Universal self-supervised learning Raman spectrum noise reduction method, device and equipment

Abstract

The invention discloses a universal self-supervised learning Raman spectrum noise reduction method, which comprises the following steps of: 1, acquiring Raman data; step 2, preprocessing the Raman data, removing a baseline from the Raman data by using an adaptive iterative reweighted penalty least square method, and standardizing a spectrum after baseline removal to obtain preprocessed spectrum data; step 3, processing the preprocessed spectral data by using a down-sampling strategy H (.) to obtain two sub-spectrum sets h1 (.) and h2 (.) with independent noise, each sub-spectrum set comprising 10 sub-spectrums; and step 4, inputting the preprocessed spectrum data and the corresponding two sub-spectrum sets into a U-Net network f theta (.) to obtain a spectrum after noise reduction processing. The noise reduction method can carry out good noise reduction processing on various Raman data, and the self-supervised learning method does not need a complex long-time learning process.

Description

Technical Field

[0001] The present invention relates to the field of spectral noise reduction, and in particular to a universal self-supervised learning Raman spectral noise reduction method, device and equipment. Background Art

[0002] Raman spectroscopy (RS) is an optical mode for molecular fingerprinting in a non-destructive and label-free manner. However, only one in every 108 photons may undergo Raman scattering, i.e., the Raman scattering signal is inherently weak, which makes it sensitive to sensor defects, so it is necessary to enhance the signal and eliminate instrument and background noise. Plasmon-based nanomaterials can further enhance Raman scattering, i.e., surface-enhanced Raman spectroscopy (SERS), which makes it possible to perform real-time imaging and specific detection of single molecules during surgery. Thanks to its high specificity and sensitivity, Raman technology has been widely used in food testing, drug testing, medical testing and other fields.

[0003] In the past few decades, advances in optical instruments such as lasers, filters, spectrometers, and detection equipment have greatly improved spectral resolution and signal-to-noise ratio (SNR), promoting the development of Raman spectroscopy. However, when faced with detection scenarios with weak Raman signals, such as single-molecule detection and deep penetration detection, the noise caused by the sample itself and the instrument background often buries the effective information in the spectrum, while the experimental methods or improvements to optical devices are harsh and have little effect. Therefore, efficient, economical, and universal spectral noise reduction methods, especially the development of advanced noise reduction algorithms, are particularly important.

[0004] At present, the commonly used spectral denoising algorithms can be divided into the moving window method based on polynomial fitting and the wavelet transform method based on power spectrum. The moving window method dynamically fits the continuous data points in the spectrum according to the window university, such as the moving average method and the Savitzky-Golay (SG) filter. In contrast, the wavelet transform method uses a specific wavelet basis and introduces a threshold coefficient to reconstruct the spectral signal while retaining the key features in the spectrum, such as wavelet threshold denoising (WTD). Although these traditional algorithms are very fast for Raman spectroscopy denoising, parameters such as window size and polynomial order still need to be determined empirically during use. In addition, with the explosive growth of spectral data, it is very time-consuming to fine-tune the parameters of the data one by one, and real-time processing cannot be achieved.

[0005] In contrast, deep learning (DL) does not require manual parameter adjustment and can automatically learn and extract high-level features from big data and incorporate them into accurate predictions and decisions. In terms of spectral denoising, a large number of research works use simulated training data with added simulated noise or background to achieve denoising effects. However, these models require a large number of ideal “clean-noise” spectral pairs for training. This simulated data cannot cover all possible noise types in the real world, and the difference between simulated data and experimental data also brings uncertainty to the generalization ability of DL.

[0006] For Raman spectroscopy, real data (such as those in clinical applications) are limited, and the corresponding labels (i.e., completely noise-free "clean" spectra) are generally not available. A general noise reduction strategy that can quickly and stably reduce the noise of Raman and SERS spectra using a small amount of real data and does not rely on empirical parameter selection has not yet been developed and applied.

[0007] The current Raman spectroscopy noise reduction technology has the following defects:

[0008] 1. Traditional noise reduction algorithms rely on user experience to select parameters, which is almost ineffective for ultra-weak signal spectra.

[0009] 2. Traditional deep learning methods rely on simulated data for supervised training, which takes a long time to train and cannot fully learn the real noise and signal characteristics, and cannot perform noise reduction on a single spectrum.

[0010] 3. The self-supervised loss function is designed based on image data, lacks adaptive adjustment to spectral data, and lacks information retention items.

[0011] 4. The training process of deep learning methods is time-consuming and cannot achieve real-time noise reduction during the acquisition process.

[0012] Therefore, technicians in this field are committed to developing a universal self-supervised learning Raman spectroscopy denoising method. Summary of the invention

[0013] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is how to perform fast and effective noise reduction processing on a variety of Raman data.

[0014] To achieve the above objectives,

[0015] In a first aspect, the present invention provides a general self-supervised learning Raman spectroscopy denoising method, comprising:

[0016] Step 1: Acquire Raman data.

[0017] Step 2: Raman data preprocessing,

[0018] The baseline of Raman data was removed using the adaptive iterative reweighted penalized least squares method, and the spectrum after baseline removal was normalized to obtain the preprocessed spectrum data.

[0019] Step 3: Use the downsampling strategy H(·) to process the preprocessed spectral data to obtain two noise-independent sub-spectral sets h1(·) and h2(·), each of which includes 10 sub-spectra.

[0020] Step 4: Input the preprocessed spectral data and the corresponding two sub-spectral sets into the U-Net network θ (·) to obtain the spectrum after noise reduction.

[0021] Furthermore, step 2 is specifically set as follows: the downsampling strategy H(·) slides on the spectrum with a window size of 3. For the three Raman shifts covered by the window, H(·) randomly selects two different shifts and copies their corresponding intensity values ​​to create a sub-spectrum with the same dimension as the original spectrum. The above operation is repeated twice to finally obtain the sub-spectrum sets h1(·) and h2(·).

[0022] Furthermore, the U-Net network f θ (·) contains 5 modules,

[0023] The first module includes 2 one-dimensional convolutional layers, 2 ReLU activation layers, 1 maximum pooling layer and 1 copy and splicing layer. The one-dimensional convolutional layer, ReLU activation layer and maximum pooling layer are used to extract features. The extracted features are passed to the second module. At the same time, the extracted features are passed to the fifth module by retaining low-level features through the copy and splicing layer.

[0024] The second module includes 2 one-dimensional convolution layers, 2 ReLU activation layers, 1 maximum pooling layer and 1 copy and splicing layer. The one-dimensional convolution layer, ReLU activation layer and maximum pooling layer are used to extract features. The extracted features are passed to the third module. At the same time, the extracted features are passed to the fourth module by retaining the low-level features through the copy and splicing layer.

[0025] The third module includes 2 one-dimensional convolution layers, 2 ReLU activation layers and 1 upsampling layer. The one-dimensional convolution layer and the ReLU activation layer extract features. The extracted features are restored through the upsampling layer and passed to the fourth module.

[0026] The fourth module includes 2 one-dimensional convolution layers, 2 ReLU activation layers and 1 upsampling layer. The convolution layer and the ReLU activation layer concatenate the features transmitted by the second module with the features transmitted by the third module, and then restore the concatenated features through the upsampling layer and pass them to the fifth module.

[0027] The fifth module includes three one-dimensional convolutional layers and three ReLU activation layers. It concatenates the features delivered by the first module with the features delivered by the fourth module to reconstruct the spectrum.

[0028] Furthermore, the U-Net network f θ The training method of (·) is:

[0029] Use stochastic gradient descent SGD to train the U-Net network f θ (·) for training, the loss function L is:

[0030] L=λ1L reg +λ2(L rec1 +L rec2 )

[0031] Where:

[0032] L reg =||yf θ (y)||2

[0033] L rec1 =||h1(y)-f θ (h2(y))||2

[0034] L rec2 =||h2(y)-f θ (h1(y))||2

[0035] Where: λ = {λ1, λ2} is the weight of the corresponding loss term, represents the input Raman spectrum as an n-dimensional real vector, and h1(·) and h2(·) represent the two sub-spectrum sets obtained by downsampling in step 2.

[0036] Furthermore, the learning rate of the stochastic gradient descent method SGD is 0.003.

[0037] Furthermore, the U-Net network f θ The training process of (·) is judged to converge based on the difference of the loss function of two iterations. When the loss difference is less than 10 -4 The model stops training when .

[0038] Technical Effects

[0039] 1. By building a deep learning denoising model, the U-net neural network is used to automatically learn the noise distribution and signal characteristics in real spectral data. After the denoising model is established, only the noisy spectrum is required as input. The deep learning model is composed of multiple layers of neurons, which can effectively capture the complex structure and abstract features in the data when facing complex real spectra. There is no need to select parameters during use, and the input spectrum is directly denoised, which avoids dependence on user experience and reduces the subjective bias of manual parameter adjustment; the denoising model is adaptive to different data and has a stable denoising effect on spectral data with different signal strengths.

[0040] 2. Design a denoising task objective function based on self-supervised learning, use the characteristics of the data itself for effective supervision, and achieve single spectrum denoising; introduce regularization terms in the objective function to emphasize the full retention of effective information in one-dimensional spectral data. Shot noise is the main type of noise that affects the spectrum. This type of noise has a consistent effect on the entire spectrum and has a mean of zero. The self-supervised denoising model uses a downsampling strategy to construct a noise-independent sub-spectrum set. Different sub-spectrum sets are mutually constrained and integrated using mean square error loss. The self-supervised strategy only uses the input spectrum itself to construct supervision, and the downsampling strategy also takes into account the data amplification effect, which can achieve single spectrum denoising. No training process is required during use; the downsampling strategy constructs a sub-spectrum set with prominent independent and identically distributed noise, highlights potential bias interference, and can obtain a more reliable and stable denoising effect; directly uses the real spectrum to supervise itself and fully learns the real noise distribution; achieves single spectrum denoising with stable effect.

[0041] 3. In addition to the commonly used regularization term, the method adds the MSE term between the input and the output to emphasize the importance of information. The regularization term uses the mean square error to construct constraints between the original spectrum and the denoised spectrum, retaining the effective information in the denoised spectrum. The added information regularization term achieves comprehensive information retention. At the same time, this term improves the SNR of the denoised spectrum and reduces the ED between the result and the reference spectrum, making this method more suitable for the characteristics of spectral data.

[0042] 4. The training process uses an adaptive convergence judgment method. The training process is based on the difference in the loss function of two iterations for convergence judgment. When the loss difference is <10-4, the model stops training. This method performs adaptive training and denoising on spectra of different qualities to achieve real-time denoising.

[0043] 5. The present invention uses a self-supervised learning strategy to apply deep learning to the denoising task of one-dimensional spectral data, avoiding the parameter selection and experience dependence of traditional denoising algorithms and the requirement of supervised learning for a large amount of training data, and the use process is simple and fast. The denoising effect of this method has been evaluated by multiple indicators, and it has obvious effects at multiple characteristic peaks. The SNR improvement effect is significantly better than the current traditional algorithm. The denoising result is close to the average spectrum and has good stability; at the same time, the peak position and half-peak width are unchanged, proving that this method effectively reduces noise on the basis of ensuring important information in the spectrum. The operation process of this method does not require manual parameter selection, only the spectrum to be denoised needs to be input, which is convenient for combining with existing Raman spectroscopy analysis software for spectral processing. The efficiency of this method in processing a single spectrum is about 1.84 frames / second, which matches the commonly used test time of Raman technology, and can be combined with Raman spectroscopy acquisition software to realize real-time spectral denoising. The model of this method is lightweight, about 1.76M, and the amount of calculation is much smaller than other deep learning methods (e.g., CNN, 128.43M), which is convenient for combining with various software.

[0044] In a second aspect, the present invention provides a general self-supervised learning Raman spectroscopy denoising device, characterized in that it comprises:

[0045] A data acquisition module, used for acquiring Raman data;

[0046] A first computing module, used for Raman data preprocessing;

[0047] The second computing module is used to input the preprocessed spectral data and the corresponding two sub-spectral sets into the U-Net network. θ (·), obtain the spectrum after noise reduction;

[0048] A noise reduction processing module is used to input the pre-processed spectral data and the corresponding two sub-spectral sets into the U-Net network f θ (·) to obtain the spectrum after noise reduction.

[0049] In a third aspect, the present invention provides a computer device, characterized in that it includes:

[0050] Memory;

[0051] One or more processors coupled to the memory;

[0052] One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by one or more processors, and the one or more applications are configured to execute the general self-supervised learning Raman spectroscopy denoising method provided in the first aspect above.

[0053] In a fourth aspect, the present invention provides a computer-readable storage medium having a program code stored therein, wherein the program code can be called by a processor to execute the general self-supervised learning Raman spectroscopy denoising method provided in the first aspect.

[0054] In a fifth aspect, the present invention provides a computer program, including a computer program, which, when executed by a processor, implements the general self-supervised learning Raman spectroscopy denoising method provided in the first aspect.

[0055] It can be understood that the beneficial effects of the second, third, fourth and fifth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here.

[0056] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flow chart of a self-supervised learning Raman spectroscopy denoising method according to a preferred embodiment of the present invention;

[0058] Figure 2 It is a structural diagram of a self-supervised U-Net denoising model of a preferred embodiment of the present invention;

[0059] Figure 3 It is a schematic diagram of a downsampling strategy of a preferred embodiment of the present invention;

[0060] Figure 4 Schematic diagram of a self-supervised denoising objective function of a preferred embodiment of the present invention;

[0061] Figure 5 It is a noise reduction result diagram of simulation data, Raman spectrum of ethanol, and open source HeLa cell Raman spectrum of a preferred embodiment of the present invention;

[0062] Figure 6 is a diagram showing changes in spectral shape due to different downsampling strategies in a preferred embodiment of the present invention;

[0063] Figure 7 It is a result diagram of characteristic peak intensity and Euclidean distance of various Raman data in a preferred embodiment of the present invention;

[0064] Figure 8 This is a noise reduction result diagram of deep penetration data according to a preferred embodiment of the present invention;

[0065] Fig. 9 It is a graph showing the denoising results and Euclidean distance results of different characteristic peaks of deep penetration data using different denoising methods in a preferred embodiment of the present invention;

[0066] Fig.10 This is a diagram of the occlusion method analysis results of a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0067] The following describes several preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0068] In the drawings, components with the same structure are denoted by the same numerical numerals, and components with similar structure or function are denoted by similar numerical numerals.

[0069] Figure 1 A general Raman spectroscopy noise reduction method is shown, including:

[0070] Acquire various Raman data, preprocess Raman data, construct sub-spectrum sets, and perform denoising using the self-supervised U-Net model.

[0071] The obtained Raman data can include: simulation data, open source Hela cell Raman spectrum (Nat. Commun. 2024, 15 (1), 754), SERS spectrum of ethanol molecules and SERS spectrum of GERRTs particles under the interference of muscle tissue of different thicknesses. The data types cover Raman spectrum and SERS spectrum and different detection scenarios, fully demonstrating the versatility of the method of the present invention.

[0072] Raman data preprocessing: The collected Raman data (except simulation data) usually contains the fluorescence background of the sample. The baseline of the collected spectrum is removed using the adaptive iterative reweighted penalized least squares method (Automatic Iterative Removal of Pump-Laser Line from Raman Spectra, airPLS), and the spectrum after baseline removal is normalized to obtain the preprocessed spectrum data.

[0073] Constructing sub-spectrum sets: Since shot noise has a uniform effect on the entire spectrum, the present invention uses a downsampling strategy to process the preprocessed spectrum to obtain two noise-independent sub-spectrum sets, each of which includes 10 sub-spectra.

[0074] Self-supervised U-Net model denoising: The pre-processed spectrum and the corresponding sub-spectrum set are constrained in the network to obtain the denoised spectrum. The U-Net network uses a one-dimensional convolution layer and a pooling layer to construct a network module, and uses a step size of 3 to avoid the loss of spectral information, which is suitable for spectral processing with high information density. The self-supervised denoising model uses a downsampling strategy to construct a noise-independent sub-spectrum set. Different sub-spectrum sets are mutually constrained and integrated using mean square error loss. We added a special regularization term and used mean square error to constrain the original spectrum and the denoised spectrum, thereby retaining the effective information in the denoised spectrum. The self-supervised strategy only uses the input spectrum itself to construct supervision, and the downsampling strategy takes into account the data amplification effect at the same time, which can achieve single spectrum denoising. The method performs adaptive training and denoising on spectra of different qualities. The training process is judged to converge based on the difference in the loss function of two iterations. When the loss difference is <10-4, the model stops training to achieve real-time denoising.

[0075] Multi-index evaluation: The method is compared with the commonly used SG filter, Wavelet Threshold Denoising (WTD) and self-supervised CNN methods. The SNR, half-peak width, peak shape and ED between the corresponding average spectrum of multiple characteristic peaks are used for qualitative and quantitative evaluation. The occlusion method is used to analyze the interpretability of the model denoising process.

[0076] Network architecture θ (·) Based on U-Net framework ( Figure 2 ), a single Raman spectrum is denoised by the network weight parameter θ. In use, the input high-dimensional spectrum is first encoded into a low-dimensional tensor and then reconstructed into a denoised spectrum. To achieve fast processing, a lightweight network containing only 5 modules is used, the first and second modules constitute the encoder, and the fourth and fifth modules constitute the decoder.

[0077] The first module includes 2 one-dimensional convolutional layers, 2 ReLU activation layers, 1 maximum pooling layer and 1 copy and splicing layer. The one-dimensional convolutional layer, ReLU activation layer and maximum pooling layer are used to extract features. The extracted features are passed to the second module. At the same time, the extracted features are passed to the fifth module by retaining low-level features through the copy and splicing layer.

[0078] The second module includes 2 one-dimensional convolution layers, 2 ReLU activation layers, 1 maximum pooling layer and 1 copy and splicing layer. The one-dimensional convolution layer, ReLU activation layer and maximum pooling layer are used to extract features. The extracted features are passed to the third module. At the same time, the extracted features are passed to the fourth module by retaining the low-level features through the copy and splicing layer.

[0079] The third module includes 2 one-dimensional convolution layers, 2 ReLU activation layers and 1 upsampling layer. The one-dimensional convolution layer and the ReLU activation layer extract features. The extracted features are restored through the upsampling layer and passed to the fourth module.

[0080] The fourth module includes 2 one-dimensional convolution layers, 2 ReLU activation layers and 1 upsampling layer. The convolution layer and the ReLU activation layer concatenate the features transmitted by the second module with the features transmitted by the third module, and then restore the concatenated features through the upsampling layer and pass them to the fifth module.

[0081] The fifth module includes three one-dimensional convolutional layers and three ReLU activation layers. It concatenates the features delivered by the first module with the features delivered by the fourth module to reconstruct the spectrum.

[0082] The decoder is symmetrical to the encoder and reuses the features from the encoder to focus on the locations of feature peaks in the spectrum.

[0083] The downsampling strategy H(·) is similar to the maximum pooling function ( Figure 3 ). H(·) slides on the spectrum with a window size of k. For the k Raman shifts covered by the window, H(·) randomly selects two different shifts and copies their corresponding intensity values ​​to create a sub-spectrum with the same dimension as the original spectrum. Repeat the above operation twice, and finally obtain the sub-spectrum sets h1(·) and h2(·). Multiple sub-spectrum sets can be constrained to obtain the characteristics of different Raman shifts. In addition to suppressing bias interference, this downsampling strategy can also effectively remove zero-mean shot noise. Assuming that the noise at each Raman shift is independent, that is, there is no correlation between the noise at different positions, the noise of the two sub-spectrum sets obtained by downsampling is independent and identically distributed. Using the mean square error loss ensemble for training, this downsampling strategy can converge to a noise-free signal faster.

[0084] This example develops a general one-dimensional spectral data denoising loss function ( Figure 4 ). Denoting the input Raman spectrum as an n-dimensional real vector, we first define the regression loss L between the original spectrum and the denoised spectrum reg , use the mean square error (MSE, ||·||2) to ensure important peaks:

[0085] L reg =||yf θ (y)||2

[0086] The input Raman spectrum is downsampled to compress the noise, therefore, we use L rec1 With L rec2 The sub-spectrum is denoised by constructing constraints through MSE, and the calculation formula is:

[0087] L rec1 =||h1(y)-f θ (h2(y))||2

[0088] L rec2 =||h2(y)-f θ (h1(y))||2

[0089] Among them, h1(·) and h2(·) represent the two sub-spectral sets obtained by downsampling.

[0090] The objective function is defined as a combination of the above three loss terms to achieve significant denoising of the spectrum while avoiding distortion:

[0091] L=λ1×L reg +λ2×(L rec1 +L rec2 )

[0092] Where λ = {λ1, λ2} is the weight of the corresponding loss term, which is used to balance the noise reduction effect and maintain the spectral shape.

[0093] We use SGD as the gradient optimization method for the self-supervised U-Net denoising network with a learning rate of 0.003. In order to adapt to spectra of different qualities, model training stops when the loss function changes by less than 10-4.

[0094] A variety of evaluation indicators are used to perform qualitative and quantitative analysis on the noise reduction effect of the noise reduction method provided in this embodiment.

[0095] Signal-to-noise ratio (SNR): For the characteristic peak in the reference spectrum, select the 5 wave numbers near the peak as the range for calculating the average peak intensity. The average peak intensity calculation formula is as follows:

[0096] I signal =∑ signal range I i / N

[0097] Among them, I i is the signal intensity corresponding to the ith Raman shift in the signal range, and N=5 is the number of Raman shifts contained in the signal range.

[0098] The frequency band without characteristic peaks in the spectrum is selected as the range of the noise standard deviation. The calculation formula for the noise standard deviation is:

[0099]

[0100] Among them I j represents the intensity corresponding to the jth Raman shift within the noise range, represents the average intensity within the noise range, and M=50 represents the number of Raman shifts contained in the noise range.

[0101] The signal-to-noise ratio calculation formula of the characteristic peak is as follows:

[0102]

[0103] If the obtained characteristic peak signal-to-noise ratio is greater than 3, the confidence level is 99%, indicating that the characteristic peak at that location is different from random noise.

[0104] Euclidean distance (ED): Define a high-quality spectrum (i.e., a spectrum of a high-concentration substance or an average spectrum) as the reference spectrum, and calculate the ED between the denoised spectrum and the reference spectrum. The formula is as follows:

[0105]

[0106] Among them, x i，Reference represents the intensity corresponding to the ith Raman shift from the reference spectrum, X i，Denoised Represents the intensity corresponding to the i-th Raman shift in the de-noised spectrum.

[0107] Characteristic peak position: Find the Raman shift corresponding to the strongest signal in the characteristic peak range of the spectrum as the peak position, and compare the changes in the characteristic peak position of the light after noise reduction with that of the reference spectrum.

[0108] Full Width Half Maximum (FWHM): Find the Raman shift of half the strongest signal of the corresponding characteristic peak in the spectrum, and calculate the change in FWHM between the spectrum after noise reduction and the reference spectrum.

[0109] This embodiment proposes a self-supervised U-Net denoising network that reduces the noise in Raman spectra by downsampling without the need for a large amount of training data and manual parameter setting. A single Raman spectrum is used as input and two sub-spectrum sets are obtained after subsampling. Thanks to self-supervised learning and downsampled data, our U-Net denoising network can quickly and effectively remove noise from spectra.

[0110] The performance of the proposed general denoising algorithm is verified by simulation data. The Lorentz function is used to generate Raman characteristic peaks, and the main peaks are located at 600, 800, 1200, and 1270 cm-1. Subsequently, Gaussian white noise with a noise level of 60 is added to the simulated spectrum. Figure 5 It can be seen that the small peak at 800cm-1 is masked and the adjacent peaks (1200 and 1270cm-1) are difficult to distinguish. The spectrum after denoising by the algorithm can consistently identify four peaks and can well separate the two peaks at 1200 and 1270cm-1. Figure 5 a (left), shaded in grey) is used to calculate δ noiseThe results show that the signal-to-noise ratio of each peak is significantly enhanced. Typically, the signal-to-noise ratio is 3 ( Figure 5 a (right), black dashed line) is considered as the detection threshold of the characteristic Raman peak.

[0111] Raman spectrum of ethanol molecule ( Figure 5 b (left)) The three main characteristic peaks at 879, 1050 and 1087 cm-1 are CCO symmetric stretching vibration, antisymmetric stretching vibration and COH vibration (light gray). The range of 78-245 cm-1 is used ( Figure 5 b (left), grey shading) to calculate δ noise Our method shows a consistent increase in the signal-to-noise ratio at three main peaks, such as Figure 5 b (right). Therefore, both self-supervised learning and downsampling strategies enable the U-Net denoising network to operate effectively on a single spectrum.

[0112] The open-source spectrum of HeLa cells was also used to further consolidate the generalization ability and robustness of the proposed method. The Raman shift at 2800-3000cm-1 represents the CH signal of organic matter in cells. The spectral noise after denoising was significantly compressed, the characteristic peak at 2932cm-1 was more clearly identified, and the signal-to-noise ratio was improved from 3.31 to 18.19±1.94. In addition, the characteristic peak intensities and EDs of data from different sources showed that the denoising results of the proposed method were closer to the corresponding reference spectra ( Figure 6 ).

[0113] Table 1 (objective function ablation experiment) shows that the three loss terms of the objective function all contribute to the significant denoising performance of our algorithm.

[0114] Table 1

[0115]

[0116] Furthermore, our downsampling strategy has negligible effects on the position, intensity, and width of each Raman peak (Table 2 (peak positions for different downsampling strategies), Table 3 (half-peak width for different downsampling strategies), and Figure 7 ).

[0117] Table 2

[0118]

[0119]

[0120] Table 3

[0121]

[0122] The effects of different noise conditions are shown in Table 4 (SNR and ED after adding spectral denoising with different noise levels), which shows that the proposed denoising method has strong generalization ability.

[0123] Table 4

[0124]

[0125] We evaluate the effectiveness of the proposed algorithm in improving deep lesion detection, i.e., detecting SERS tags buried in thick biological tissues ( Figure 8 a). We used gap-enhanced resonance Raman probes (GERRTs) with IR780 Raman reporter molecules as the signal source and embedded them into ex vivo porcine tissue. To collect Raman spectra with different signal-to-noise ratios, porcine muscle tissues of different thicknesses were used to simulate conditions under different noise levels. The overall tissue thickness was adjusted by adding tissue slices ( Figure 8 b). A transmission Raman device was used to irradiate and collect signals on both sides of the tissue. The laser powers used were 100, 200, 400, and 600 mW, and the thickness of the pig muscle tissue was 4, 6, 8, 10, 12, 14, 16, and 18 cm, respectively. GERRTs were placed on the excitation surface, and SERS spectra were collected by changing the laser power and increasing the tissue thickness on the GERRTs. A total of 160 SERS spectra were obtained, with a signal-to-noise ratio range of 0-190 ( Figure 8 d).

[0126] GERRTs have characteristic Raman peaks of IR780 at 520, 931, 1203, and 1580 cm-1, which can be clearly distinguished from the background ( Figure 8 c, light grey). 2000-2060cm-1( Figure 8 c, dark grey shading) is used to calculate δ noise Because light penetration in biological tissues undergoes exponential attenuation, even the strongest characteristic peak will be greatly weakened as the tissue thickness increases. When the tissue thickness increases from 10 cm to 12 cm, we observe that the SNR of the characteristic peak at 1203 cm-1 decays from 10.1 ± 2.3 to 2.7 ± 1.2 ( Figure 8 e), making the signal undetectable.

[0127] In order to test the robustness of the proposed method for noise reduction of SERS spectra of different qualities, all spectra were independently repeated three times. Figure 8d, it can be seen that the SNR of all spectra is improved by about 401.57% on average compared with the original spectra. At the same time, the corresponding error bar is narrow (±18.71%), which proves the stability of the proposed method in the denoising task. This paper focuses on denoising of ultra-weak signal spectra (SNR<10), which is a challenging task, but there are few studies on it.

[0128] Finally, we compared the detection depth improvement before and after using the U-Net denoising network. The characteristic peak 1203cm-1 is the strongest Raman peak of GERRTs, and we used this peak to verify the denoising effect and optical penetration limit of the method. The limit detection depth (LODT) is defined as the maximum tissue thickness corresponding to the SNR ≥ 3 at the characteristic peak 1203cm-1. The LODT based on the original spectrum is 10cm, while our denoising algorithm increases the LODT to more than 18cm ( Figure 8 ef). After using the self-supervised U-Net for noise reduction, we further applied the deep penetration equation to fit and estimated the value of LODT to be about 19.53cm. The dorsal-ventral distance of the human body is 15-30cm. Therefore, these results show that this method makes optical technology closer to the real scene in biomedical deep detection.

[0129] Furthermore, we compared the U-Net denoising algorithm with two commonly used denoising methods, SG and WTD, and the self-supervised CNN denoising algorithm using the deep penetration SERS data collected above. The SG algorithm is a typical algorithm based on moving window smoothing, while the WTD algorithm is an algorithm based on power spectrum estimation. The order and window size used by SG are 3 and 5, respectively, and WTD uses the Daubechies wavelet basis (db3) of 2 decomposition layers and the VisuShrink threshold function. The self-supervised CNN contains 4 blocks, each of which contains a convolutional layer with a kernel size of 2, a rectified linear unit and a pooling layer with a kernel size of 2, and finally 3 fully connected layers and uses the same self-supervised objective function as the method described above.

[0130] In order to more intuitively demonstrate the noise reduction effect, we define the signal-to-noise ratio improvement (iSNR), the formula is as follows:

[0131] iSNR=(SNR Denoised -SNR ori ) / SNR ori ×100%

[0132] Among them, SNR Denoised represents the SNR of the spectrum after noise reduction, SNR ori represents the SNR of the undenoised spectrum.

[0133] Fig. 9Four characteristic peaks (521, 932, 1203, 1581 cm-1) are included, which comprehensively demonstrates the noise reduction effect of different methods on the full spectrum. All spectra were independently repeated three times. Obviously, our method outperforms the other three methods on all characteristic peaks, and shows a uniform and reproducible noise reduction effect on a set of SERS spectra of different masses at different positions. The iSNRs values ​​of the four characteristic peaks are 723.41±181.28%, 533.17±117.71%, 540.37±120.33% and 617.66±160.13%, respectively. The iSNRs of the characteristic peaks after noise reduction by the three methods are shown in Table 5 (iSNR (%) of different noise reduction methods for deep penetration data).

[0134] Table 5

[0135]

[0136] SG uses an order of 3 and a window size of 5; WTD uses a wavelet basis of db3 and a decomposition level of 2

[0137] The best results are in bold.

[0138] For ultra-weak signals, we Fig. 9 The results of different methods when the signal-to-noise ratio is <10 are clearly shown in b. For such a challenging task, the performance of WTD drops significantly, and the performance of SG and CNN models is comparable. In contrast, our method achieves more than 500% improvement in all feature peaks, indicating that its stability and noise reduction effect are significantly better than other algorithms (see Table 6)

[0139] Table 6

[0140]

[0141] SG uses an order of 3 and a window size of 5; WTD uses a wavelet basis of db3 and a decomposition level of 2

[0142] like Fig.10 As shown, the masking method is used to enhance the interpretability of our method by calculating the Raman shift that contributes to the spectral denoising. The masking window size used is 100, and the corresponding area is zeroed. The resulting masked spectrum is input into the training model based on the unmasked spectrum. The Pearson correlation coefficient (PCC) is used to calculate the similarity between the masked spectrum result and the original denoised spectrum result: the smaller the absolute value of the PCC, the greater the difference between the two spectra, indicating that the contribution of certain areas is more important. We performed importance calculations on spectra with different signal-to-noise ratios, Fig.10Three representative results are shown in Figure BD. For SERS spectra with a signal-to-noise ratio of 0-190, as the signal-to-noise ratio increases, the CNN model pays more attention to the region with larger Raman intensity, while the U-Net network's attention is always more focused on learning the locations of characteristic peaks, which explains the superior denoising performance of the proposed method. It is noteworthy that even for ultra-weak signal spectra with SNR < 10, the proposed method can automatically identify most of the characteristic peaks that the model pays attention to. This attention encourages the network to fully utilize and learn the patterns (e.g., shape, half-peak width) of key peaks, thereby preserving them well. At the same time, since the noise affects the entire spectrum uniformly, the relative perturbation at the characteristic peaks is weak, which is conducive to better peak preservation. Although the model is less interested in the signal-free region at the end of the spectrum (Raman shift > 1700cm-1), it will eventually converge to the desired direction. The unique combination of self-supervised learning and downsampling strategy makes an important contribution to the stable denoising performance and high interpretability.

[0143] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A general self-supervised learning Raman spectroscopy denoising method, characterized by: include, Step 1: Acquire Raman data; Step 2: Raman data preprocessing, The baseline of Raman data was removed by using adaptive iterative reweighted penalized least squares method, and the spectrum after baseline removal was normalized to obtain preprocessed spectrum data; Step 3: Use the downsampling strategy H(·) to process the preprocessed spectral data to obtain two noise-independent sub-spectral sets h1(·) and h2(·), each of which includes 10 sub-spectra; Step 4: Input the preprocessed spectral data and the corresponding two sub-spectral sets into the U-Net network θ (·) to obtain the spectrum after noise reduction.

2. A general self-supervised learning Raman spectroscopy denoising method as claimed in claim 1, characterized in that: In step 2, the downsampling strategy H(·) slides on the spectrum with a window size of 3. For the three Raman shifts covered by the window, H(·) randomly selects two different shifts and copies their corresponding intensity values ​​to create a sub-spectrum with the same dimension as the original spectrum. The above operation is repeated twice to finally obtain the sub-spectrum sets h1(·) and h2(·).

3. A general self-supervised learning Raman spectroscopy denoising method as claimed in claim 1, characterized in that: The U-Net network f θ (·) contains 5 modules, The first module includes 2 one-dimensional convolution layers, 2 ReLU activation layers, 1 maximum pooling layer and 1 copy and splicing layer. The one-dimensional convolution layer, ReLU activation layer and maximum pooling layer are used to extract features. The extracted features are passed to the second module. At the same time, the extracted features are passed to the fifth module by retaining low-level features through the copy and splicing layer. The second module includes 2 one-dimensional convolution layers, 2 ReLU activation layers, 1 maximum pooling layer and 1 copy and splicing layer. The one-dimensional convolution layer, ReLU activation layer and maximum pooling layer are used to extract features. The extracted features are passed to the third module. At the same time, the extracted features are passed to the fourth module by retaining low-level features through the copy and splicing layer. The third module includes 2 one-dimensional convolution layers, 2 ReLU activation layers and 1 upsampling layer. The one-dimensional convolution layer and the ReLU activation layer extract features, and the extracted features are restored through the upsampling layer and passed to the fourth module. The fourth module includes 2 one-dimensional convolution layers, 2 ReLU activation layers and 1 upsampling layer. The convolution layer and the ReLU activation layer concatenate the features transmitted by the second module with the features transmitted by the third module, and then restore the concatenated features through the upsampling layer and pass them to the fifth module. The fifth module includes three one-dimensional convolutional layers and three ReLU activation layers. It concatenates the features delivered by the first module with the features delivered by the fourth module to reconstruct the spectrum.

4. A general self-supervised learning Raman spectroscopy denoising method as claimed in claim 1, characterized in that: The U-Net network f θ The training method of (·) is: Use stochastic gradient descent SGD to train the U-Net network f θ (·) for training, the loss function L is: L=λ1L reg +λ2(L rec1 +L rec2 ) Where: L reg =||y-f θ (y)||2 L rec1 =||h1(y)-f θ (h2(y))||2 L rec2 =||h2(y)-f θ (h1(y))||2 Where: λ = {λ1,λ2} is the weight of the corresponding loss term, represents the input Raman spectrum as an n-dimensional real vector, and h1(·) and h2(·) represent the two sub-spectrum sets obtained by downsampling in step 2.

5. A general self-supervised learning Raman spectroscopy denoising method as claimed in claim 4, characterized in that: The learning rate of the stochastic gradient descent method SGD is 0.

003.

6. A general self-supervised learning Raman spectroscopy denoising method as claimed in claim 4, characterized in that: U-Net network θ The training process of (·) is judged to converge based on the difference of the loss function of two iterations. When the loss difference is less than 10 -4 The model stops training when .

7. A general self-supervised learning Raman spectroscopy denoising device, characterized in that: include: A data acquisition module, used for acquiring Raman data; A first computing module, used for Raman data preprocessing; The second computing module is used to input the preprocessed spectral data and the corresponding two sub-spectral sets into the U-Net network. θ (·), obtain the spectrum after noise reduction; A noise reduction processing module is used to input the pre-processed spectral data and the corresponding two sub-spectral sets into the U-Net network f θ (·) to obtain the spectrum after noise reduction.

8. A computer device, characterized in that: include: Memory; One or more processors coupled to the memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by one or more processors, and the one or more applications are configured to execute the general self-supervised learning Raman spectroscopy denoising method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program codes, which can be called by a processor to execute the general self-supervised learning Raman spectroscopy denoising method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the computer program implements the general self-supervised learning Raman spectroscopy denoising method according to any one of claims 1 to 6.

Images