, and removes residual small high-frequency jitter; specifically, the first-order difference penalty term can be defined as the sum of squares of the differences between adjacent points of the model output signal, and the calculation formula is as follows:
[0051] (3) The variation amplitude penalty term constrains the signal curvature from the perspective of the second-order rate of change, focusing on suppressing local abrupt changes resembling sharp peak artifacts, while preserving the smooth rise-peak-fall process of the HRF itself; the variation amplitude penalty term can be defined as the sum of squares of the second-order differences of the model output signal, and the calculation formula is:
[0052] in, For the model in the first The output value of each sampling point The average value of the output signal. This is the signal length.
[0053] The three penalty terms mentioned above apply to the zero-order (energy), first-order (slope), and second-order (curvature) characteristics of the signal, respectively. They are mathematically independent and together constitute a multi-scale constraint on the temporal smoothness of the output signal, avoiding oversmoothing or undersmoothing problems under a single constraint.
[0054] In one possible implementation, the weights of each hyperparameter all range from 10. -5 ~10 - ³.
[0055] For example, each hyperparameter has a weight of 10. 4 While ensuring that MSE loss dominates the training direction and does not damage the true HRF morphology, it can effectively suppress residual noise and spikes in the output signal, and achieve a denoising effect that combines high fidelity and high smoothness.
[0056] This application employs a data-driven approach to simulate and generate training sample sets. Without any manual labeling, it can generate an unlimited number of diverse and highly realistic training samples, ensuring the data diversity and effectiveness of the training sample set. Simultaneously, the combination of a two-stage training process effectively improves the model's denoising performance. The training process requires no human intervention, the model inference speed is fast, and it is suitable for denoising large-scale fNIRS data.
[0057] Corresponding to the above embodiments, this invention also provides a functional near-infrared spectral signal denoising method, comprising: The functional near-infrared spectral signal to be denoised is input into the target denoising model provided in the above embodiment to obtain the denoised functional near-infrared spectral signal.
[0058] Using the target denoising model trained above, the denoising effect is good while ensuring signal integrity.
[0059] The functional near-infrared spectral signal denoising method provided in this embodiment of the invention is applied to denoise the functional near-infrared spectral signal (Original Scale) to be denoised, resulting in a denoised functional near-infrared spectral signal (Model). Figure 4 It can be seen that the denoised functional near-infrared spectral signal (Model) is smooth and very close to the true clean signal (Clean HRF) used as the reference, effectively removing the large-amplitude interference in the original functional near-infrared spectral signal to be denoised.
[0060] Meanwhile, the functional near-infrared spectral signal denoising method provided in this embodiment of the invention is compared with the wavelet denoising model. By detecting the number of points exceeding a preset threshold in the first-order difference of the signal, the average number of motion artifacts remaining after processing all signals in the entire test set using different methods is counted, and the result is obtained. Figure 5 Similar to comparisons with bandpass filtering and PCA methods, this method exhibits significantly fewer residual artifacts. Figure 5 It can be seen that the original functional near-infrared spectral signal to be denoised without any correction has an average of 1186.40 artifacts. After denoising using the wavelet denoising model, the average number of artifacts is reduced to 871.11. After denoising using the functional near-infrared spectral signal denoising method (DAE Fine-tuned) provided in this embodiment of the invention, the average number of artifacts is only 305.57. Figure 5 It can be seen that the functional near-infrared spectral signal denoising method provided by the embodiments of the present invention has significant advantages over existing denoising methods in removing motion artifacts.
[0061] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.
[0062] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.
[0063] Figure 6 A schematic diagram of the denoising model training device provided in an embodiment of the present invention is shown. For ease of explanation, only the parts related to the embodiment of the present invention are shown, and are described in detail below: like Figure 6 As shown, the denoising model training device includes: The sample generation module 21 is used to simulate and generate multiple pure hemodynamic response signals, and add noise to each pure hemodynamic response signal to form multiple noisy functional near-infrared spectral signals; wherein each pure hemodynamic response signal corresponds one-to-one with each noisy functional near-infrared spectral signal to form a training sample set. The first training module 22 is used to train the initial denoising model with the first loss function using the training sample set to obtain the intermediate denoising model; The second training module 23 is used to train the intermediate denoising model with the training sample set and the second loss function to obtain the target denoising model. The first loss function is different from the second loss function.
[0064] In one possible implementation, the sample generation module 21 may include: The signal synthesis unit is used to randomly generate motion artifacts within a physiologically reasonable range for any pure hemodynamic response signal; wherein the motion artifacts include: spike artifacts and baseline drift artifacts; extract the resting-state functional near-infrared spectral signal from the real functional near-infrared spectral signal, and generate physiological noise based on the resting-state functional near-infrared spectral signal; superimpose the motion artifacts, physiological noise and the pure hemodynamic response signal to obtain the noisy functional near-infrared spectral signal corresponding to the pure hemodynamic response signal.
[0065] In one possible implementation, the signal synthesis unit may include: The noise parameter determination subunit is used to fit the resting-state functional near-infrared spectral signal using an autoregressive model to obtain noise parameters; the noise parameters include: autoregressive model coefficients and residual standard deviations; The first noise generation subunit is used to generate physiological noise based on noise parameters.
[0066] In one possible implementation, the first loss function can be the mean squared error loss function.
[0067] In one possible implementation, the second loss function can be:
[0068] in, For the second loss function, For the first loss function, , , These are the signal variance penalty term, the first-order difference penalty term, and the change amplitude penalty term, respectively. , , For hyperparameter weights.
[0069] In one possible implementation, the values of each hyperparameter weight range from 10. -5 ~10 - ³.
[0070] In one possible implementation, the initial denoising model can be a denoising autoencoder with a U-Net structure.
[0071] In one possible implementation, the sample generation module 21 may include: The first sample generation unit is used to generate multiple inactive samples; The second sample generation unit is used to generate multiple positive samples with amplitude, shape and delay that vary randomly within a preset range based on the dual gamma function. The sample synthesis unit is used to generate multiple pure hemodynamic response signals from multiple inactive samples and multiple positive samples.
[0072] This invention also provides a functional near-infrared spectral signal denoising device, comprising: The denoising module is used to input the functional near-infrared spectral signal to be denoised into the target denoising model provided in the above embodiment to obtain the denoised functional near-infrared spectral signal.
[0073] Figure 7 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. For example... Figure 7 As shown, the electronic device 3 of this embodiment includes a processor 30 and a memory 31. The memory 31 stores a computer program 32. When the processor 30 executes the computer program 32, it implements the steps in the various method embodiments described above. Alternatively, when the processor 30 executes the computer program 32, it implements the functions of each module / unit in the various device embodiments described above.
[0074] For example, computer program 32 may be divided into one or more modules / units, which are stored in memory 31 and executed by processor 30 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 32 in electronic device 3.
[0075] Electronic device 3 may include, but is not limited to, processor 30 and memory 31. Those skilled in the art will understand that... Figure 7 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device 3 may also include input / output devices, network access devices, buses, etc.
[0076] The processor 30 can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.
[0077] The memory 31 can be an internal storage unit of the electronic device 3, such as a hard disk or memory of the electronic device 3. The memory 31 can also be an external storage device of the electronic device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 3. Furthermore, the memory 31 can include both internal and external storage units of the electronic device 3. The memory 31 is used to store the computer program 32 and other programs and data required by the electronic device 3. The memory 31 can also be used to temporarily store data that has been output or will be output.
[0078] For the sake of simplicity and clarity, only the above-described functional modules / units are used as examples. In practical applications, the functions described above can be assigned to different functional modules / units as needed. These modules / units can be implemented in hardware, software, or a combination of both.
[0079] This invention also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.
[0080] This invention also provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the methods described in the above-described method embodiments.
[0081] Computer programs include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.
[0082] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. Unless otherwise specified or in conflict with logic, the terminology and / or descriptions between different embodiments are consistent and can be referenced interchangeably. Technical features in different embodiments can be combined to form new embodiments based on their inherent logical relationships.
[0083] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.
Claims
1. A method for training a denoising model, characterized in that, include: Multiple pure hemodynamic response signals are simulated and generated, and noise is added to each pure hemodynamic response signal to form multiple noisy functional near-infrared spectral signals; each pure hemodynamic response signal corresponds one-to-one with each noisy functional near-infrared spectral signal to form a training sample set. Using the training sample set, the initial denoising model is trained with the first loss function to obtain an intermediate denoising model; Using the training sample set, the intermediate denoising model is trained with the second loss function to obtain the target denoising model; The first loss function is different from the second loss function.
2. The denoising model training method according to claim 1, characterized in that, The process involves adding noise to each pure hemodynamic response signal to form multiple noisy functional near-infrared spectral signals, including: For any pure hemodynamic response signal, motion artifacts are randomly generated within a physiologically reasonable range; wherein, the motion artifacts include: spike artifacts and baseline drift artifacts; a resting-state functional near-infrared spectral signal is extracted from the true functional near-infrared spectral signal, and physiological noise is generated based on the resting-state functional near-infrared spectral signal; the motion artifacts, the physiological noise, and the pure hemodynamic response signal are superimposed to obtain the noisy functional near-infrared spectral signal corresponding to the pure hemodynamic response signal.
3. The denoising model training method according to claim 2, characterized in that, The generation of physiological noise based on the resting-state functional near-infrared spectral signal includes: An autoregressive model was used to fit the resting-state functional near-infrared spectral signal to obtain noise parameters; wherein, the noise parameters include: autoregressive model coefficients and residual standard deviation; The physiological noise is generated based on the noise parameters.
4. The denoising model training method according to any one of claims 1 to 3, characterized in that, The first loss function is the mean squared error loss function.
5. The denoising model training method according to claim 4, characterized in that, The second loss function is: in, For the second loss function, Let the first loss function be... , , These are the signal variance penalty term, the first-order difference penalty term, and the change amplitude penalty term, respectively. , , For hyperparameter weights.
6. The denoising model training method according to claim 5, characterized in that, The value range of each hyperparameter weight is 10. -5 ~10 - ³.
7. The denoising model training method according to any one of claims 1 to 3, characterized in that, The initial denoising model is a denoising autoencoder with a U-Net structure.
8. The denoising model training method according to any one of claims 1 to 3, characterized in that, The simulation generates multiple pure hemodynamic response signals, including: Generate multiple inactive samples; Based on the dual gamma function, multiple positive samples are generated, with amplitude, shape, and delay all varying randomly within a preset range. The plurality of inactive samples and the plurality of positive samples form the plurality of pure hemodynamic response signals.
9. A method for denoising functional near-infrared spectral signals, characterized in that, include: The functional near-infrared spectral signal to be denoised is input into the target denoising model as described in any one of claims 1 to 8 to obtain the denoised functional near-infrared spectral signal.
10. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the denoising model training method as described in any one of claims 1 to 8 and the functional near-infrared spectral signal denoising method as described in claim 9.